Learning semantic image similarity
View Patent ↗Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for identifying similar images. In some implementations, a method is provided that includes receiving a collection of images and data associated with each image in the collection of images; generating a sparse feature representation for each image in the collection of images; and training an image similarity function using image triplets sampled from the collection of images and corresponding sparse feature representations.
1. A method comprising:
receiving an image search query;
providing image search results responsive to the image search query;
receiving a request for images similar to an identified image in the provided image search results;
determining similar images to the identified image using the identified image and obtained data generated from a plurality of image triplets, each image triplet including an image, a more relevant image to the image, and a less relevant image to the image, and wherein each image triplet has a corresponding vector representation; and
providing one or more similar images.
2. The method of claim 1 , wherein determining similar images includes identifying the image in a similarity matrix and identifying images in the similarity matrix having a threshold similarity to the image.
3. The method of claim 2 , wherein the similarity matrix is generated from the plurality of image triplets.
4. The method of claim 1 , wherein determining similar images to the identified image includes using a similarity function trained using the plurality of image triplets and the corresponding vector representations.
5. The method of claim 4 , wherein training the image similarity function further comprises iteratively sampling image triplets p, p i + , p i − from a collection of images such that the relative similarity, r, for r(p, p i + )>r(p, p i − ).
6. The method of claim 4 , wherein the plurality of image triplets is identified using collection of image search queries and a specified number of image resources identified in ranked results responsive to each image query.
7. The method of claim 1 , wherein the similar images are provided in a search results interface along with the identified image.
8. The method of claim 1 , wherein the vector representation for each image of the plurality of images is determined from a sparse feature representation generated for the image.
9. The method of claim 8 , wherein generating the sparse feature representation for the image comprises:
dividing the image into blocks;
generating an edge histogram and a color histogram for each block to determine local descriptors for the image; and
combining the local descriptors to obtain a sparse feature representation of the image.
10. A system comprising:
one or more computers configured to perform operations comprising:
receiving an image search query;
providing image search results responsive to the image search query;
receiving a request for images similar to an identified image in the provided image search results;
determining similar images to the identified image using the identified image and obtained data generated from a plurality of image triplets, each image triplet including an image, a more relevant image to the image, and a less relevant image to the image, and wherein each image triplet has a corresponding vector representation; and
providing one or more or similar images.
11. The system of claim 10 , wherein determining similar images includes identifying the image in a similarity matrix and identifying images in the similarity matrix having a threshold similarity to the image.
12. The system of claim 11 , wherein the similarity matrix is generated from the plurality of image triplets.
13. The system of claim 10 , wherein determining similar images to the identified image includes using a pairwise similarity function trained using the plurality of image triplets and the corresponding vector representations.
14. The system of claim 13 , wherein training the image similarity function further comprises iteratively sampling image triplets p, p i + , p i − from a collection of images such that the relative similarity, r, for r(p, p i + )>r(p i , p i − ).
15. The system of claim 13 , wherein the plurality of image triplets is identified using collection of image search queries and a specified number of image resources identified in ranked results responsive to each image query.
16. The system of claim 10 , wherein the similar images are provided in a search results interface along with the identified image.
17. The system of claim 10 , wherein the vector representation for each image of the plurality of images is determined from a sparse feature representation generated for the image.
18. The method of claim 17 , wherein generating the sparse feature representation for the image comprises:
dividing the image into blocks;
generating an edge histogram and a color histogram for each block to determine local descriptors for the image; and
combining the local descriptors to obtain a sparse feature representation of the image.
19. A non-transitory computer storage medium encoded with a computer program, the program comprising instructions that when executed by data processing apparatus cause the data processing apparatus to perform operations comprising:
receiving an image search query;
providing image search results responsive to the image search query;
receiving a request for images similar to an identified image in the provided image search results;
determining similar images to the identified image using the identified image and obtained data generated from a plurality of image triplets, each image triplet including an image, a more relevant image to the image, and a less relevant image to the image, and wherein each image triplet has a corresponding vector representation; and
providing one or more similar images.