Zero-shot sketch-based image retrieval techniques using neural networks for sketch-image recognition and retrieval
This disclosure relates to improved sketch-based image retrieval (SBIR) techniques. The SBIR techniques utilize an architecture comprising three interconnected neural networks to enable zero-shot image recognition and retrieval based on free-hand sketches. Zero-shot learning may be implemented to retrieve one or more images corresponding to the sketches without prior training on all categories of the sketches. The neural network architecture may do so, at least in part, by training encoder hashing functions to mitigate heterogeneity of sketches and images, and by applying semantic knowledge that is learned during a limited training phase to unknown categories.
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:
encode images using an image encoding network that has been trained to mitigate heterogeneity associated with sketches and the images;
receive a query comprising a sketch from an unknown category;
encode the sketch using a sketch encoding network that has been trained to mitigate heterogeneity associated with the sketches and the images; and
utilize the encoded sketch to execute a zero-shot retrieval function for retrieving one or more of the images corresponding to the unknown category associated with the sketch;
wherein a multi-modal neural training network is utilized to train the sketch encoding network and the image encoding network to generate binary codes for the sketches and the images in a manner that mitigates heterogeneity of the sketches and the images; and
wherein a fused representation layer of the multi-modal neural training network is used to merge outputs from attention models associated with the sketch encoding network and the image encoding network to derive the binary codes for the sketches and the images.
2. The system of claim 1 , wherein the fused representation layer is a Kronecker product fusion layer.
3. The system of claim 1 , wherein executing the zero-shot retrieval function to retrieve the one or more images includes utilizing semantic knowledge derived from one or more known categories of information to identify the one or more images corresponding to the unknown category.
4. The system of claim 3 , wherein the multi-modal neural training network trains the sketch encoding network and the image encoding network to utilize the semantic knowledge via a graph convolutional network and a decoding model.
5. The system of claim 1 , wherein the attention models that are configured to identify portions of the images and the sketches that are informative for cross-modal mapping.
6. The system of claim 1 , wherein the sketch encoding network and the image encoding network are implemented as convolutional neural networks.
7. A method for performing sketch-based image retrieval (SBIR) comprising:
encoding images using an image encoding network that has been trained to mitigate heterogeneity associated with sketches and the images;
receiving a query comprising a sketch from an unknown category;
encoding the sketch using a sketch encoding network that has been trained to mitigate heterogeneity associated with the sketches and the images; and
utilizing the encoded sketch to execute a zero-shot retrieval function for retrieving one or more of the images corresponding to the unknown category associated with the sketch;
wherein a multi-modal neural training network is utilized to train the sketch encoding network and the image encoding network to generate binary codes for the sketches and the images in a manner that mitigates heterogeneity of the sketches and the images; and
wherein a fused representation layer of the multi-modal neural training network is used to merge outputs from attention models associated with the sketch encoding network and the image encoding network to derive the binary codes for the sketches and the images.
8. The method of claim 7 , wherein the fused representation layer is a Kronecker product fusion layer.
9. The method of claim 7 , wherein executing the zero-shot retrieval function to retrieve the one or more images includes utilizing semantic knowledge derived from one or more known categories of information to identify the one or more images corresponding to the unknown category.
10. The method of claim 9 , wherein the multi-modal neural training network trains the sketch encoding network and the image encoding network to utilize the semantic knowledge via a graph convolutional network and a decoding model.
11. The method of claim 7 , wherein the attention models are configured to identify portions of the images and the sketches that are informative for cross-modal mapping.
12. The method of claim 7 , wherein the sketch encoding network and the image encoding network are implemented as convolutional neural networks.
13. A computer program product for performing sketch-based image retrieval (SBIR), the computer program product comprising a non-transitory computer-readable medium including codes for causing a computer to:
encode images using an image encoding network that has been trained to mitigate heterogeneity associated with sketches and the images;
receive a query comprising a sketch from an unknown category;
encode the sketch using a sketch encoding network that has been trained to mitigate heterogeneity associated with the sketches and the images; and
utilize the encoded sketch to execute a zero-shot retrieval function for retrieving one or more of the images corresponding to the unknown category associated with the sketch;
wherein a multi-modal neural training network is utilized to train the sketch encoding network and the image encoding network to generate binary codes for the sketches and the images in a manner that mitigates heterogeneity of the sketches and the images; and
wherein a fused representation layer of the multi-modal neural training network is used to merge outputs from attention models associated with the sketch encoding network and the image encoding network to derive the binary codes for the sketches and the images.
14. The computer program product of claim 13 , wherein executing the zero-shot retrieval function to retrieve the one or more images includes utilizing semantic knowledge derived from one or more known categories of information to identify the one or more images corresponding to the unknown category.
15. 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:
encode images using an image encoding network that has been trained to mitigate heterogeneity associated with sketches and the images;
receive a query comprising a sketch from an unknown category;
encode the sketch using a sketch encoding network that has been trained to mitigate heterogeneity associated with the sketches and the images; and
utilize the encoded sketch to execute a zero-shot retrieval function for retrieving one or more of the images corresponding to the unknown category associated with the sketch;
wherein executing the zero-shot retrieval function to retrieve the one or more images includes utilizing semantic knowledge derived from one or more known categories of information to identify the one or more images corresponding to the unknown category; and
wherein a multi-modal neural training network trains the sketch encoding network and the image encoding network to utilize the semantic knowledge via a graph convolutional network and a decoding model.
16. A method for performing sketch-based image retrieval (SBIR) comprising:
encoding images using an image encoding network that has been trained to mitigate heterogeneity associated with sketches and the images;
receiving a query comprising a sketch from an unknown category;
encoding the sketch using a sketch encoding network that has been trained to mitigate heterogeneity associated with the sketches and the images; and
utilizing the encoded sketch to execute a zero-shot retrieval function for retrieving one or more of the images corresponding to the unknown category associated with the sketch;
wherein executing the zero-shot retrieval function to retrieve the one or more images includes utilizing semantic knowledge derived from one or more known categories of information to identify the one or more images corresponding to the unknown category; and
wherein a multi-modal neural training network trains the sketch encoding network and the image encoding network to utilize the semantic knowledge via a graph convolutional network and a decoding model.
17. A computer program product for performing sketch-based image retrieval (SBIR), the computer program product comprising a non-transitory computer-readable medium including codes for causing a computer to:
encode images using an image encoding network that has been trained to mitigate heterogeneity associated with sketches and the images;
receive a query comprising a sketch from an unknown category;
encode the sketch using a sketch encoding network that has been trained to mitigate heterogeneity associated with the sketches and the images; and
utilize the encoded sketch to execute a zero-shot retrieval function for retrieving one or more of the images corresponding to the unknown category associated with the sketch;
wherein executing the zero-shot retrieval function to retrieve the one or more images includes utilizing semantic knowledge derived from one or more known categories of information to identify the one or more images corresponding to the unknown category; and
wherein a multi-modal neural training network trains the sketch encoding network and the image encoding network to utilize the semantic knowledge via a graph convolutional network and a decoding model.