System and method for superresolution imaging from a sequence of color filter array (CFA) lowresolution images

0Associated
Cases 
0Associated
Defendants 
0Accused
Products 
29Forward
Citations 
0
Petitions 
1
Assignment
First Claim
1. A system for improving picture quality of color images by combining the content of a plurality of frames of the same subject comprising:
 at least one processor;
the at least one processor operatively connected to a memory for storing a plurality of frames of a subject;
the at least one processor operating to combine the content of plurality of frames of the subject into a combined color image by performing;
a process in which at least two multicolored frames are converted to monochromatic predetermined color frames;
a gross shift process in which the gross shift translation of one monochromatic predetermined color frame is determined relative to a reference monochromatic predetermined color frame;
a subpixel shift process utilizing a correlation method to determine the translational and/or rotational differences of one monochromatic predetermined color frame to the reference monochromatic predetermined color frame to estimate subpixel shifts and/or rotations between the frames; and
an error reduction process to determine whether the resolution of the resulting combined color image is of sufficient resolution;
the error reduction process comprising applying at least one spatial frequency domain constraint and at least one spatial domain constraint to the combined color image to produce at least one highresolution full color image.
1 Assignment
0 Petitions
Accused Products
Abstract
A method and system for improving picture quality of color images by combing the content of a plurality of frames of the same subject; comprising: at least one processor; the at least one processor comprising a memory for storing a plurality of frames of a subject; the at least one processor operating to combine the content of plurality of frames of the subject into a combined color image by performing: a process in which at least two multicolored frames are converted to monochromatic predetermined color frames; a gross shift process in which the gross shift translation of one monochromatic predetermined color frame is determined relative to a reference monochromatic predetermined color frame; a subpixel shift process utilizing a correlation method to determine the translational and/or rotational differences of one monochromatic predetermined color frame to the reference monochromatic predetermined color frame to estimate subpixel shifts and/or rotations between the frames; and an error reduction process to determine whether the resolution of the resulting combined color image is of sufficient resolution; the error reduction process comprising applying at least one spatial frequency domain constraint and at least one spatial domain constraint to the combined color image to produce at least one highresolution full color image.
62 Citations
View as Search Results
Algorithm for calculating high accuracy image slopes  
Patent #
US 9,218,641 B1
Filed 06/19/2012

Current Assignee
Harris Corporation

Sponsoring Entity
Harris Corporation

Method and apparatus for receiving a location of a vehicle service center from an image  
Patent #
US 9,558,419 B1
Filed 05/19/2015

Current Assignee
Blinker Inc.

Sponsoring Entity
Blinker Inc.

Method and apparatus for recovering a vehicle identification number from an image  
Patent #
US 9,563,814 B1
Filed 05/19/2015

Current Assignee
Blinker Inc.

Sponsoring Entity
Blinker Inc.

Method and apparatus for recovering a vehicle value from an image  
Patent #
US 9,589,201 B1
Filed 05/19/2015

Current Assignee
Blinker Inc.

Sponsoring Entity
Blinker Inc.

Method and apparatus for receiving an insurance quote from an image  
Patent #
US 9,589,202 B1
Filed 05/19/2015

Current Assignee
Blinker Inc.

Sponsoring Entity
Blinker Inc.

Method and apparatus for receiving listings of similar vehicles from an image  
Patent #
US 9,594,971 B1
Filed 05/19/2015

Current Assignee
Blinker Inc.

Sponsoring Entity
Blinker Inc.

Method and apparatus for receiving car parts data from an image  
Patent #
US 9,600,733 B1
Filed 05/19/2015

Current Assignee
Blinker Inc.

Sponsoring Entity
Blinker Inc.

Method and apparatus for providing loan verification from an image  
Patent #
US 9,607,236 B1
Filed 05/19/2015

Current Assignee
Blinker Inc.

Sponsoring Entity
Blinker Inc.

Method and apparatus for receiving vehicle information from an image and posting the vehicle information to a website  
Patent #
US 9,754,171 B1
Filed 05/19/2015

Current Assignee
Blinker Inc.

Sponsoring Entity
Blinker Inc.

Method and apparatus for obtaining a vehicle history report from an image  
Patent #
US 9,760,776 B1
Filed 05/19/2015

Current Assignee
Blinker Inc.

Sponsoring Entity
Blinker Inc.

Method and apparatus for receiving a broadcast radio service offer from an image  
Patent #
US 9,773,184 B1
Filed 05/19/2015

Current Assignee
Blinker Inc.

Sponsoring Entity
Blinker Inc.

Method and apparatus for verifying vehicle ownership from an image  
Patent #
US 9,779,318 B1
Filed 05/19/2015

Current Assignee
Blinker Inc.

Sponsoring Entity
Blinker Inc.

System and method for electronic processing of vehicle transactions based on image detection of vehicle license plate  
Patent #
US 9,818,154 B1
Filed 11/29/2016

Current Assignee
Blinker Inc.

Sponsoring Entity
Blinker Inc.

Method and apparatus for receiving a refinancing offer from an image  
Patent #
US 9,892,337 B1
Filed 05/19/2015

Current Assignee
Blinker Inc.

Sponsoring Entity
Blinker Inc.

Method and apparatus for recovering a vehicle identification number from an image  
Patent #
US 10,163,026 B2
Filed 02/07/2017

Current Assignee
Blinker Inc.

Sponsoring Entity
Blinker Inc.

Method and apparatus for receiving a location of a vehicle service center from an image  
Patent #
US 10,163,025 B2
Filed 01/30/2017

Current Assignee
Blinker Inc.

Sponsoring Entity
Blinker Inc.

Method and apparatus for receiving listings of similar vehicles from an image  
Patent #
US 10,169,675 B2
Filed 03/10/2017

Current Assignee
Blinker Inc.

Sponsoring Entity
Blinker Inc.

Method and apparatus for receiving an insurance quote from an image  
Patent #
US 10,176,531 B2
Filed 03/06/2017

Current Assignee
Blinker Inc.

Sponsoring Entity
Blinker Inc.

Method and apparatus for recovering a vehicle value from an image  
Patent #
US 10,192,130 B2
Filed 03/06/2017

Current Assignee
Blinker Inc.

Sponsoring Entity
Blinker Inc.

Method and apparatus for obtaining a vehicle history report from an image  
Patent #
US 10,192,114 B2
Filed 08/21/2017

Current Assignee
Blinker Inc.

Sponsoring Entity
Blinker Inc.

Method and apparatus for verifying vehicle ownership from an image  
Patent #
US 10,204,282 B2
Filed 09/22/2017

Current Assignee
Blinker Inc.

Sponsoring Entity
Blinker Inc.

Method and apparatus for receiving a refinancing offer from an image  
Patent #
US 10,210,417 B2
Filed 01/25/2018

Current Assignee
Blinker Inc.

Sponsoring Entity
Blinker Inc.

Method and apparatus for receiving vehicle information from an image and posting the vehicle information to a website  
Patent #
US 10,210,396 B2
Filed 08/21/2017

Current Assignee
Blinker Inc.

Sponsoring Entity
Blinker Inc.

Method and apparatus for receiving a broadcast radio service offer from an image  
Patent #
US 10,210,416 B2
Filed 09/22/2017

Current Assignee
Blinker Inc.

Sponsoring Entity
Blinker Inc.

Method and apparatus for providing loan verification from an image  
Patent #
US 10,242,284 B2
Filed 03/22/2017

Current Assignee
Blinker Inc.

Sponsoring Entity
Blinker Inc.

Method and apparatus for blocking information from an image  
Patent #
US 10,515,285 B2
Filed 06/07/2017

Current Assignee
Blinker Inc.

Sponsoring Entity
Blinker Inc.

Method and apparatus for identifying vehicle information from an image  
Patent #
US 10,540,564 B2
Filed 06/20/2018

Current Assignee
Blinker Inc.

Sponsoring Entity
Blinker Inc.

Method and apparatus for receiving a financing offer from an image  
Patent #
US 10,572,758 B1
Filed 05/19/2015

Current Assignee
Blinker Inc.

Sponsoring Entity
Blinker Inc.

Method and apparatus for recovering license plate information from an image  
Patent #
US 10,579,892 B1
Filed 02/03/2015

Current Assignee
Blinker Inc.

Sponsoring Entity
Blinker Inc.

