Detecting objects of interest in still images

• US 8,401,250 B2
• Filed: 04/05/2010
• Issued: 03/19/2013
• Est. Priority Date: 02/19/2010
• Status: Active Grant

• Current Assignee: LTIMindtree Ltd. (Larsen & Toubro Limited)
• Original Assignee: Mindtree Limited (Larsen & Toubro Limited)
• Inventors: Sangappa, Hemanth Kumar, Kumaraswamy, Suresh Kirthi, Chandrashekhar, Puneeth B
• Primary Examiner(s): Carter, Aaron W
• Application Number:
  US12/753,914
• Publication Number:
  US 20110243431A1
• Time in Patent Office: 1,079 Days
• Field of Search: 382/118, 382/159, 382/195
• US Class Current: 382/118
• CPC Class Codes: G06V 40/161 Detection; Localisation; Normalisation

Abstract

A computer implemented method and system for detecting interest sections in a still image are provided. One or more sub images of the still image is subjected to segmentation. A gray scale version of interest sub images and/or a binary image version of the interest sub images are matched with a predefined template for filtering the interest sub images. Multiple prospective image sections comprising one or more of prospective interest sections and prospective near interest sections are determined by performing discriminative feature analyses of the filtered interest sub images using a gabor feature filter. The discriminative feature analyses are processed by a boosted cascade of classifiers. The boosted cascade of classifiers detects the interest sections in the still image from the prospective interest sections and the prospective near interest sections. The detected interest sections are subjected to a support vector machine classifier for further detecting interest sections.

Claims

1. A computer implemented method for detecting interest sections in a still image, comprising:

segmenting one or more sub images of said still image based on pixels of interest in one or more color spaces, said one or more sub images being categorized into one or more of interest sub images and a plurality of non interest sub images based on said segmentation;

matching one of a gray scale version of said interest sub images and a binary image version of said interest sub images with a predefined template for filtering said interest sub images based on relative densities of said pixels of interest and relative average intensities of predetermined features in said interest sub images;

determining a plurality of prospective image sections comprising one or more of a plurality of prospective interest sections and a plurality of prospective near interest sections by performing discriminative feature analyses of said filtered interest sub images, wherein said discriminative feature analyses are processed by a boosted cascade of classifiers; and

detecting said interest sections in said still image from said prospective interest sections and said prospective near interest sections by said boosted cascade of classifiers.

2. The computer implemented method of claim 1, wherein said discriminative feature analyses of said filtered interest sub images are performed using a gabor feature filter, wherein said discriminative feature analyses of said filtered interest sub images comprises:

resizing and normalizing said filtered interest sub images to a predetermined dimension;

subjecting said resized interest sub images to a gabor feature transform by passing said resized interest sub images through said gabor feature filter for obtaining first feature vectors corresponding to said prospective interest sections and said prospective near interest sections;

subjecting said obtained first feature vectors to a principal component analysis for obtaining two sets of prominent eigen vectors;

projecting said obtained first feature vectors on said prominent eigen vectors for obtaining final feature vectors; and

classifying said final feature vectors using one or more of said boosted cascade of classifiers for obtaining said interest sections and non interest sections.

3. The computer implemented method of claim 1, further comprising subjecting said detected interest sections to a support vector machine classifier, if said boosted cascade of classifiers detects at least one of said interest sections in said still image.

4. The computer implemented method of claim 3, wherein subjecting said detected interest sections to said support vector machine classifier comprises:

resizing and normalizing said detected interest sections to a predetermined dimension;

compensating brightness variations on said resized interest sections using a least squares fit plane;

performing histogram equalization for said compensated interest sections for adjusting contrast of grayscale versions of said compensated interest sections;

training said support vector machine classifier with labeled training samples as data points; and

subjecting said histogram equalized interest sections as data points to said support vector machine classifier for further detecting interest sections.

5. The computer implemented method of claim 1, further comprising estimating a tilt in each of said filtered interest sub images and correcting said estimated tilt based on said relative average intensities of said predetermined features in said filtered interest sub images for obtaining a plurality of corrected interest sub images.

6. The computer implemented method of claim 5, wherein said estimation and correction of said tilt in each of said filtered interest sub images comprises:

performing histogram equalization for said filtered interest sub images using a plurality of image histograms for adjusting contrast of said filtered interest sub images;

computing a horizontal intensity projection for said filtered interest sub images;

identifying locations of said predetermined features in said filtered interest sub images relative to one another using said horizontal intensity projection;

measuring average intensities on said filtered interest sub images;

comparing said average intensities corresponding to said predetermined features in said filtered interest sub images at different predefined tilt angles, wherein a maximum relative average intensity of two or more of said predetermined features corresponds to a compensation tilt in each of said filtered interest sub images; and

correcting said tilt by introducing said compensation tilt in each of said filtered interest sub images for obtaining said corrected interest sub images.

7. The computer implemented method of claim 1, further comprising obtaining said binary image version of said interest sub images from said still image, comprising:

analyzing said still image in one or more color spaces, comprising;

comparing each color component of each of a plurality of pixels in said still image with a threshold value for said color component based on predetermined criteria;

determining at least one of a plurality of pixels of interest and a plurality of pixels of non interest from among said pixels based on said comparison; and

obtaining said binary image version of said interest sub images comprising said determined pixels of interest based on said analysis of said still image.

