Bags of visual context-dependent words for generic visual categorization

• US 7,885,466 B2
• Filed: 09/19/2006
• Issued: 02/08/2011
• Est. Priority Date: 09/19/2006
• Status: Active Grant

• Current Assignee: Xerox Corporation (Xerox Holdings Corp.)
• Original Assignee: Xerox Corporation (Xerox Holdings Corp.)
• Inventors: Perronnin, Florent
• Primary Examiner(s): Bhatnagar; Anand
• Assistant Examiner(s): Liew; Alex
• Application Number:
  US11/524,100
• Publication Number:
  US 20080069456A1
• Time in Patent Office: 1,603 Days
• Field of Search: 382224-228, 382181-231
• US Class Current: 382/224
• CPC Class Codes: G06F 18/24 Classification techniques
• CPC Class Codes: G06V 10/462 Salient features, e.g. scale invariant feature transforms [SIFT]
• CPC Class Codes: G06V 10/464 using a plurality of salient features, e.g. bag-of-words [BoW] representations
• CPC Class Codes: G06V 10/764 using classification, e.g. of video objects

Abstract

Category context models (64) and a universal context model (62) are generated including sums of soft co-occurrences of pairs of visual words in geometric proximity to each other in training images (50) assigned to each category and assigned to all categories, respectively. Context information (76) about an image to be classified (70) are generated including sums of soft co-occurrences of pairs of visual words in geometric proximity to each other in the image to be classified. For each category (82), a comparison is made of (i) closeness of the context information about the image to be classified with the corresponding category context model and (ii) closeness of the context information about the image to be classified with the universal context model. An image category (92) is assigned to the image to be classified being based on the comparisons.

Claims

1. An image classifier comprising:

a vocabulary of visual words;

a patch context analyzer configured to generate a context representation for each of a plurality of patches of an image, each context representation being indicative of occurrence probabilities of context words in a plurality of neighboring patches wherein the plurality of neighboring patches are defined as one of (i) patches satisfying selected distance and size constraints respective to the patch whose context representation is being generated and (ii) pre-selected patches at pre-selected positions respective to the patch whose context representation is being generated; and

an image labeler configured to assign an image category to an image based at least on the context representations of a plurality of patches of the image.

2. The image classifier as set forth in claim 1, wherein the context words consist of a sub-set of visual words of the vocabulary that is substantially smaller than the total number of words in the vocabulary.

3. The image classifier as set forth in claim 1, wherein the context words include all visual words of the vocabulary.

4. The image classifier as set forth in claim 1, wherein the context words include at least some derived words that are not included in the vocabulary of visual words but that are derived from visual words that are included in the vocabulary of visual words.

5. The image classifier as set forth in claim 1, wherein the patch context analyzer generates the context representation for each patch as a plurality of values, each value being indicative of a statistical occupancy probability of a corresponding context word in neighboring patches weighted by closeness of the neighboring patches to the patch for which the context histogram is generated.

6. The image classifier as set forth in claim 1, wherein the image labeler applies a category context model for each image category, the category context model indicating probabilities of context words being in a neighborhood of an occurrence of a vocabulary word for images of that image category.

7. The image classifier as set forth in claim 1, further comprising:

a category context model generator configured to generate each category context model as sums of soft co-occurrences of pairs of words in geometric proximity to each other in training images assigned to the category.

8. The image classifier as set forth in claim 6, wherein the image labeler further applies a universal context model, the universal context model indicating probabilities of context words being in a neighborhood of an occurrence of a vocabulary word for images regardless of image category.

9. The image classifier as set forth in claim 8, further comprising:

a category context model generator configured to generate each category context model as sums of soft co-occurrences of pairs of words in geometric proximity to each other in training images assigned to the category; and

a universal context model generator configured to generate the universal context model as sums of soft co-occurrences of pairs of words in geometric proximity to each other in training images assigned to all categories.

10. The image classifier as set forth in claim 1, wherein the image labeler includes a plurality of comparators each comparing (i) closeness of the context representations of the patches of the image with a category context model and (ii) closeness of the context representations of the patches of the image with a universal context model, the image category being assigned based on the outputs of the comparators.

11. An image classifier comprising:

a patch context analyzer configured to generate a context representation for each of a plurality of patches of an image, the patch context analyzer comprising a context histogram generator configured to generate the context representation as a plurality of values each indicative of probabilistic number of occurrences of a corresponding context visual word in a neighborhood of a patch whose context representation is being generated, wherein the context histogram generator determines the probabilistic number of occurrences of each context visual word based on occupancy probabilities of the context visual word in a plurality of neighboring patches that neighbor the patch whose context representation is being generated, wherein the plurality of neighboring patches are defined as one of;

patches satisfying selected distance and size constraints respective to the patch whose context representation is being generated, and

pre-selected patches at pre-selected positions respective to the patch whose context representation is being generated; and

an image labeler including a plurality of comparators each comparing (i) closeness of context representations of a plurality of patches of an image with a category context model and (ii) closeness of the context representations of the plurality of patches of the image with a universal context model, the image labeler being configured to assign an image category to the image based on the outputs of the comparators.

12. An image classification method comprising:

for each patch of a plurality of patches of an image;

selecting a plurality of neighboring patches using a closeness measure that is based on criteria including (i) distance between patch centers; and

(ii) similarity of patch sizes or scales, and

generating a context representation for the patch based at least on occupancy probabilities of context words in the plurality of neighboring patches;

for each of a plurality of categories, generating a comparison of (i) closeness of the context representations of the image with a category context model representative of the category and (ii) closeness of the context representations of the image with a universal context model representative of all categories; and

assigning an image category to the image based on the generated comparisons;

wherein the image classification method is performed by a computing apparatus.

13. The image classification method as set forth in claim 12, wherein the generating of the context representations comprises:

generating a context histogram for each patch, each element of the context histogram being indicative of a likelihood of occurrence of a corresponding context word in a neighborhood of the patch.

14. The image classification method as set forth in claim 12, wherein the generating of the comparisons comprises:

generating each category context model as sums of soft co-occurrences of pairs of words in geometric proximity to each other in training images assigned to the category; and

generating the universal context model as sums of soft co-occurrences of pairs of words in geometric proximity to each other in training images assigned to all categories.