Method for creating high resolution color image, system for creating high resolution color image and program creating high resolution color image  
Patent #
US 7,515,747 B2
Filed 10/08/2003

Current Assignee
Tokyo Institute of Technology

Sponsoring Entity
Tokyo Institute of Technology

Image enhancement by subpixel imaging  
Patent #
US 7,323,681 B1
Filed 12/28/2006

Current Assignee
Applera Corporation

Sponsoring Entity
Applera Corporation

Dynamic reconstruction of highresolution video from colorfiltered lowresolution videotovideo superresolution  
Patent #
US 7,379,612 B2
Filed 10/19/2006

Current Assignee
Regents of the University of California

Sponsoring Entity
Regents of the University of California

Generating super resolution digital images  
Patent #
US 7,218,751 B2
Filed 06/29/2001

Current Assignee
Digimarc Corporation

Sponsoring Entity
Digimarc Corporation

Bayesian approach for sensor superresolution  
Patent #
US 20070130095A1
Filed 09/30/2005

Current Assignee
Teledyne Scientific Imaging LLC

Sponsoring Entity
Teledyne Scientific Imaging LLC

Method for operating an Xray diagnostic device for the generation of highresolution images  
Patent #
US 20070196009A1
Filed 02/09/2007

Current Assignee
Siemens Healthcare GMBH

Sponsoring Entity
Siemens Healthcare GMBH

Method for obtaining a highresolution digital image  
Patent #
US 7,260,277 B2
Filed 04/22/2003

Current Assignee
Stmicroelectronics SRL

Sponsoring Entity
Stmicroelectronics SRL

Robust reconstruction of high resolution grayscale images from a sequence of low resolution frames  
Patent #
US 20070217713A1
Filed 11/16/2006

Current Assignee
Sina Farsiu, Peyman Milanfar, Michael Elad, Michael Robinson

Sponsoring Entity
Sina Farsiu, Peyman Milanfar, Michael Elad, Michael Robinson

Image processing method and image processing device  
Patent #
US 20070206678A1
Filed 03/02/2007

Current Assignee
Panasonic Corporation

Sponsoring Entity
Panasonic Corporation

Image processing system and method using subsampling with constraints such as time and uncertainty constraints  
Patent #
US 6,157,732 A
Filed 04/07/1999

Current Assignee
Cognex Corporation

Sponsoring Entity
Cognex Corporation

Method for creating high resolution color image, system for creating high resolution color image and program creating high resolution color image  
Patent #
US 20060038891A1
Filed 10/08/2003

Current Assignee
Tokyo Institute of Technology

Sponsoring Entity
Tokyo Institute of Technology

Synchronization of signals  
Patent #
US 20060053455A1
Filed 11/26/2003

Current Assignee
Entropic Communications Incorporated

Sponsoring Entity
Entropic Communications Incorporated

Method of superresolving images  
Patent #
US 20060159369A1
Filed 01/19/2005

Current Assignee
The United States of America As Represented By The Secretary of Agriculture

Sponsoring Entity
The United States of America As Represented By The Secretary of Agriculture

Bayesian image super resolution  
Patent #
US 7,106,914 B2
Filed 02/27/2003

Current Assignee
Microsoft Technology Licensing LLC

Sponsoring Entity
Microsoft Corporation

System and method for robust multiframe demosaicing and color super resolution  
Patent #
US 20060279585A1
Filed 08/17/2006

Current Assignee
University of California Santa Cruz

Sponsoring Entity
University of California Santa Cruz

Method of restoring and reconstructing superresolution image from lowresolution compressed image  
Patent #
US 20050019000A1
Filed 06/25/2004

Current Assignee
Yonsei University, SUNGJIN CC CO. LTD.

Sponsoring Entity
Yonsei University, SUNGJIN CC CO. LTD.

System and method for increasing space or time resolution in video  
Patent #
US 20050057687A1
Filed 06/17/2004

Current Assignee
Yeda Research and Development Co. Ltd.

Sponsoring Entity
Yeda Research and Development Co. Ltd.

Superresolution overlay in multiprojector displays  
Patent #
US 20040239885A1
Filed 04/19/2004

Current Assignee
University Of Kentucky Research Foundation

Sponsoring Entity
University Of Kentucky Research Foundation

Synthetic aperture MRI  
Patent #
US 20040263169A1
Filed 06/25/2004

Current Assignee
Calgary Scientific Inc.

Sponsoring Entity
Calgary Scientific Inc.

Method for subpixel registration of images  
Patent #
US 6,628,845 B1
Filed 03/15/2000

Current Assignee
NEC Corporation

Sponsoring Entity
NEC Corporation

Method of producing a high quality, high resolution image from a sequence of low quality, low resolution images that are undersampled and subject to jitter  
Patent #
US 6,650,704 B1
Filed 10/25/1999

Current Assignee
Nytell Software LLC

Sponsoring Entity
Irvine Sensors Corporation

System for improving image resolution via sensor rotation  
Patent #
US 6,642,497 B1
Filed 06/14/2002

Current Assignee
HewlettPackard Development Company L.P.

Sponsoring Entity
HewlettPackard Development Company L.P.

Superresolving imaging system  
Patent #
US 6,344,893 B1
Filed 06/19/2000

Current Assignee
Ramot University Authority For Applied Research and Industrial Development Limited

Sponsoring Entity
Ramot University Authority For Applied Research and Industrial Development Limited

Method and Arrangement for creating a highresolution still picture  
Patent #
US 6,349,154 B1
Filed 12/18/1998

Current Assignee
Uniloc 2017 LLC

Sponsoring Entity
US Philips Corporation

System and method for scaling combined video and computer generated imagery  
Patent #
US 6,424,749 B1
Filed 03/30/1999

Current Assignee
Godo Kaisha IP Bridge 1

Sponsoring Entity
Matsushita Electric Industrial Company Limited

Super resolution methods for electrooptical systems  
Patent #
US 6,483,952 B2
Filed 05/16/2001

Current Assignee
Harris Corporation

Sponsoring Entity
ITT Manufacturing Enterprises Inc.

Method and apparatus for improving image resolution  
Patent #
US 6,208,765 B1
Filed 06/19/1998

Current Assignee
Sarnoff Corporation

Sponsoring Entity
Sarnoff Corporation

Resolution improvement from multiple images of a scene containing motion at fractional pixel values  
Patent #
US 6,285,804 B1
Filed 12/21/1998

Current Assignee
Sharp Electronics Corporation

Sponsoring Entity
Sharp Laboratories of America Incorporated

Methods and systems for reproducing a high resolution image from sample data  
Patent #
US 6,023,535 A
Filed 08/30/1996

Current Assignee
Ricoh Company Limited

Sponsoring Entity
Ricoh Company Limited

Wavelet transform of superresolutions based on radar and infrared sensor fusion  
Patent #
US 5,952,957 A
Filed 05/01/1998

Current Assignee
United States Of America As Represented By The Secretary Of The Navy

Sponsoring Entity
United States Of America As Represented By The Secretary Of The Navy

Method and apparatus for combining multiple image scans for enhanced resolution  
Patent #
US 5,767,987 A
Filed 09/26/1994

Current Assignee
Ricoh Company Limited, Ricoh Corporation

Sponsoring Entity
Ricoh Company Limited, Ricoh Corporation

System for creating a high resolution image from a sequence of lower resolution motion images  
Patent #
US 5,696,848 A
Filed 01/17/1996

