Image processing method, apparatus, and device, and storage medium

1. An image processing method, performed by an image processing device, comprising:

obtaining an image, and performing feature extraction on the image, to obtain a feature map of the image;

generating at least one candidate region of the image, and mapping the at least one candidate region to the feature map of the image, each candidate region comprising one instance;

processing the mapped feature map based on a target network for instance segmentation; and

indicating an overall structure of an occluded instance in the image by using a perspective mask, and indicating an invisible part of the occluded instance by using a non-perspective mask, the perspective mask and the non-perspective mask representing a relative occlusion relationship of the occluded instance, wherein:

the target network comprises a first branch structure configured to obtain the non-perspective mask and the perspective mask, a second branch structure configured to output occlusion classification information of an instance comprised in the candidate region, and a third branch configured to output a semantic label and position information of the instance in the image; and

processing the mapped feature map based on the target network for instance segmentation comprises: fusing feature information output by the second branch structure, feature information output by the third branch structure, and feature information corresponding to the first branch structure to obtain fused feature information; and applying the fused feature information in the first branch structure to obtain the perspective mask and the non-perspective mask.

2. The method according to claim 1 , wherein after the mapping the at least one candidate region to the feature map of the image, the method further comprises:

performing pixel correction on each candidate region, to obtain a feature map of the each candidate region,

the feature map of the each candidate region comprising a first feature map, a second feature map, and a third feature map, a size of the first feature map and a size of the second feature map being the same and being less than a size of the third feature map.

3. The method according to claim 2 , wherein the processing the mapped feature map based on a target network for instance segmentation comprises:

performing, for the each candidate region, feature fusion on the first feature map, the second feature map, and the third feature map; and

inputting an obtained fused feature map into the first branch structure of the target network for processing.

4. The method according to claim 3 , wherein the performing feature fusion on the first feature map, the second feature map, and the third feature map comprises:

performing concatenation on the first feature map and the second feature map, to obtain a first intermediate feature map;

processing the first intermediate feature map by using one deconvolutional layer and two convolutional layers in sequence, to obtain a second intermediate feature map; and

performing concatenation on the second intermediate feature map and the third feature map, to obtain a third intermediate feature map, and processing the third intermediate feature map by using three convolutional layers in sequence, to obtain the fused feature map.

5. The method according to claim 3 , wherein the inputting an obtained fused feature map into a first branch structure of the target network for processing comprises:

inputting the fused feature map into a first deconvolutional layer for processing, to obtain a first processed feature map; and

inputting the first processed feature map into a first mask branch comprised in the first branch structure, and obtaining, when an instance comprised in a candidate region is an occluded instance, a non-perspective mask of the instance comprised in the candidate region.

6. The method according to claim 5 , wherein the inputting an obtained fused feature map into a first branch structure of the target network for processing further comprises:

inputting the fused feature map into a second deconvolutional layer for processing, to obtain a second processed feature map; and

inputting the second processed feature map into a second mask branch comprised in the first branch structure, to obtain a perspective mask of the instance comprised in the candidate region.

7. The method according to claim 2 , wherein the method further comprises:

inputting, for the each candidate region, the first feature map into the second branch structure for processing, to obtain the occlusion classification information of the instance comprised in the candidate region.

8. The method according to claim 7 , wherein the inputting the first feature map into the second branch structure for processing, to obtain occlusion classification information of an instance comprised in the candidate region comprises:

determining, based on the second branch structure and the first feature map, whether an occluded area of the instance comprised in the candidate region reaches a target threshold; and

determining, when the occluded area of the instance comprised in the candidate region reaches the target threshold, that the occlusion classification information indicates that the instance comprised in the candidate region is occluded.

9. The method according to claim 7 , wherein the inputting the first feature map into the second branch structure for processing comprises:

inputting the first feature map into a first fully connected layer for processing; and

inputting the first feature map processed in the first fully connected layer into the second branch structure for processing.

10. The method according to claim 2 , wherein the method further comprises:

inputting, for the each candidate region, the second feature map into the third branch structure for processing, to obtain the semantic label and the position information of the instance comprised in the candidate region.

11. The method according to claim 10 , wherein the inputting the second feature map into the third branch structure for processing comprises:

inputting the second feature map into a second fully connected layer for processing;

inputting the second feature map processed in the second fully connected layer into a classification branch comprised in the third branch structure for processing, to obtain classification information;

inputting the second feature map into a third fully connected layer for processing; and

inputting the second feature map processed in the third fully connected layer into a regression branch comprised in the third branch structure for processing, to obtain the position information.

