Fast image classification by vocabulary tree based image retrieval

1. A method to categorize images, comprising:

detecting local features for each image;

applying a tree structure to index local features in the images;

extracting a rank list of candidate images with category tags based on a tree indexing structure to estimate a label of a query image, and

determining a category of the query image C(q) by

C

⁡

(

q

)

=

arg

⁢

⁢

max

N

c

=

1

⁢

∑

i

=

1

K

⁢

⁢

1

⁢

(

C

⁡

(

I

i

)

=

c

)

i

,

where retrieved top K candidate images I i are sorted according to matching score s i in descending order {I i ,s i } 1 K , s i >s j , ∀i<j and where C(I i ) is an image category index of I i ι.

wherein the above steps are computer implemented.

2. The method of claim 1 , comprising applying a weighted voting according to a rank.

3. The method of claim 1 , comprising performing vocabulary tree based image retrieval in a large-scale image category classifier.

4. The method of claim 3 , wherein the vocabulary tree based image retrieval comprises determining image content based nearest neighbor searching.

5. The method of claim 3 , wherein the vocabulary tree based image retrieval comprises classifying with a linear support vector machine (SVM).

6. The method of claim 1 , comprising generating inverted indices representing local invariant feature of an image.

7. The method of claim 1 , comprising performing a sematic category classification to generate a semantic label for the query image.

8. A system to categorize images, comprising:

means for detecting local features for each image;

means for applying a tree structure to index local features in the images;

means for extracting a rank list of candidate images with category tags based on a tree indexing structure to estimate a label of a query image, and

means for determining a category of the query image C(q)

C

⁡

(

q

)

=

arg

⁢

⁢

max

N

c

=

1

⁢

∑

i

=

1

K

⁢

⁢

1

⁢

(

C

⁡

(

I

i

)

=

c

)

i

,

where retrieved top K candidate images I i are sorted according to matching score s i in descending order {I i ,s i } 1 K , s i >s j , ∀i<j and where C(I i ) is an image category index of I i ,

wherein the above means are computer implemented.

9. The method of claim 1 , comprising ranking scores of all categories to train a multi-class linear SVM.

10. The method of claim 1 , comprising concatenating ranking scores build a feature xεR N for an image q:

x

=

{

∑

i

=

1

K

⁢

⁢

1

⁢

(

C

⁡

(

I

i

)

=

c

)

i

}

c

=

1

N

.

11. The system of claim 8 , comprising means for concatenating ranking scores build a feature xεR N for an image q:

x

=

{

∑

i

=

1

K

⁢

⁢

1

⁢

(

C

⁡

(

I

i

)

=

c

)

i

}

c

=

1

N

.

12. The system of claim 8 , comprising means for applying a weighted voting according to a rank.

13. The system of claim 8 , comprising means for performing vocabulary tree based image retrieval in a large-scale image category classifier.

14. The system of claim 8 , wherein the vocabulary tree based image retrieval means comprises means for determining image content based nearest neighbor searching.

15. The system of claim 13 , wherein the vocabulary tree based image retrieval means comprises means for classifying with a linear support vector machine (SVM).

16. The system of claim 8 , comprising means for generating inverted indices representing local invariant feature of an image.

17. The system of claim 8 , comprising means for performing a sematic category classification to generate a semantic label for the query image.

18. The system of claim 8 , comprising means for ranking scores of all categories to train a multi-class linear SVM.