Current Assignee
Intellectual Ventures Fund 83 LLC

Sponsoring Entity
Eastman Kodak Company

Color imaging array  
Patent #
US 3,971,065 A
Filed 03/05/1975

Current Assignee
Eastman Kodak Company

Sponsoring Entity
Eastman Kodak Company

22 Claims
 1. A system for improving picture quality of color images by combining the content of a plurality of frames of the same subject comprising:
at least one processor;
the at least one processor operatively connected to a memory for storing a plurality of frames of a subject;
the at least one processor operating to combine the content of plurality of frames of the subject into a combined color image by performing;a process in which at least two multicolored frames are converted to monochromatic predetermined color frames; a gross shift process in which the gross shift translation of one monochromatic predetermined color frame is determined relative to a reference monochromatic predetermined color frame; a subpixel shift process utilizing a correlation method to determine the translational and/or rotational differences of one monochromatic predetermined color frame to the reference monochromatic predetermined color frame to estimate subpixel shifts and/or rotations between the frames; and an error reduction process to determine whether the resolution of the resulting combined color image is of sufficient resolution;
the error reduction process comprising applying at least one spatial frequency domain constraint and at least one spatial domain constraint to the combined color image to produce at least one highresolution full color image. View Dependent Claims (2, 3, 4, 5, 6, 7)
 8. A system for superresolving images from color filter array (CFA) lowresolution sequences comprising:
an image processor having an input for inputting multiple CFA images of a scene with subpixel translations and rotations;
one of the images of the scene being a reference image; andat least one memory comprising; a monochromatic color image constructing algorithm for creating at least one monochromatic color image for each image in the input sequence; a gross translation estimation algorithm for estimating overall translations of the at least one monochromatic color image with respect to the reference image on the image processor; a subpixel estimation algorithm for estimating the subpixel translations and/and or rotations of at least one monochromatic color image with respect to the reference image; and an error reduction algorithm for applying at least one spatial frequency domain constraint and at least one spatial domain constraint to the CFA low resolution images to produce at least one highresolution full color image.  View Dependent Claims (9, 10, 11, 12, 13, 14, 15)
 16. A method of superresolving CFA images affected by movement including translation and rotation comprising:
inputting frames of lowresolution CFA images of a scene into a processor;
at least one of the frames being affected by movement;obtaining at least one monochromatic color image for each image in the input frames; estimating the gross translations and subpixel shifts and rotations due to the movement using the monochromatic color images; and applying an error reduction algorithm utilizing at least one spatial frequency domain constraint and at least one spatial domain constraint to produce a highresolution full color output image.  View Dependent Claims (17, 18, 19, 20, 21, 22)
1 Specification
This application is a ContinuationInPart of U.S. patent application Ser. No. 12/576,132, filed Oct. 8, 2009, entitled System and Method of SuperResolution Imaging from a Sequence of Translated and Rotated LowResolution Images, and which in turn is a ContinuationInPart of application Ser. No. 11/038,401, filed on Jan. 19, 2005, now U.S. Pat. No. 7,602,997, entitled “Method of SuperResolving Images.” Priority is being claimed under 35 U.S.C. §120 to both application Ser. Nos. 12/576,132 and 11/038,401, and both are hereby incorporated by reference. This application is related to U.S. application Ser. No. 12/968,881, filed Dec. 15, 2010, hereby incorporated by reference.
The invention described herein may be manufactured, used, and/or licensed by or for the United States Government.
This invention relates in general to a method of digital image processing, and more specifically relates to superresolution image reconstruction.
The terminology “subpixel” or “subpixels” as used herein refers to one or more components of a single pixel. In a color subpixelated image, the components of a single pixel comprise several primary color primaries, which may be three ordered color elements such as blue, green, and red (BGR), or red, green, and blue (RGB), or green, blue, red. Some image displays have more than three primaries, further including yellow (RGBY), or white (RGBW), or yellow and cyan (RGBYC). Subpixels may appear as a single color when viewed directly by the human eye due to spatial integration by the eye. However, the components may be visible under a magnification. Upon achieving a predetermined resolution, the colors in the subpixels are not visible, but the relative intensity of the components may cause a shift in the apparent position or orientation of a line. The resolution at which subpixels go unnoticed may depend upon the viewer or the application as the visual processes of humans vary; whereby the colored “fringes” resulting from subpixel rendering may cause distraction in some individuals or applications but not others. Methods that account for subpixel rendering are generally called subpixel rendering algorithms.
Image resolution relates to the detail that an image possesses. For satellite images, the resolution generally correlates to the area represented by each pixel. Generally speaking, an image is considered to be more accurate and detailed as the area represented by each pixel is decreased. As used herein, the term images include digital images, film images, and/or other types of images.
As used herein, the terminology “sampling” refers to the practice of collecting a subset of individual observations intended to yield some knowledge about a subject. “Sampling” includes the act, process, or technique of selecting a representative part of a population or sequence of moving images for the purpose of determining parameters or characteristics of the whole population or moving image sequence. For example, accord ing to the 1949 sampling theorem by C E Shannon, in order to recover a signal function ƒ(t) precisely, the sampling rate must be done at a rate greater than twice the signal'"'"'s highest frequency component.
Many lowcost sensors (or cameras) may spatially or electronically undersample an image. Similarly, cameras taking pictures from great distances, such as aerial photos, may not obtain detailed information about the subject matter. This may result in aliased images in which the high frequency components are folded into the low frequency components in the image. Consequently, subtle or detail information (high frequency components) are not present in the images.
Many digital cameras acquire images using a single image sensor overlaid with a color filter array (CFA), and demosaicing is required to render these images into a viewable format. “Demosaicing process” is a digital image process of reconstructing a full color image from incomplete color samples output from an image sensor or camera overlaid with a color filter array (CFA).
When an image is captured by a monochrome camera, a single chargecoupled device (CCD) or complementary metaloxide semiconductor (CMOS) sensor is used to sample the light intensity projected onto the sensor. Color images are captured in much the same way, except that the light intensity is measured in separate color channels, usually red, green, and blue. In order to do this, three separate sensors could be used in conjunction with a beam splitter to accurately measure each of the three primary colors at each pixel. However, this approach is expensive and mechanically difficult to implement, making its use in commercial imaging systems infeasible. To overcome this obstacle, the color filter array (CFA) was introduced to capture a color image using only one sensor.
A CFA is an array that is used in front of the image sensors to alternate color filters that samples only one color channel at each pixel location. The most popular and common CFA is arranged in mosaic pattern, called the Bayer pattern, as described in U.S. Pat. No. 3,971,065, “Color image array,” July 1976, B. E. Bayer, which is illustrated in
This pattern results in half of the image resolution being dedicated to accurate measurement of the green color channel and quarter of the image resolution of the red or blue color channel. The peak sensitivity of the human visual system lies in the medium wavelengths, justifying the extra green sampling as described in X. Li, B. Gunturk, and L. Zhang, “Image demosaicing: a systematic survey,” Proc. SPIE, Vol. 6822, 68221J, 2008. Because each pixel now has only one color sampled, in order to produce a threechannel full color image, that is, each pixel contains three color channel values, missing information needs to be estimated from surrounding pixels of CFA pattern raw data. This is called demosaicing algorithm or process. The simplest demosaicing algorithm is linear interpolation applied to every color channel. More advanced demosaicing methods are summarized in X. Li, B. Gunturk, and L. Zhang, “Image demosaicing: a systematic survey,” Proc. SPIE, Vol. 6822, 68221J, 2008; D. Alleysson and B. C. de Lavarene, “Frequency selection demosaicking: a review and a look ahead,” Proc. SPIEIS&T Electonic Imaging, Vol. 6822, 68221M, 2008.
Sampling by color filter array (CFA) causes severe aliasing in addition to the aliasing caused by undersampling of many lowcost sensors, e.g., the CCD aperture.
The resolution of a color image reconstructed from the demosaicing method is equal to the physical resolution of a CCD. However, the resolution of an image sensor can be improved by a digital image processing algorithm: superresolution image reconstruction.
Superresolution image reconstruction generally increases image resolution without necessitating a change in the design of the optics and/or detectors by using a sequence (or a few snapshots) of lowresolution images. Superresolution image reconstruction algorithms effectively dealias undersampled images to obtain a substantially aliasfree or, as identified in the literature, a superresolved image.
When undersampled images have subpixel shifts between successive frames, they contain different information regarding the same scene. Superresolution image reconstruction involves, inter alfa, combining information contained in undersampled images to obtain an aliasfree (highresolution) image. Superresolution image reconstruction from multiple snapshots, taken by a detector which has shifted in position, provides far more detail information than any interpolated image from a single snapshot.
Methods for superresolving images that are acquired from CFA sensors are in the following references.
One of these methods is to capture multiple appropriately positioned CFA images that are used to fill in the “holes” in a Bayer pattern, as for example in U.S. Pat. No. 7,218,751, “Generating super resolution digital images,” May 15, 2007, A. M. Reed and B. T. Hannigan. Values of the pixels in multiple images which are appropriately aligned to each pixel position are averaged to generate a better value for each pixel position. In this method, in order to produce a useful result, information carried by a digital watermark is used to determine the alignment of the images.
