Sketch-based image retrieval techniques using generative domain migration hashing
This disclosure relates to improved sketch-based image retrieval (SBIR) techniques. The SBIR techniques utilize a neural network architecture to train a domain migration function and a hashing function. The domain migration function is configured to transform sketches into synthetic images, and the hashing function is configured to generate hash codes from synthetic images and authentic images in a manner that preserves semantic consistency across the sketch and image domains. The hash codes generated from the synthetic images can be used for accurately identifying and retrieving authentic images corresponding to sketch queries, or vice versa.
1. A system for performing sketch-based image retrieval (SBIR) comprising:
one or more computing devices comprising one or more processors and one or more non-transitory storage devices for storing instructions, wherein execution of the instructions by the one or more processors causes the one or more computing devices to:
receive a query comprising a sketch;
execute a domain migration function configured to generate a synthetic image corresponding to the sketch;
execute a hashing function configured to generate a hash code corresponding to the synthetic image;
utilize the hash code to identify retrieval results comprising one or more authentic images corresponding to the sketch; and
wherein the domain migration function is learned using a neural network that includes a generative adversarial network that is trained to migrate the sketch into an image domain, at least in part, by applying an upsampling process to the sketch.
2. The system of claim 1 , wherein the generative adversarial network includes at least one generative network and at least one discriminative network that cooperate to learn the domain migration function that is used to translate the sketch into the synthetic image.
3. The system of claim 2 , wherein the at least one generative network and the at least one discriminative network compete in a two-player minimax game such that:
the at least one generative network is configured to generate synthetic images; and
the at least one discriminative network is configured to evaluate whether received inputs correspond to authentic images or synthetic images.
4. The system of claim 1 , wherein the generative adversarial network is trained to learn the hashing function that generates the hash code from the synthetic image in a manner that preserves semantic consistency across image and sketch domains.
5. The system of claim 1 , wherein:
the hash code generated by the hashing function is a binary code; and
a neural network is trained, at least in part, using a loss function that is configured to minimize a distance between binary codes of authentic images and synthetic images that are within the same category, and to maximize the distance between binary codes that are not within the same category.
6. The system of claim 1 , wherein the neural network is trained, at least in part, using a loss function that is configured to preserve visual similarities of intra-class sketches and authentic images.
7. The system of claim 1 , wherein execution of the instructions by the one or more processors further causes the one or more computing devices to:
generate hash codes for a plurality of authentic images using the hashing function; and
utilize the hash code generated for the synthetic image and the hash codes generated for the plurality of authentic images to identify the one or more authentic images to be included in the retrieval results.
8. The system of claim 7 , wherein generating the hash codes for the plurality of authentic images includes applying an attention model to suppress background information in the plurality of authentic images and to identify features that are representative of categories associated with the plurality of authentic images.
9. The system of claim 1 , wherein the system is integrated into one or more of: a law enforcement image retrieval system; an e-commerce system; or a surveillance system.
10. A method for performing sketch-based image retrieval (SBIR) comprising:
receiving a query comprising a sketch;
executing, with one or more processors, a domain migration function configured to generate a synthetic image corresponding to the sketch;
executing, with the one or more processors, a hashing function configured to generate a hash code corresponding to the synthetic image;
utilizing the hash code to identify retrieval results comprising one or more authentic images corresponding to the sketch; and
wherein the domain migration function is learned using a neural network that includes a generative adversarial network that is trained to migrate the sketch into an image domain, at least in part, by applying an upsampling process to the sketch.
11. The method of claim 10 , wherein the generative adversarial network includes at least one generative network and at least one discriminative network that cooperate to learn the domain migration function that is used to translate the sketch into the synthetic image.
12. The method of claim 11 , wherein the at least one generative network and the at least one discriminative network compete in a two-player minimax game such that:
the at least one generative network is configured to generate synthetic images; and
the at least one discriminative network is configured to evaluate whether received inputs correspond to authentic images or synthetic images.
13. The method of claim 10 , wherein the generative adversarial network is utilized to learn the hashing function that generates the hash code from the synthetic image in a manner that preserves semantic consistency across image and sketch domains.
14. The method of claim 10 , wherein:
the hash code generated by the hashing function is a binary code; and
the neural network is trained, at least in part, using a loss function that is configured to minimize a distance between binary codes of authentic images and synthetic images that are within the same category, and to maximize the distance between binary codes that are not within the same category.
15. The method of claim 10 , wherein the neural network is trained, at least in part, using a loss function that is configured to preserve visual similarities of intra-class sketches and authentic images.
16. The method of claim 10 further comprising:
generating hash codes for a plurality of authentic images using the hashing function; and
utilizing the hash code generated for the synthetic image and the hash codes generated for the plurality of authentic images to identify the one or more authentic images to be included in the retrieval results.
17. The method of claim 16 , wherein generating the hash codes for the plurality of authentic images includes applying an attention model to suppress background information in the plurality of authentic images and to identify features that are representative of categories associated with the plurality of authentic images.
18. A computer program product for performing sketch-based image retrieval (SBIR), the computer program product comprising a physical, non-transitory computer-readable medium including codes for causing a computer to:
receive a query comprising a sketch;
utilize a domain migration function configured to generate a synthetic image corresponding to the sketch;
execute a hashing function configured to generate a hash code corresponding to the synthetic image;
utilize the hash code to identify retrieval results comprising one or more authentic images corresponding to the sketch; and
wherein the domain migration function is learned using a neural network that includes a generative adversarial network that is trained to migrate the sketch into an image domain, at least in part, by applying an upsampling process to the sketch.