System and method for face detection using candidate image region selection

• US 6,463,163 B1
• Filed: 01/11/1999
• Issued: 10/08/2002
• Est. Priority Date: 01/11/1999
• Status: Expired due to Term

• Current Assignee: ASML Netherlands BV (ASML Holding NV), Carl Zeiss AG (Carl-Zeiss-Stiftung), Tarsium B.V.
• Original Assignee: Hewlett-Packard Company (HP Inc.)
• Inventors: Kresch, Renato
• Primary Examiner(s): Boudreau, Leo
• Assistant Examiner(s): CHOOBIN, BARRY
• Application Number:
  US09/229,052
• Time in Patent Office: 1,366 Days
• Field of Search: 382/103, 382/157, 382/228, 382/190, 382/118, 382/209, 382/100, 348/169
• US Class Current: 382/103
• CPC Class Codes: G06V 40/161 Detection; Localisation; Normalisation

Abstract

A face detection system and a method of pre-filtering an input image for face detection utilize a candidate selector that selects candidate regions of the input image that potentially contains a picture of a human face. The candidate selector operates in conjunction with an associated face detector that verifies whether the candidate regions contain a human face. In the preferred embodiment, the candidate selector includes a linear matched filter and a non-linear filter that operate in series to select the candidate regions from the input image. Initially, the linear matched filter performs a linear correlation on the input image using a filtering kernel to derive a correlation image. The linear matched filter selects regions of the input image that have a local maximum in the correlation image and have correlation values greater than a threshold correlation value. Preferably, the linear correlation is performed in the discrete cosine transform (DCT) domain. The non-linear filter then examines contrast values from various sub-regions of the image regions that were selected by the linear matched filter to screen for suitable candidate regions. The filtering kernel used by the linear matched filter is calculated during a training period, or a non-face detecting period, by a filtering-kernel generator. The filtering kernel is preferably computed utilizing a database of training face images that have been averaged and then masked to remove DC, illumination and noise components.

Claims

1. A method of pre-filtering an input image to determine whether said input image includes a target image pattern of interest comprising steps of:

receiving said input image;

correlating said input image with preselected reference data to derive a correlation image, said preselected reference data being representative of said target image pattern of interest, said correlation image having image data indicative of correlations between said input image and said preselected reference data; and

selecting a candidate region of said input image having a likelihood of containing said target image pattern of interest on a basis of correlating said input image with said preselected reference data, said step of selecting including identifying potential candidate regions of said input image based on detecting a local maximum in said correlation image and screening said potential candidate regions by analyzing a grayscale characteristic of said potential candidate regions, said screening including discarding said potential candidate regions that do not satisfy a predetermined criterion with respect to said grayscale characteristic in order to increase a probability that said selected candidate region contains said target image pattern of interest.

2. The method of claim 1 wherein said step of screening said potential candidate regions further includes comparing a contrast value calculated for a sub-region of each potential candidate region to a threshold contrast value, said target image pattern of interest being a human face.

3. The method of claim 2 wherein said step of comparing said contrast value is a step of comparing said contrast value from said sub-region of said each potential candidate region that is located in an upper-half portion of said each potential candidate region.

4. The method of claim 1 further comprising a step of calculating said preselected reference data utilizing a plurality of training images containing said target image pattern of interest, thereby forming a convolution kernel.

5. The method of claim 4 wherein said step of forming said convolution kernel includes masking a particular frequency component of an averaged image derived from said training images to remove an undesired image characteristic from said averaged image.

6. The method of claim 1 wherein said step of correlating said input image with said preselected reference data includes removing an undesired image characteristic from said input image, said removal being effectuated by a mathematical property of said preselected reference data.

7. The method of claim 1 wherein said step of selecting said candidate region further includes comparing correlation values associated with each of said potential candidate regions to a threshold correlation value.

8. The method of claim 1 wherein said step of correlating said input image with said preselected reference data is performed in a discrete cosine transform (DCT) domain.

9. A method of pre-filtering an input image to select a candidate region of interest that potentially contains a predetermined image pattern comprising steps of:

receiving said input image;

examining a grayscale characteristic of a particular image region of said input image as a determination factor in determining whether said particular image region has a likelihood of containing said predetermined image pattern, including comparing a contrast value from a sub-region of said particular image region to a threshold contrast value as a basis for said determination, said particular image region being identified by correlating said input image with a reference image; and

designating said particular image region as said candidate region for said image pattern detection.

10. The method of claim 9 further comprising a step of selecting said particular image region from a plurality of image regions of said input image, said selection being at least partly determined by identifying which of said image regions includes a local maximum in a correlation image that is generated by correlating said input image with a convolution kernel that is representative of said predetermined image pattern, said predetermined image pattern being indicative of a human face.

11. The method of claim 10 wherein said step of selecting said particular image region from said plurality of image regions includes comparing correlation values that correspond to each identified image region having said local maximum to a correlation threshold value.

12. The method of claim 10 further comprising a step of generating said convolution kernel by utilizing a plurality of training images, said convolution kernel having a correlation property to filter an undesired image characteristic of said input image during a correlation operation.

13. The method of claim 12 wherein said step of generating said convolution kernel includes masking a frequency component of each training image, said frequency component corresponding to said undesired image characteristic.

14. The method of claim 9 wherein said step of comparing said contrast value from said sub-region is a step of comparing said contrast value from said sub-region located in an upper-half portion of said image region.

15. An image processing system for detecting a target image pattern in an input image comprising:

first filtering means configured for identifying a plurality of image regions of said input image that have a likelihood of containing said target image pattern, said filtering means being configured to correlate said input image with a reference image pattern to select said image regions, said reference image pattern being indicative of said target image pattern;

second filtering means operatively coupled to and operatively following said first filtering means for screening said image regions to select a candidate region that has a high likelihood of containing said target image pattern, said second filtering means being configured to examine a grayscale characteristic of said image regions as a determination factor in selecting said candidate region from among said image regions; and

an image pattern detector connected to and operatively following said second filtering means to receive said candidate region, said image pattern detector designed to analyze said candidate region to verify whether said candidate region contains said target image pattern.

16. The system of claim 15 wherein said first filtering means includes a linear matched filter configured to perform a linear convolution operation on said input image with a convolution kernel, said convolution kernel being a numerical representation of said reference image pattern, said reference image pattern being representative of a human face.

17. The system of claim 16 wherein said linear matched filter is a discrete cosine transform (DCT) linear matched filter that is designed to perform said linear convolution operation in a DCT domain.

18. The system of claim 16 further comprising a calculating means operatively associated with said linear matched filter for generating said convolution kernel from a plurality of training images, said calculating means including a masker to remove a particular frequency component of said training images that corresponds to an undesired image characteristic.

19. The system of claim 15 wherein said second filtering means includes a non-linear filter coupled in series to said first filtering means, said non-linear filter configured to select said candidate region by comparing a computed contrast value from a sub-region of each identified image region to a threshold contrast value.