Another type of these methods is called separate approach or twopass algorithm. The CFA images are first demosaicked and then followed by the application of superresolution, as for example in U.S. Pat. No. 7,260,277, “Method for obtaining a highresolution digital image,” Aug. 21, 2007, G. Messina, S. Battiato, and M. Mancuso. Each CFA input image is subjected to an interpolation phase to generate a complete lowresolution with three color channels containing three primary colors in RGB format and is thus linearly transformed into a complete lowresolution image in the YCrCb format, where Y represents the luminance component, Cr and Cb represent two chrominance components. After this, a modified back projection superresolution approach, based in M. Irani and S. Peleg, “Super resolution from image sequence,” Proceedings of the 10th International Conference on Pattern Recognition, Vol. 2, pages 115120, is applied only to the luminance component Y of the multiple images.
The third type of these methods is called joint or onepass method in which demosaicing and superresolution algorithms are simultaneously carried for a CFA image sequence, as for example in U.S. Pat. No. 7,515,747, “Method for creating high resolution color image, system for creating high resolution color image and program creating high resolution color image,” Apr. 7, 2009, M. Okutomi and T. Goto; U.S. Pat. No. 7,379,612, “Dynamic reconstruction of highresolution video from colorfiltered lowresolution videotovideo superresolution,” May 27, 2008, P. Milanfar, S. Farsiu, and M. Elad. In this method, an observation model is formulated to relate the original highresolution image to the observed lowresolution images to incorporate color measurements encountered in video sequences. Then, the highresolution full color image is obtained by solving an inverse problem by minimizing a cost function which is the function of the difference between the estimated lowresolution input images and the measured input images by imposing penalty terms, such as smoothness of chrominance and intercolor dependencies.
There is a need to produce highresolution three channel full color images from a sequence of lowresolution images captured by a lowcost CFA imaging device that has experienced translations and/or rotations. The amount of subpixel translation and/or rotation may be unknown and it creates a need to estimate subpixel translation and rotation. In order to produce highresolution images from a sequence of CFA undersampled (lowresolution) images, there exists a need to eliminate aliasing, while taking into account natural jitter.
The present invention provides a method and system for the superresolving images reconstructed from sequences of CFA lower resolution imagery affected by undersampled color filter array sensor.
An embodiment of the present invention utilizes a sequence of undersampled CFA lowresolution images from the same scene with subpixel translations and/or rotations among the images to reconstruct an aliasfree highresolution threechannel full color image, or to produce a sequence of such aliasfree highresolution full color frames as in a video sequence. The undersampled CFA low resolution images may be captured in the presence of natural jitter or some kind of controlled motion of the camera. For example, the natural jitter may be the result of a vehicle traveling down a bumpy road or an aircraft experiencing turbulence. The present invention is adaptable to situations where prior knowledge about the highresolution full color image is unknown, as well when unknown, irregular or uncontrolled subpixel translations and/or rotations have occurred among the lowresolution CFA images.
As depicted in
According to a preferred embodiment of the present invention, there is provided a system for improving picture quality of color images by combing the content of a plurality of frames of the same subject; comprising:
 at least one processor; the at least one processor comprising a memory for storing a plurality of frames of a subject; the at least one processor operating to combine the content of plurality of frames of the subject into a combined color image by performing:
 a process in which at least two multicolored frames are converted to monochromatic predetermined color frames;
 a gross shift process in which the gross shift translation of one monochromatic predetermined color frame is determined relative to a reference monochromatic predetermined color frame;
 a subpixel shift process utilizing a correlation method to determine the translational and/or rotational differences of one monochromatic predetermined color frame to the reference monochromatic predetermined color frame to estimate subpixel shifts and/or rotations between the frames; and
 an error reduction process to determine whether the resolution of the resulting combined color image is of sufficient resolution; the error reduction process comprising applying at least one spatial frequency domain constraint and at least one spatial domain constraint to the combined color image to produce at least one highresolution full color image.
The present invention can best be understood when reading the following specification with reference to the accompanying drawings, which are incorporated in and form a part of the specification, illustrate alternate embodiments of the present invention, and together with the description, serve to explain the principles of the invention. In the drawings:
The present invention now will be described more fully hereinafter with reference to the accompanying drawings, in which embodiments of the invention are shown. However, this invention should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art. In the drawings, the thickness of layers and regions may be exaggerated for clarity. Like numbers refer to like elements throughout. As used herein the term “and/or” includes any and all combinations of one or more of the associated listed items.
The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the full scope of the invention. As used herein, the singular forms “a”, “an” and “the” are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms “comprises” and/or “comprising,” when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and/or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof.
Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein.
With reference to
Subpixel transformation, also known as fraction pixel displacement, subpixel shift, or subpixel displacement, is estimated by a variety of methods in both patents and literature. Frametoframe motion detection based on gradient decent methods are most commonly used (see: (1) U.S. Pat. No. 5,767,987, “Method and apparatus for combining multiple image scans for enhanced resolution,” Jun. 16, 1998, G. J. Wolff and R. J. Van Steenkiste; (2) U.S. Pat. No. 6,650,704, “Method of producing a high quality, high resolution image from a sequence of low quality, low resolution images that are undersampled and subject to jitter,” Nov. 18, 2003, R. S. Carlson, J. L. Arnold, and V. G. Feldmus; (3) U.S. Pat. No. 6,285,804, “Resolution improvement from multiple images of a scene containing motion at fractional pixel values,” Sep. 4, 2001, R. J. Crinon and M. I. Sezan; (4) U.S. Pat. No. 6,349,154, “Method and arrangement for creating a highresolution still picture,” Feb. 19, 2002, R. P. Kleihorst). One of the variations of these methods is to estimate the velocity vector field measurement based on spatiotemporal image derivative (see U.S. Pat. No. 6,023,535, “Methods and systems for reproducing a high resolution image from sample data,” Feb. 8, 2000, S. Aoki). Most of these methods need to calculate matrix inversion or use iterative methods to calculate the motion vectors. Bergen (see U.S. Pat. No. 6,208,765, “method and apparatus for improving image resolution,” Mar. 27, 2001, J. R. Bergen) teaches a method to use warp information to obtain subpixel displacement. Stone et al (see U.S. Pat. No. 6,628,845, “Method for subpixel registration of images,” Sep. 30, 2003, H. S. Stone, M. T. Orchard, EC Chang, and S. Martucci) teaches another method that is to estimate the phase difference between two images to obtain subpixel shifts. In this method, the minimum least square solution has to be obtained to find the linear Fourier phase relationship between two images. U.S. Pat. No. 7,602,997, invented by the inventor herein, utilizes a correlation method without solving the minimum least square problem to explore the translational differences (shifts in x and y domains) of Fourier transforms of two images to estimate subpixel shifts between two low resolution aliased images.
Translation and Rotation Estimation
Two image frames of a lowresolution image sequence may be designated ƒ_{1}(x,y) and ƒ_{2}(u,v). Because of the conditions under which the frames are acquired, the second image is a shifted and/or rotated version of the first image, that is,
where (x,y) represent the coordinate plane in the first image, and (u,v) represent the coordinate plane in the shifted and rotated second image, as shown in
If the relative shifts (u_{s},v_{s}) are obtained, the shifts (x_{s}, y_{s}) can be obtained. Hereinafter, both (x_{s}, y_{s}) and (u_{s}, v_{s}) may be used for the relative shifts. The transformation parameters between images including translation (shift) (x_{s}, y_{s}) and rotation θ are estimated using a signal registration method and properties of Fourier transform.
Signal Registration—
The signal registration method is described in the following. Almost all of the transformation parameter estimation methods are based on the signal registration method. It is assumed that the signals to be matched differ only by an unknown translation. In practice, beside translation, there may be rotation, noise, and geometrical distortions. Nevertheless, the signal registration description offers a fundamental solution to the problem. The procedure for pixellevel registration is to first calculate the correlation between two images ƒ_{1}(x,y) and ƒ_{2}(x,y) as:
where 1≦k≦N, 1≦l≦M and R is a similarity measure, such as a normalized crosscorrelation or an absolute difference function. For the (N−1, M−1) samples of r(k, l), there is a maximum value of r(x_{s}, y_{s}), where (x_{s}, y_{s}) is the displacement in sampling intervals of ƒ_{2}(x,y) with respect to ƒ_{1}(x,y). Many transformation parameter estimation methods are based on the variations of the signal registration.
A signal registration based correlation method is used for translation parameter estimation between two images based on 1) the correlation theorem; and 2) the shift property of Fourier transform. This method begins with the crosscorrelation between two images. The terminology “crosscorrelation” or “cross correlation” as used herein is a measure of similarity of two frames in terms of a function of one of them. The twodimensional normalized crosscorrelation function measures the similarity for each translation:
where 1≦k≦N, 1≦l≦M and (N, M) is the dimension of the image. If ƒ_{1}(x,y) matches ƒ_{2}(x,y) exactly, except for an intensity scale factor, at a translation of (x_{s}, y_{s}), the crosscorrelation has its peak at C(x_{s}, y_{s}). Various implementations and discussions can be found for example in A. Rosenfeld and A. C. Kak, Digital Picture Processing, Vol. II pp. 3639, Orlando, Fla.: Academic Press, 1982; L. G. Brown, “A survey of image registration techniques,” ACM Computing Survey, Vol. 24, No. 4, December 1992, pp. 325376, both of which are hereby incorporated by reference.