8. The computer implemented method of claim 7, wherein said predetermined criteria for said comparison comprise a range of potential color tones of said prospective interest sections in said interest sub images and contrast quality of said still image.

9. The computer implemented method of claim 1, wherein said matching of said gray scale version of said interest sub images with said predefined template comprises:

obtaining a horizontal projection of intensities of a plurality of gray pixels in said gray scale version of said interest sub images;

measuring average intensities of said predetermined features in said interest sub images based on measurements from said horizontal projection; and

comparing said average intensities corresponding to said predetermined features in said interest sub images for mapping locations of said predetermined features relative to one another.

10. The computer implemented method of claim 1, wherein said boosted cascade of classifiers comprises one or more linear classifiers.

11. The computer implemented method of claim 1, wherein said pixels of interest are skin pixels.

12. The computer implemented method of claim 1, wherein said predetermined features in said interest sub images comprise eyes, nose, and cheek regions of a face.

13. The computer implemented method of claim 1, wherein said prospective interest sections are prospective face sections, and said prospective near interest sections are prospective near face sections.

14. The computer implemented method of claim 1, further comprising gabor feature extraction for training one or more of said boosted cascade of classifiers for determining a plurality of prospective image sections, comprising:

resizing and normalizing said filtered interest sub images to a predetermined dimension;

subjecting said resized interest sub images to a gabor feature transform by passing said resized interest sub images through a gabor feature filter for obtaining first feature vectors corresponding to said prospective interest sections and said prospective near interest sections;

obtaining a predetermined number of training samples from said first feature vectors corresponding to said prospective interest sections and said prospective near interest sections;

subjecting said training samples corresponding to said prospective interest sections and said prospective near interest sections to a principal component analysis for obtaining two sets of prominent eigen vectors;

projecting said training samples on said obtained prominent eigen vectors for obtaining gabor discriminative features; and

training said one or more of said boosted cascade of classifiers to recognize said interest sections and said non interest sections from said gabor discriminative features.

15. The computer implemented method of claim 14, wherein said two sets of prominent eigen vectors correspond to prominent eigen vectors of a face class and a near face class.

16. The computer implemented method of claim 14, wherein said interest sections correspond to face sections, and said non interest sections correspond to non face sections.

17. A computer implemented system for detecting interest sections in a still image, comprising:

an image segmentation module that segments one or more sub images of said still image based on pixels of interest in one or more color spaces, said one or more sub images being categorized into one or more of interest sub images and a plurality of non interest sub images based on said segmentation;

a template matching engine that matches one of a gray scale version of said interest sub images and a binary image version of said interest sub images with a predefined template for filtering said interest sub images based on relative densities of said pixels of interest and relative average intensities of predetermined features in said interest sub images;

a feature analyses module that determines a plurality of prospective image sections comprising one or more of a plurality of prospective interest sections and a plurality of prospective near interest sections by performing discriminative feature analyses of said filtered interest sub images using a gabor feature filter; and

a boosted cascade of classifiers that detects said interest sections in said still image from said prospective interest sections and prospective near interest sections.

18. The computer implemented system of claim 17, further comprising a support vector machine classifier that further detects interest sections in said still image, if said boosted cascade of classifiers detects at least one of said interest sections in said still image.

19. The computer implemented system of claim 17, further comprising a tilt estimation and correction module that estimates a tilt in each of said filtered interest sub images and corrects said estimated tilt based on said relative average intensities of said predetermined features in said filtered interest sub images for obtaining a plurality of corrected interest sub images.

20. The computer implemented system of claim 17, wherein said boosted cascade of classifiers comprises one or more linear classifiers.

21. The computer implemented system of claim 17, wherein said pixels of interest are skin pixels.

22. The computer implemented system of claim 17, wherein said predetermined features in said interest sub images comprise eyes, nose, and cheek regions of a face.

23. A computer program product comprising computer executable instructions embodied in a non-transitory computer readable storage medium, wherein said computer program product comprises:

a first computer parsable program code for segmenting one or more sub images of said still image based on pixels of interest in one or more color spaces, said one or more sub images being categorized into one or more of interest sub images and a plurality of non interest sub images based on said segmentation;

a second computer parsable program code for matching one of a gray scale version of said interest sub images and a binary image version of said interest sub images with a predefined template for filtering said interest sub images based on relative densities of said pixels of interest and relative average intensities of predetermined features in said interest sub images;

a third computer parsable program code for determining a plurality of prospective image sections comprising one or more of a plurality of prospective interest sections and a plurality of prospective near interest sections by performing discriminative feature analyses of said filtered interest sub images using a gabor feature filter, wherein said discriminative feature analyses are processed by a boosted cascade of classifiers; and

a fourth computer parsable program code for detecting interest sections in said still image from said prospective interest sections and said prospective near interest sections by said boosted cascade of classifiers.

24. The computer program product of claim 23, further comprising a fifth computer parsable program code for subjecting said detected interest sections to a support vector machine classifier, if said boosted cascade of classifiers detects at least one of said interest sections in said still image.

25. The computer program product of claim 23, further comprising a sixth computer parsable program code for estimating a tilt in each of said filtered interest sub images and correcting said estimated tilt based on said relative average intensities of said predetermined features in said filtered interest sub images for obtaining a plurality of corrected interest sub images.