12. The method according to claim 1 , further comprising:

obtaining training sample images, annotation information of each instance in the training sample images comprising at least a semantic label used for indicating a category, a relative occlusion order, and a perspective mask;

determining, based on relative occlusion orders and perspective masks of occluded instances in the training sample images, non-perspective masks of the occluded instances in the training sample images; and

performing model training according to the training sample images annotated with the perspective masks, the non-perspective masks, and semantic labels, to obtain the target network.

13. An image processing apparatus, comprising one or more processors and a memory, the memory storing at least one computer-readable instruction, the at least one computer-readable instruction being loaded and executed by the one or more processors to:

obtain an image;

perform feature extraction on the image, to obtain a feature map of the image;

generate at least one candidate region of the image;

map the at least one candidate region to the feature map of the image, each candidate region comprising one instance; and

process the mapped feature map based on a target network for instance segmentation, indicate an overall structure of an occluded instance in the image by using a perspective mask, and indicate an invisible part of the occluded instance by using a non-perspective mask, the perspective mask and the non-perspective mask representing a relative occlusion relationship of the occluded instance, wherein:

the target network comprises a first branch structure configured to obtain the non-perspective mask and the perspective mask, a second branch structure configured to output occlusion classification information of an instance comprised in the candidate region, and a third branch configured to output a semantic label and position information of the instance in the image; and

the one or more processors are further configured to: fuse feature information output by the second branch structure, feature information output by the third branch structure, and feature information corresponding to the first branch structure to obtain fused feature information; and apply the fused feature information in the first branch structure to obtain the perspective mask and the non-perspective mask.

14. The apparatus according to claim 13 , wherein the one or more processors are further configured to perform pixel correction on each candidate region, to obtain a feature map of the each candidate region, the feature map of the each candidate region comprising a first feature map, a second feature map, and a third feature map, a size of the first feature map and a size of the second feature map being the same and being less than a size of the third feature map.

15. The apparatus according to claim 14 , wherein the one or more processors are further configured to perform, for the each candidate region, feature fusion on the first feature map, the second feature map, and the third feature map; and input an obtained fused feature map into the first branch structure of the target network for processing.

16. The apparatus according to claim 15 , wherein the one or more processors are further configured to perform concatenation on the first feature map and the second feature map, to obtain a first intermediate feature map; process the first intermediate feature map by using one deconvolutional layer and two convolutional layers in sequence, to obtain a second intermediate feature map; and perform concatenation on the second intermediate feature map and the third feature map, to obtain a third intermediate feature map, and process the third intermediate feature map by using three convolutional layers in sequence, to obtain the fused feature map.

17. The apparatus according to claim 15 , wherein the one or more processors are further configured to input the fused feature map into a first deconvolutional layer for processing, to obtain a first processed feature map; and inputting the first processed feature map into a first mask branch comprised in the first branch structure, and obtain, when an instance comprised in a candidate region is an occluded instance, a non-perspective mask of the instance comprised in the candidate region.

18. The apparatus according to claim 14 , wherein the one or more processors are further configured to input, for the each candidate region, the first feature map into the second branch structure for processing, to obtain the occlusion classification information of the instance comprised in the candidate region.

19. The apparatus according to claim 14 , wherein the one or more processors are further configured to input, for the each candidate region, the second feature map into the third branch structure for processing, to obtain the semantic label and the position information of the instance comprised in the candidate region.

20. One or more non-transitory computer-readable storage media,

each storage medium storing at least one computer-readable instruction, the at least one computer-readable instruction being loaded and executed by one or more processors to implement:

obtaining an image, and performing feature extraction on the image, to obtain a feature map of the image;

generating at least one candidate region of the image, and mapping the at least one candidate region to the feature map of the image, each candidate region comprising one instance;

processing the mapped feature map based on a target network for instance segmentation; and

indicating an overall structure of an occluded instance in the image by using a perspective mask, and indicating an invisible part of the occluded instance by using a non-perspective mask, the perspective mask and the non-perspective mask representing a relative occlusion relationship of the occluded instance, wherein:

the target network comprises a first branch structure configured to obtain the non-perspective mask and the perspective mask, a second branch structure configured to output occlusion classification information of an instance comprised in the candidate region, and a third branch configured to output a semantic label and position information of the instance in the image; and

processing the mapped feature map based on the target network for instance segmentation comprises: fusing feature information output by the second branch structure, feature information output by the third branch structure, and feature information corresponding to the first branch structure to obtain fused feature information; and applying the fused feature information in the first branch structure to obtain the perspective mask and the non-perspective mask.