Correlation Theorem—
More efficient signal registration method for translation estimation is using the correlation method in the spatial frequency domain via the Correlation Theorem. The Correlation Theorem states that the Fourier transform of the crosscorrelation of two images is the product of the Fourier transform of one image and the complex conjugate of the Fourier transform of the other, that is,
Q_{12}(K_{x},k_{y})=ℑ[C(x,y)]=F_{2}(k_{x},k_{y})·F_{1}*(k_{x},k_{y}) Equation (7A)
where F_{1}(k_{x},k_{y}) and F_{2}(k_{x},k_{y}) are Fourier transforms of two images, ƒ_{1}(x,y) and ƒ_{2}(x,y), respectively, F_{1}*(k_{x}, k_{y}) is the conjugate of F_{1}(k_{x}, k_{y}).
Fourier Transform Properties—
The properties of Fourier transform are described in the following that include shift property of Fourier transform, polar transformation and polar function mapping, and rotation property of Fourier transform. A Fourier transform is an operation that transforms one complexvalued function of a real variable into a new function. In such applications as image processing, the domain of the original function is typically space and is accordingly called the spatial domain. The domain of the new function is spatial frequency, and so the Fourier transform is often called the spatial frequency domain (or Fourier domain) representation of the original function as it describes which frequencies are present in the original function. As used herein the terminology “Fourier transform” refers both to the frequency domain representation of a function and to the process or formula that “transforms” one function into the other.
Shift Property of Fourier Transform—
Two members of the image sequence are denoted by ƒ_{1}(x,y) and ƒ_{2}(x,y). The second image is a shifted version of the first image, such as,
ƒ_{2}(x,y)=ƒ_{1}(x−x_{s},y−y_{s}) Equation (8A)
where (x_{s}, y_{s}) are the relative shifts (translational differences) between ƒ_{1}(x,y) and ƒ_{2}(x,y). Then,
F_{2}(k_{x},k_{y})=F_{1}(k_{x},k_{y})exp [−j(k_{x}x_{s}+k_{y}y_{s})] Equation (9A)
where F_{1}(k_{x}, k_{y}) and F_{2}(k_{x}, k_{y}) are Fourier transforms of two images. Thus, a shift in one domain results in the addition of a linear phase function in the Fourier domain.
Polar Transform and Polar Function Mapping—
The polar transformation of the domain (x,y) for a 2D signal ƒ(x,y) may be represented by
ƒ_{p}(θ,r)≡ƒ(x,y) Equation (10A)
where
and
r≡√{square root over (x^{2}+y^{2})} Equation (11AB)
Note that ƒ(x,y)≠ƒ_{p}(x,y). This equation is called the polar function mapping of ƒ(x,y). The polar function mapping of F(k_{x},k_{y}) in the spatial frequency domain may be defined via
F_{p}(φ,ρ)≡F(k_{x},k_{y}) Equation (12A)
where
and
ρ≡√{square root over (k_{x}^{2}+k_{y}^{2})} Equations (13AB)
It is noted that F_{p}(·,·) is not the Fourier transform of ƒ_{p}(·,·).
Rotation Property of Fourier Transform—
Two members of the image sequence are denoted ƒ_{1}(x,y) and ƒ_{2}(u,v), the second image being a rotated version of the first image,
where θ is the relative rotation (rotational differences) between ƒ_{1}(x,y) and ƒ_{2}(u,v). Then,
F_{2}(k_{u},k_{v})=F_{1}(k_{x},k_{y}) Equation (14A)
where [k_{u},k_{v}] and [k_{x},k_{y}] have the following relation,
Using the polar function mappings of F_{1}(k_{y},k_{y}) and F_{2}(k_{u},k_{v}), then
F_{p2}(φ,ρ)=F_{p1}(φ−θ,ρ) (Equation 16A)
That is, if one image is a rotated version of another image, then the polar function mapping of one is the shifted version of the polar function mapping of the other one in the spatial frequency domain.
Gross Translation Estimation Theory
The gross translation estimation (Box 214, and, alternatively, Boxes 320, 340, 360) is described in the following. The second image can be expressed as a shifted version of the first image, such as, ƒ_{2}(u,v)=ƒ_{1}(x,y) Equation (1A)
where (x_{s}, y_{s}) are the relative shifts (translational differences) between ƒ_{1}(x,y) and ƒ_{2}(u,v). The shifts (x_{s}, y_{s}) are found according to the Correlation Theorem based on the signal registration method by exploring the translational differences (shifts in x and y domains) of Fourier transforms of two images.
According to the Correlation Theorem, the Fourier transform of the crosscorrelation of the two images, ƒ_{1}(x,y) and ƒ_{2}(x,y) can be found as follows:
Q_{12}(k_{x},k_{y})=ℑ[C(x,y)]=F_{2}(k_{x},k_{y})·F_{1}*(k_{x},k_{y}) Equation (7A)
where F_{1}(k_{x}, k_{y}) and F_{2}(k_{x},k_{y}) are Fourier transforms of two images, ƒ_{1}(x,y) and ƒ_{2}(x,y), respectively, F_{1}*(k_{x},k_{y}) is the conjugate of F_{1}(k_{y},k_{y}).
Using the shift property of Fourier transform, the Fourier transform, F_{2}(k_{x}, k_{y}), can be substituted via
F_{2}(k_{x},k_{y})=F_{1}(k_{x},k_{y})exp [−j(k_{x}x_{s}+k_{y}y_{s})] Equation (9A)
Then, the Fourier transform of the crosscorrelation of the two images, ƒ_{1}(x,y) and ƒ_{2}(x,y), that is, the product of the Fourier transform of one image and the complex conjugate of the Fourier transform of the other, can then be expressed as
Then, the shifts (x_{s}, y_{s}) are found as the peak of inverse Fourier transform of the crosscorrelation, Q_{12 }(k_{x},k_{y}).
Translation Estimation with Aliasing Consideration—
Since both images ƒ_{1}(x,y) and ƒ_{2}(u,v) are aliased, i.e., the high frequency components are folded into a portion of low frequency components. In general, for an aliased image, the image energy at lower frequency components is stronger than the folded high frequency components. Thus, the signal to aliasing power ratio is relatively high around (k_{x},k_{y})≈(0,0). Therefore, a lowpass filtering, W, is applied to the Fourier transforms F_{1}(k_{x},k_{y}) and F_{2}(k_{x}, k_{y}). A schematic diagram of the computational efficient translation estimation method as described in U.S. Pat. No. 7,602,997, “Method of SuperResolving Images,” filed Jan. 19, 2005, S. S. Young, hereby incorporated by reference, is shown in
Since the Fourier transform can be computed efficiently with existing, welltested programs of FFT (fast Fourier transform) (and occasionally in hardware using specialized optics), the use of the FFT becomes most beneficial for cases where the images to be tested are large as described in L. G. Brown, “A survey of image registration techniques,” ACM Computing Survey, Vol. 24, No. 4, December 1992, pp. 325376.
SubPixel Translation Estimation
Translation Estimation with Subpixel Accuracy and Aliasing Consideration—
In superresolution image reconstruction of a preferred embodiment, subpixel shifts (or fractional pixel displacements) among the input images are estimated using a method to estimate or calculate frametoframe motion within subpixel accuracy. In general, the smallest motion that can be obtained from conventional motion estimation is one pixel. Obtaining subpixel motion accurately is not straightforward. A method to achieve the subpixel accuracy solution has been described in the aforementioned U.S. Pat. No. 7,602,997. This method is based on resampling via frequency domain interpolation by using a Fourier windowing method as described in S. S. Young, “Aliasfree image subsampling using Fourierbased windowing methods,” Optical Engineering, Vol. 43, No. 4, April 2004, pp. 843855 (hereby incorporated by reference), called power window.
Because the lowresolution input images contain aliasing, a lowpass filter is applied to the input images first. A schematic diagram of the computational efficient subpixel translation estimation method as described in U.S. Pat. No. 7,602,997 is shown in
Theory of Translation and Rotation Estimation
The subpixel translation and rotational estimation is described in the following. Assuming the second image is a shifted and rotated version of the first image, i.e.,
where θ is the relative rotation and (u_{s}, v_{s}) the relative shifts between ƒ_{1}(x,y) and ƒ_{2}(u,v).
From the shift property of Fourier transform of these images, the following relation is derived:
F_{2}(k_{u},k_{v})=F_{1}(k_{x},k_{y})·exp [−j(k_{u}u_{s}+k_{v}v_{s})] Equation (9A)
where F_{2}(k_{u},k_{v}) and F_{1}(k_{x},k_{y}) are the Fourier transforms of the two images, and the (k_{u},k_{v}) domain is the rotated version of the (k_{x},k_{y}) domain; that is,
Then from the polar transformation, the rotation and shift properties of Fourier transform as described previously, the polar function mappings (see operation in Equation 12A) of the 2D Fourier transforms of the two images are related by
F_{p2}(φ,ρ)=F_{p1}(φ−θ,ρ)·exp [−j(k_{u}u_{s}+k_{v}v_{s})] Equation (19A)
Thus, for their magnitudes we have,
F_{p2}(φ,ρ)=F_{p1}(φ−θ,ρ) Equation (20A)
That is, if one image is a rotated and shifted version of the other image, then the magnitude of polar function mapping of one is the shifted version of the magnitude of polar function mapping of the other one in the spatial frequency domain.
Note the difference between Equation (20A) and Equation (16A). One image is represented as a rotated version of the other image in Equation (16A), while one image is represented as a rotated and shifted version of the other image in Equation (20A).
Therefore, the rotation angle difference θ is represented as a translational difference using the polar transformation and polar function mapping of Fourier transform. Using the correlation method, the rotation angle between the two images is estimated in a manner similar to the translation estimation.
The rotation estimation is described in the following in detail. Because the rotation difference is only represented in the first domain of the Fourier transform, the rotation angle θ is found as the peak of inverse of the marginal 1D Fourier transform of the crosscorrelation with respect to φ. The marginal 1D Fourier transform is an operation of performing the onedimensional Fourier transform of a 2D function f(x,y) with respect to one of the variables, x or y.
The marginal 1D Fourier transform of the crosscorrelation with respect to φ between the two images is defined as:
Q_{p12}(k_{φ},ρ)=ℑ_{p2}(k_{φ},ρ)·ℑ_{p1}*(k_{φ},ρ) Equation (21A)
where ℑ_{p1}(k_{φ}, ρ) and ℑ_{p2}(k_{φ}, ρ) are the marginal 1D Fourier transforms with respect to φ for the two images F_{p1}(φ, ρ) and F_{p2}(φ, ρ) (as defined in Equation 12A), respectively, and ℑ_{p1}(k_{φ}, ρ) is the complex conjugate of ℑ_{p1 }(k_{φ}, ρ)
After the rotation angle is estimated, the second image is rotated back by the angle −θ to form the backward rotated image ƒ_{2θ}(x,y). Then the shifts (x_{s}, y_{s}) between the two images are estimated between ƒ_{2θ}(x,y) and ƒ_{1}(x,y) according to the correlation method based on the Correlation Theorem and the shift property of Fourier transform as discussed previously.
SubPixel Rotation and Translation Estimation
Rotation and Translation Estimation with Subpixel Accuracy—
Similar as in the translation estimation with subpixel accuracy, the rotation estimation with subpixel accuracy is achieved by resampling via frequency domain interpolation. However, the resampling operation is only performed in the φ domain (as defined in Equation 13A) since the rotation angle is the only parameter to be estimated. The subsequent subpixel translation estimation, (x_{s}, y_{s}) between the backward rotated image ƒ_{2θ}(x,y) and the first image ƒ_{1}(x,y), is achieved by resampling for both k_{x }and k_{y }domains via frequency domain interpolation (see
Rotation with Aliasing Consideration
The common center area of the two images ƒ_{1}(x,y) and ƒ_{2}(u,v) contributes most to the rotation angle estimation. A radially symmetrical window is applied to both images to emphasize the center common region between the two images. This radially symmetrical window is also designed as a lowpass filter to achieve accurate rotation estimation by filtering out the aliased high frequency components in the input images. As mentioned previously, the terminology aliased refers to the high frequency components folded into a portion of low frequency components. In general, for an aliased image, the image energy at lower frequency components is stronger than the folded high frequency components. Thus, the signal to aliasing power ratio is relatively high around (k_{x},k_{y})≈(0,0).
The radially symmetrical and lowpass window, W, is applied to the Fourier transforms F_{1}(k_{x},k_{y}) and F_{2}(k_{u},k_{v}) in the first step of the rotation estimation procedure, that is,
{tilde over (F)}_{1}(k_{x},k_{y})=F_{1}(k_{x},k_{y})·W(k_{x},k_{y}) Equation (22AA)
{tilde over (F)}_{2}(k_{u},k_{v})=F_{2}(k_{u},k_{v})·W(k_{u},k_{v}) Equation (22AB)
SubPixel Rotation and Translation Estimation Procedure
A procedure to estimate translation and rotation parameters with subpixel accuracy between two images, ƒ_{1}(x,y) and ƒ_{2}(u,v), is summarized in the following.
 1) Input two lowresolution images ƒ_{1}(x,y) and ƒ_{2}(u,v) (Box 30).
 2) Obtain the 2D Fourier transforms of the two images, F_{1}(k_{x},k_{y}) and F_{2}(k_{u},k_{v}) (Box 31).
 3) Apply a radially symmetrical window W to the Fourier transforms of the two images to obtain {tilde over (F)}_{1 }(k_{x},k_{y}) and {tilde over (F)}_{2}(k_{u}, k_{v}) (Box 32), see Equation (22AA)(22AB).
 4) Construct the magnitudes of the polar function mappings of the two windowed images, F_{P1}(φ,ρ) and F_{P2}(φ,ρ) (Box 33), as defined in Equation (12A).
 5) Obtain the marginal 1D Fourier transforms with respect to φ for the magnitude of the two polar function mappings, ℑ_{P1}(k_{φ}, ρ) and ℑ_{P2}(k_{φ}, ρ) (Box 34).
 6) Form the crosscorrelation function of the two images, ℑ_{P1}(k_{φ},ρ) and ℑ_{P2}(k_{φ},ρ), to obtain Q_{P12}(k_{φ}, ρ) (Box 35), see Equation (21A).
 7) Estimate the rotation angle θ between two images ƒ_{1}(x,y) and ƒ_{2}(u,v) with subpixel accuracy from the crosscorrelation function Q_{P12 }(k_{φ}, ρ) (Box 36).
 8) Rotate the second image by the angle −θ to form the backward rotated image ƒ_{2θ}(x,y) (Box 37).
 9) Estimate the relative shift between ƒ_{2θ}(x,y) and ƒ_{1}(x,y) with subpixel accuracy to obtain the relative shifts (x_{s}, y_{s}) (Box 38).
Example of SubPixel Rotation and Translation
An illustration of translation and rotation estimation with subpixel accuracy is shown in FIGS. 7a, 7b, 8, and 9 of application Ser. No. 12/576,132. In the featured example, a highresolution forward looking infrared (FLIR) image was used to simulate a sequence of lowresolution undersampled images with translations and rotations.
An original FLIR tank image is shown in
 x_{s}=[0 3.1 7.7 12.3 14.1 17.3];
 y_{s}=[0 2.3 8.5 11.2 15.2 16.8];
 θ=[0 14.4 15.6 17.8 19.1 22.7];
The units of x_{s }and y_{s }are pixels and the unit of θ is degrees. FIG. 8 of application Ser. No. 12/576,132 shows six simulated lowresolution images with shifts and rotates among them.
After applying the correlation method for subpixel shift and rotation estimation to these 6 simulated lowresolution images, the subpixel shift and rotation values are obtained. The estimated subpixel shift and rotation values are all correct according to the actual values. FIG. 9 of application Ser. No. 12/576,132 shows the transformation of the plane of one image (image 6) into the target domain of the reference image (image 1).
The method of acquiring a sequence of images in superresolution image reconstruction in this embodiment may involve noncontrolled subpixel translations and rotations; e.g. natural jitter. This can be more cost effective and more practical than the controlled subpixel translation and rotation image acquisition. In many applications, the imager may be carried by a moving platform or the object itself may be moving. In military reconnaissance situations, the moving platform may be a person, an unmanned ground vehicle (UGV), or an unmanned aerial vehicle (UAV) flying in a circular motion or a forward motion. As a result, acquired images may contain subpixel translations and/or rotations. During the process of superresolution image reconstruction, after multiple subpixel translated and/or rotated CFA images are obtained, they are then combined to reconstruct the highresolution full color image by projecting lowresolution CFA images onto three highresolution grids representing red, green, and blue channels, respectively. Each grid is of a finer sampled grid than the CFA lower resolution grid. The subpixel translations and rotations found in images resulting from natural jitter pose a need for accurate estimation.
In Box 200 of
The procedure of obtaining the reconstructed green channel image (Box 212) is first described in the following.
Fourier Domain Reconstructing from Incomplete Sampled Data Based on Spectral Propertities of Rectangular and Hexagonal Sampling
Green Channel Candidate—
One important step in superresolution is to accurately estimate subpixel shifts and rotations among lowresolution images. Since the undersampled CFA images are captured either by natural jitter or some kind of controlled motion of the camera, the images of successive frames contain not only the subpixel shifts but also the integer pixel shifts.
Because CFA images contain severe aliasing, using the direct raw input images does not give accurate shift estimation. Using lowpass filtered version of the input images is a better choice because the lowfrequency spectrum is less corrupted by aliasing.
In general, it was found in K. A. Prabhu and A. N. Netravali, “Motion compensated component color coding,” IEEE Trans. Communications, Vol. COM30, pp. 25192527, December 1982; R. C. Tom and A. K. Katsaggelos, “Resolution enhancement of monochrome and color video using motion compensation,” IEEE Trans. Image Processing, Vol. 10, No. 2, February 2001; U.S. Pat. No. 7,515,747, “Method for creating high resolution color image, system for creating high resolution color image and program creating high resolution color image,” Apr. 7, 2009, M. Okutomi and T. Goto, that it was sufficient to estimate the motion field of a color sequence based on the luminance component only. Since a Bayer color filter array is used in this embodiment, the green channel is chosen as the luminance component for transformation parameter estimation. However, when the other CFAs are used, the luminance channel is chosen based on that CFA to be used.
The green channel in the Bayer pattern is sampled using a hexagonal sampling pattern. Obtaining the reconstructed green signal is equivalent to the process of reconstructing the original green signal at each pixel location from hexagonal sampled green signal in the Bayer pattern. Before describing the hexagonal sampling, we describe the properties of a more common sampling pattern, rectangular sampling.
TwoDimensional Rectangular Sampling—
Let ƒ_{δ}(x,y) be a twodimensional deltasampled signal of an arbitrary signal, ƒ(x,y):
where x_{m}=mΔ_{x }and y_{n}=nΔ_{y}, Δ_{x }and Δ_{y }are the sample spacings in the spatial x and y domain respectively. Then, the Fourier transform of the deltasampled signal, F_{8}(k_{x},k_{y}), is composed of repeated versions of the Fourier transform of the continuous signal F(k_{x},k_{y}), that is,
The repeated positions of this Fourier transform are at
for m=(−∞, ∞) and n=(−∞, ∞). Thus, provided that F(k_{x},k_{y}) has a finite support in the spatial frequency domain, for example,
where
then ƒ(x,y) can be recovered from ƒ_{δ}(x,y) via lowpass filter in the frequency domain.
as shown in
Where ρ(k_{x},k_{y})=√{square root over (k_{x}^{2}+k_{y}^{2})}, ρ_{0}=√{square root over ((1/Δ_{x})^{2}+(1/Δ_{y})^{2})}{square root over ((1/Δ_{x})^{2}+(1/Δ_{y})^{2})}.
Spectral Properties of CFA Blue and Red Channel with TwoDimensional Rectangular Sampling—
For CFA blue or red channel, it is a twodimensional rectangular sampling. However, if the sampling spaces of the CFA pattern in x and y domains are Δ_{x }and Δ_{y}, respectively, the sampling spaces of blue or red channel are Δ_{x c}=2Δ_{x }and Δ_{y,C}=2Δ_{y}, where C is for red or blue. Therefore, the cutoff frequency for red or blue channel is
in spatial frequency domain,
in radial spatial frequency domain.
TwoDimensional Hexagonal Sampling—
As used herein the terminology “hexagonal sampling” is depicted in the insert in
As discussed in M. Soumekh, Fourier Array Imaging, Prentice Hall, 1994, pp. 167177, the hexagonal sampling has the following linear transformation:
Note that rectangular sampling corresponds to the case of b=c=0. The inverse of the hexagonal transformation is also linear, that is,
The Jacobian of the hexagonal transformation is
Next, define
g(α,β)=ƒ(x,y)
or, equivalently
g[T_{1}(x,y),T_{2}(x,y)]=ƒ(x,y)
It is assumed that the imaging system provides evenly spaced samples of g(α, β) in the (α, β) measurement domain, for example, α_{m}=mΔ_{α} and β_{n}=nΔ_{β}. Our task is to reconstruct the function ƒ(x,y) from the available samples at (x_{mn},y_{mn}), where
α_{m}=mΔ_{α}=T_{1}(x_{mn},y_{mn})
β_{n}=nΔ_{β}=T_{2}(x_{mn},y_{mn})
It is shown following in M. Soumekh, Fourier Array Imaging, Prentice Hall, 1994, pp. 167177,
δ(x−x_{m},y−y_{n})=T′(x,y)δ(α−α_{m},β−β_{n}).
The deltasampled signal is defined via
Using above equation, ƒ_{δ}(x,y) is rewritten as follows:
Then, the Fourier transform of the deltasampled signal, F_{δ}(k_{x},k_{y}), is composed of repeated versions of the Fourier transform of the continuous signal F(k_{x},k_{y}), that is,
Using T_{1}(x,y)=ax+by and T_{2}(x,y)=cx+dy, the above equation can be rewritten as:
where
The repeated positions of this Fourier transform are at (k_{xmn}, k_{ymn}) for m=(−∞,∞) and n=(−∞, ∞).
Spectral Property of CFA Green Channel with TwoDimensional Hexagonal Sampling—
In the case of CFA green channel, the hexagonal sampling has the following linear transformation:
That is, a=1, b=1, c=−1, and d=1.
for m=(−∞, ∞) and n=(−∞, ∞).
The cutoff frequency is designed as
in spatial frequency domain, and
in the radial spatial frequency as shown in
Reconstructing CFA Green Channel and Red/Blue Channel—
The spectral property of the hexagonal sampled data provides the basis for reconstructing the original signal to avoid nonlinear aliasing.
in spatial frequency domain, and
in the radial spatial frequency (Box 630) based on the spectrum property of the CFA green signal. A lowpass filtering W_{g}(k_{x},k_{y}) with the cutoff frequency ρ_{s0,G }(Box 640) is applied to the Fourier transform F_{gδ}(k_{x},k_{y}) (Box 620) to generate F_{gL}(k_{x},k_{y}). That is,
F_{gL}(k_{x},k_{y})=F_{gδ}(k_{x},k_{y})W_{g}(k_{x},k_{y}).
Then the inverse Fourier transform of the lowpassed signal (Box 650), F_{gL}(k_{x},k_{y}), is denoted as {tilde over (ƒ)}_{g}(x,y), which is the reconstructed CFA green signal. The lowpass filter is obtained by a Fourier windowing method described in S. S. Young, “Aliasfree image subsampling using Fourierbased windowing methods,” Optical Engineering, Vol. 43, No. 4, pp. 843855, April 2004, hereby incorporated by reference.
Reconstructing red or blue channel is similar to the procedure of reconstructing the green channel except the cutoff frequency for the lowpass filter is designed as
in spatial frequency domain,
in radial spatial frequency domain, where C for red or blue channel.
Example of Reconstructing Green and Red/Blue Channels
An illustration of reconstructing green and red/blue channels is shown in
The original gray scale image of a nature scene is shown in
The hexagonal sampling grid for the green image is shown in
Gross Translation Estimation for Color Filter Array Image Sequences
The gross translation estimation (
SUBPIXEL ROTATION AND TRANSLATION ESTIMATION FOR COLOR FILTER ARRAY IMAGE SEQUENCES The subpixel translation and rotation estimation (Box 216) as described in the foregoing and in U.S. patent application Ser. No. 12/576,132, “System and Method of SuperResolution Imaging from a Sequence of Translated and Rotated LowResolution Images,” filed Oct. 8, 2009, S. S. Young is illustrated in
ErrorEnergy Reduction Algorithm
ErrorEnergy Reduction Algorithm for CFA Input Images—
An errorenergy reduction algorithm for CFA images is derived to reconstruct the highresolution aliasfree output image that is a threechannel full color image containing red, green and blue channels. A primary feature of this proposed errorenergy reduction algorithm for CFA images is that, for each color channel, we treat the spatial samples from lowresolution images that possess unknown and irregular (uncontrolled) subpixel shifts and/or rotations as a set of constraints to populate an oversampled (sampled above the desired output bandwidth) processing array. The estimated subpixel shifted and rotated locations of these samples and their values constitute a spatial domain constraint. Furthermore, the bandwidth of the aliasfree image (or the sensor imposed bandwidth) is the criterion used as a spatial frequency domain constraint on the oversampled processing array.
In an errorenergy reduction algorithm, the highresolution image is reconstructed by removing aliasing from low resolution images. The highresolution image is reconstructed by applying nonuniform interpolation using the constraints both in the spatial and spatial frequency domains. The algorithm is based on the concept that the errorenergy is reduced by imposing both spatial and spatial frequency domain constraints: the samples from the lowresolution images; and the bandwidth of the highresolution, aliasfree output image.
In accordance with the errorenergy reduction algorithm of the present invention, spatial samples from low resolution images that possess unknown and irregular (uncontrolled) subpixel shifts and rotations are treated as a set of constraints to populate an oversampled (sampled above the desired output bandwidth) processing array. The estimated subpixel locations of these samples and their values constitute a spatial domain constraint. Furthermore, the bandwidth of the aliasfree image (or the sensor imposed bandwidth) is the criterion used as a spatial frequency domain constraint on the oversampled processing array. One may also incorporate other spatial or spatial frequency domain constraints that have been used by others; e.g., positivity of the image in the spatial domain. Since each color channel presents a unique spatial and spatial frequency properties as described earlier, the spatial and spatial frequency domain constraints are designed accordingly.
Image Acquisition Model—
G(k_{x})=F(k_{x})H(k_{x})
where G(k_{x}), F(k_{x}), and H(k_{x}) are the Fourier transforms of g(x), ƒ(x), and h(x), respectively.
B_{o}=min(B_{t},B_{s})
where B_{o}, B_{t }and B_{s }are the bandwidths of the output target signature, the ideal target signature and the sensor PSF, respectively. The proposed superresolution image reconstruction algorithm cannot produce an image whose bandwidth exceeds B_{o}, because the dealiasing approach cannot recover the bandwidth beyond the sensor'"'"'s diffraction limit.
Assuming the measurement system stores the incoming images at the sample spacing of Δ_{x}, the measured target signature, g(x), is then sampled at the integer multiples of Δ_{x}, e.g., x_{n}=nΔ_{x}, n=0, ±1, ±2, . . . , ±N, to obtain the undersampled (aliased) image; the resultant can be modeled as the following deltasampled signal in the spatial domain (see
The Fourier transform of the above deltasampled signal is
An acquired image is aliased when
As mentioned previously, many low resolution sensors possess a spatial sample spacing Δ_{x }that violates the Nyquist rate. The emphasis of the superresolution image reconstruction algorithm in a preferred embodiment is to dealias a sequence of undersampled aliased images to obtain the aliasfree or a superresolved image. Again it should be emphasized that the resolution of the output aliasfree image is limited by the minimum bandwidth of the sensor and the ideal target signature.
One of the considerations in superresolution image reconstruction is the number of the undersampled images that are required to recover the aliasfree image. For monochrominance images, let the sampling space of the original undersampled images be A_{x}. From an informationtheory point of view, the sampling space of the output highresolution image that is reconstructed from k frames (with distinct subpixel shifts) is
in onedimensional imaging systems, and
in twodimensional imaging systems.
A numerical example is given in the following. An image array of 100×100 is initially acquired to represent a scene. The total number of pixels is 10,000. If the resolution of the image is desired to be 4 times in each direction, the image array of 400×400 is required. The total number of pixels for the highresolution image is 160,000. That means that it requires 16 lowresolution images to achieve the resolution improvement factor of 4. Therefore, the number of input lowresolution image frames is the square of the resolution improvement factor. In other words, the resolution improvement factor is the square root of the number of the input images.
Now consider a CFA image, an image array of n_{x}*n_{y }is initially acquired to represent a scene. The total number of pixels is N=n_{x}*n_{y}. The total number pixels in the green channel is N_{G}=N/2, while the total number of pixels in the blue or red channel is N_{C}=N/4, where C represents for red or blue channel. When k frames of CFA images are acquired to reconstruct the highresolution full color image, the total number of pixels of the green channel of the output highhighresolution image is
the one for the blue or red channel is
where C represents for red or blue channel. Therefore, the resolution improvement factor in the green channel is √{square root over (k/2)} in both x and y directions, and the one in the red or blue channel √{square root over (k/4)}. That is, for CFA images, let the sampling space of the original undersampled images be Δ_{x}. From an informationtheory point of view, the sampling space of the output highresolution full color image that is reconstructed from k frames (with distinct subpixel shifts) is
in the green channel,
in the red or blue channel in twodimensional imaging systems. In another words, the resolution improvement factor of the green channel is √{square root over (k/2)} and the one of the red or blue channel is √{square root over (k/4)}.
Theoretically, more frames should provide more information and, thus, larger bandwidth and/or higher resolution. However, the superresolution image quality does not improve when the number of processed undersampled images exceeds a certain value. This is due to the fact that the bandwidth recovery is limited by the minimum of the bandwidths of the sensor or ideal target signature.
Implementation of ErrorEnergy Reduction Algorithm for CFA Images—
For each color channel, the bandwidth of the input undersampled (aliased) images is designated B_{i}. The bandwidth of the aliasfree (highresolution) output image for that channel is denoted as B_{o}. In order to recover the desired (aliasfree) bandwidth, it is important to select a processing array with a sample spacing smaller than the sample spacing of the desired highresolution image. In this case, the 2D FFT of the processing array yields a spectral coverage B_{p }that is larger than the bandwidth of the output image.
The highresolution aliasfree image on the processing array (grid) is designated as p_{D}(x,y) where D E {G, C}, G for the green channel and C E {R,B} for the red or blue channel. The lowresolution (aliased) snapshots only provide a subset of samples of this image. The location of each aliased image on the processing grid is dictated by its corresponding subpixel shifts and rotations in the (x,y) domain.
In the following, the errorenergy reduction algorithm is described for the CFA input images containing translational shifts and/or rotational differences.
Inputs to the error energy reduction algorithm are a sequence of lowresolution CFA input images. In the preferred embodiment, the estimated gross shifts and subpixel shifts and/or rotations of each image with respect to the reference image are obtained for each color channel; usually the first image is chosen as the reference image. Denote ƒ_{1}(u_{i}, v_{j}) to be L CFA images from the input image sequence, where l=1, 2, . . . , L, i=1, 2, . . . , I, j=1, 2, . . . , J; I×J is the dimension of the input image. After the transformation parameters ({circumflex over (θ)}_{i},{circumflex over (x)}_{si},ŷ_{si}) are estimated for all L images, the transform coordinates are calculated from the estimated subpixel shifts and rotation angles as follows:
In the preferred embodiment, the transform coordinates are the same for all three color channels.
In the initializing step (Box 402), three processing arrays are initialized for three color channels by populating each grid using the input images according to the estimates of shifts and rotation angles, respectively. Each processing array is selected with a sample spacing that is smaller than the sample spacing of the desired highresolution image.
In this procedure, the available CFA image values (from multiple snapshots) for three color channels are assigned to each shifted and rotated grid location of three processing arrays, respectively. At the other grid locations of each processing array, zero values are assigned. This can be written as
where Dε{G,C}, G for the green channel and Cε{R,B} for the red or blue channel, (X_{m}, Y_{n}) is the grid location that is closest to the transformation coordinates [{circumflex over (x)}_{ijl}, ŷ_{ijl}], and m=1, 2, . . . , M, n=1, 2, . . . , N; M×N is the dimension of the processing grid. The input green channel ƒ_{1,G }follows the sampling pattern of the hexagonal sampling shown in
In Box 404, a 2D Fourier transform p_{1,D}(k_{x},k_{y}) is performed on the processing array, where Dε{G, C}, G for the green channel and Cε{R,B} for the red or blue channel. Its spectrum has a wider bandwidth than the true (desired) output bandwidth. Therefore, the spatial frequency domain constraint (Box 408) is applied, that is, replacing zeros outside the desired bandwidth.
In this step, the desired bandwidth for the green channel and the one for red or blue channel are determined first (Box 406). Based on the spectral properties of the hexagonal sampling and the resolution improvement factor for the green channel when L CFA images are acquired, the desired bandwidth for the green channel is obtained as
in the spatial frequency domain. From the spectral properties of the rectangular sampling and the resolution improvement factor for the red or blue channel when L CFA images acquired, the desired bandwidth for the red or blue channel is obtained as
in the spatial frequency domain.
The functions
P_{2,D}(k_{x},k_{y})=P_{1,D}(k_{x},k_{y})·W_{p,D}(k_{x},k_{y})
is formed where W_{p,D}(k_{x},k_{y}) is a windowing function of realizing the spatial frequency domain constraint using (B_{Ox,G},B_{Oy,G}) for the green channel and (B_{Ox,C},B_{Oy,C}) for the red or blue channel, respectively. In Box 410, the inverse 2D Fourier transform p_{2,D}(x,y) of the resultant array p_{2},D(k_{x},k_{y}) is performed for all three color channels, respectively.
The resultant inverse array is not a true image because the image values at the known grid locations are not equal to the original image values. Then the spatial domain constraint is applied for all three color channels (Box 412), that is, replacing those image values at the known grid locations with the known original image values.
In this step of applying spatial domain constraints using shifts and rotations (Box 412), we have the function
to form the current iteration output image p_{k+1,D}(X_{m},Y_{n}) from the previous iteration output image p_{k,D}(X_{m}, Y_{n}) where Dε{G, C}, G for the green channel and Cε{R,B} for the red or blue channel. That is, the known grid locations are replaced with the known original image values and other grid locations are kept as the previous iteration output image values.
Similarly, the use of the available information in the spatial and spatial frequency domains results in a reduction of the energy of the error at each iteration step. At the odd steps of the iteration, the error for each of three color channels is defined by
where Dε{G, C}, G for the green channel and Cε{R, B} for the red or blue channel.
In Box 414, the condition of the errorenergy reduction is examined by defining the following ratio:
where Dε{G,C}, G for the green channel and Cε{R,B} for the red or blue channel. If SNR_{2k+1,D}<SNR_{max,D }(where SNR_{max,D }is a preassigned threshold value), and the maximum iteration is not reached (Box 416), the iteration continues. If SNR_{2k+1,D}>SNR_{max,D}, that is, the stopping criterion is satisfied, the iteration is terminated. Before the final superresolution image is generated, the bandwidth of the output image is reduced (Box 418) from B_{p }to B_{o }using the Fourier windowing method (see S. S. Young 2004) for the green, red, and blue channel, respectively. Then the final superresolved full color image with a desired bandwidth is saved for the output (Box 420).
An illustration of superresolving images from CFA lowresolution sequences is shown in
In order to protect the license plate owner'"'"'s privacy, part of the license plate is covered.
Another example of scene of a car emblem is illustrated in
Potential Applications
The potential extent of possible use of this invention is described in the following. However, this description should not be construed as limited to the statements.
Speed Traffic Enforcement—
Current local governments use traffic enforcement cameras to enforce traffic regulation violations. A traffic enforcement camera (also road safety camera, road rule camera, photo radar, speed camera, Gatso) is an automated ticketing machine. It may include a camera which may be mounted besides, on, or over a highway or installed in an enforcement vehicle to detect traffic regulation violations, including speeding, vehicles going through a red traffic light, unauthorized use of a bus lane, for recording vehicles inside a congestion charge area and others. The latest automatic number plate recognition (ANPR) systems can be used for the detection of average speeds and use optical character recognition on images to read the license plates on vehicles. There are a number of possible factors that affect the ANPR software performance. One of these important factors is poor image resolution, usually because the plate is too far away but sometimes resulting from the use of a lowquality camera. In the case of camera recording a video (a sequence of images), this invention can process the recorded images to improve image quality of the license plate on vehicle. The enhanced license plate images are used to improve the performance of ANPR software. The invention is especially useful when the images are acquired from a far away distance and/or from a lowquality camera.
Large Crowd Event Security and Management—
In an event involving a large crowd, especially the types of events including circuses, sporting events, theatrical events, concerts, rallies, parades, etc., the security task is to prevent, where possible, crimes including theft, vandalism or assault through the deployment of trained and licensed security personnel. Beside the large number of event staff members needed to oversee this important task, camera monitoring is an important component in this type of event security and management. The invention can be used to improve image details of a human face, nomenclature on a jersey, or a moving object/vehicle, etc., from a distance, or from the periphery of the event location. Also at football games, a preferred embodiment could be used to enhance the readability of numbers and/or names on football uniforms.
Industrial Production Line and Fruit Inspection—
To measure the quality of a product, a product is examined for color and shape as it passes the inspection station, and accepted or rejected based on those factors. The components are placed at any angle on the conveyer while they are imaged from above. The inspection algorithm is used to detect defects based on features that are extracted from images. These extracted features include the color, the bonding box, the bounding circle, the perimeter, the convex hull (mathematical extraction), and the area. Similarly, to measure the quality of fruit, the colors and shapes are used. The fruit is imaged with a color digital imaging system, the images are analyzed, and results are reported to a host computer to determine the final disposition of each piece of fruit. In this type of application, the object to be examined is placed on a conveyer. A camera takes images of the object showing the effects of natural motion among them and subpixel shifts exist among them. A preferred embodiment can be used to combine the images to generate a higher resolution image of the object. That enhances the detail appearance of the object area and therefore improves the inspection algorithm performance by improving the features of the images inputted to the inspection algorithm.
Although various preferred embodiments of the present invention have been described herein in detail to provide for complete and clear disclosure, it will be appreciated by those skilled in the art, that variations may be made thereto without departing from the spirit of the invention.
As used herein, the terminology “subject” means: an area, a scene, an object or objects, a landscape, overhead view of land or an object or objects, or a combination thereof.
As used herein, the terminology “frame” means: a picture, an image or one of the successive pictures on a strip of film or video.
As used herein, the terminology “process” means: an algorithm, software, subroutine, computer program, or methodology.
As used herein, the terminology “algorithm” means: sequence of steps using computer software, process, software, subroutine, computer program, or methodology.
As used herein, the terminology “demosaicing process” is a digital image process of reconstructing a full color image from incomplete color samples output from an image sensor or camera overlaid with a color filter array (CFA). Demosaicing or demosaicking is also referred to as CFA interpolation or color reconstruction.
As used herein, the terminology “image sensor” means: a camera, charge coupled device (CCD), video device, spatial sensor, or range sensor. The image sensor may comprise a device having a shutter controlled aperture that, when opened, admits light enabling an object to be focused, usually by means of a lens, onto a surface, thereby producing a photographic image OR a device in which the picture is formed before it is changed into electric impulses.
The terminology “transformation” as used herein (not to be confused with the terminology Fourier transform) refers to changes induced by movement of the image capture device, including translation and rotation.
The terminology “processor” or “image processor” as used in the following claims includes a computer, multiprocessor, CPU, minicomputer, microprocessor or any machine similar to a computer or processor which is capable of processing algorithms.
The terminology “luminescence” or “luminescence contrast” refers to the difference in luminosity (or lightness) of an object as compared to the background. A more detailed description of the human visual reaction to various levels of contrast is found at NASA'"'"'s Applied Color Science site http://colorusage.arc.nasa.gov/design_lum_{—}1.php, hereby incorporated by reference. In general, the primary colors are been given luminances by design, to achieve white alignment of the monitor. Early designers of computer displays decided that equal R, G, and B digital codes should produce a particular neutral chromaticity, the “alignment white” of the monitor (usually 6500K or 5300K). Once the white alignment color was decided and the chromaticities of the three primaries were chosen the maximum luminances of the primaries were set so that luminances of the primaries would mix to give the alignment white. For example, the human eye is more sensitive to the color green. For example, the color green is used in the color array of
As used herein, the terminology “spatial antialiasing” refers to a technique in which distortion artifacts (also refer to as aliasing) are minimized when creating a highresolution image from a sequence of lower resolution images. Antialiasing relates to the process of removing signal components of a frequency higher than is able to be properly resolved by the optical recording or sampling device.
The terminology “operations” as used in the following claims includes steps, a series of operations, actions, processes, subprocesses, acts, functions, and/or subroutines.
As used herein the terminology “color channel” is used in the context of a grayscale image of the same size as a color image, made of just one of the primary colors. For example, an image from a standard digital camera generally has green, red, and blue channels. A grayscale image has just one channel.
The terminology “full color” as used herein includes three primary color channels where three colors are utilized to produce the color image or four or more primary color channels where four or more colors are used to produce the color image.
It should be emphasized that the abovedescribed embodiments are merely possible examples of implementations. Many variations and modifications may be made to the abovedescribed embodiments. All such modifications and variations are intended to be included herein within the scope of the disclosure and protected by the following claims. The invention has been described in detail with particular reference to certain preferred embodiments thereof, but it will be understood that variations and modifications can be effected within the spirit and scope of the invention.