Instance segmentation method and system for enhanced image, and device and medium

1. A method for enhanced image instance segmentation, comprising:

setting an auxiliary training network with an identical architecture to a main training network in the main training network, and performing data enhancement on an original image to obtain an enhanced image;

inputting the original image into the main training network, and inputting the enhanced image into the auxiliary training network;

determining whether an intersection over union value of a second prediction bounding box generated by the auxiliary training network and a target bounding box is greater than an intersection over union value of a first prediction bounding box generated by the main training network and the target bounding box; and

replacing the intersection over union value of the first prediction bounding box and the target bounding box with the intersection over union value of the second prediction bounding box and the target bounding box in response to the intersection over union value of the second prediction bounding box and the target bounding box being greater than the intersection over union value of the first prediction bounding box and the target bounding box, so as to conduct image instance segmentation;

wherein performing data enhancement on an original image to obtain an enhanced image comprises:

displacing a coordinate bounding box of the original image by a preset distance.

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

generating a multi-scale first semantic feature map based on the main training network, generating a multi-scale second semantic feature map based on the auxiliary training network, and exchanging the first semantic feature map and the second semantic feature map of the same scales with a predetermined probability.

3. The method according to claim 2 , further comprising:

adding a main branch loss of the main training network and an auxiliary branch loss of the auxiliary training network to the image instance segmentation.

4. The method according to claim 3 , wherein adding a main branch loss of the main training network and an auxiliary branch loss of the auxiliary training network to the image instance segmentation comprises:

combining an initial coordinate bounding box with the first semantic feature map to obtain a first classification feature map, combining the initial coordinate bounding box with the second semantic feature map to obtain a second classification feature map, and acquiring a divergence loss according to the first classification feature map and the second classification feature map.

5. The method according to claim 4 , wherein acquiring a divergence loss according to the first classification feature map and the second classification feature map comprises:

calculating a sub-divergence loss of the first classification feature map and the second classification feature map of the same scales, and adding up the sub-divergence losses of all the scales to obtain the divergence loss.

6. The method according to claim 5 , wherein an equation for calculating the KL divergence, denoted as D KL , is as follows:

D

KL

(

p

⁢



q

)

=

∑

N

i

=

1

p

⁡

(

x

i

)

⁢

log

⁡

(

p

⁡

(

x

i

)

q

⁡

(

x

i

)

)

where p and q represent the feature maps of the main branch and the auxiliary branch, respectively, and x i represents an i-th scale.

7. The method according to claim 4 , further comprising:

inputting the first classification feature map or the second classification feature map into a classification network to determine a type of a candidate bounding box;

inputting the first classification feature map or the second classification feature map into a regression network to locate the candidate bounding box; and

inputting the first classification feature map or the second classification feature map into a segmentation network to perform pixel-level segmentation on instances in the candidate bounding box.

8. The method according to claim 3 , wherein adding a main branch loss of the main training network and an auxiliary branch loss of the auxiliary training network to the image instance segmentation comprises:

obtaining a main regression loss and an auxiliary regression loss according to information of a distance and scale of a coordinate bounding box.

9. The method according to claim 3 , wherein the main branch loss and the auxiliary branch loss are both a classification loss and a regression loss; and

the classification loss uses a cross-entropy loss, and the regression loss uses an intersection over union loss.

10. The method according to claim 9 , wherein an equation of the intersection over union loss is as follows:

Ne

⁢

w

Loss

=

1

-

·

Intersection

⁡

(

p

,

g

)

Union

⁡

(

p

,

g

)

+

d

⁡

(

p

c

,

g

c

)

c

where Intersection represents an intersection of the prediction bounding box and the ground-truth bounding box, Union represents a union of the prediction bounding box and the ground-truth bounding box, P c is a coordinate of a center point of the prediction bounding box, g c is a coordinate of a center point of the ground-truth bounding box, and c is a diagonal distance of a minimum closure of the prediction bounding box and the ground-truth bounding box.

11. The method according to claim 1 , wherein horizontal flipping is used for the data enhancement during training.

12. A computer device, comprising:

at least one processor; and

a memory having processor-executable computer instructions stored thereon, causes the processor to:

set an auxiliary training network with an identical architecture to a main training network in the main training network, and perform data enhancement on an original image to obtain an enhanced image;

input the original image into the main training network, and input the enhanced image into the auxiliary training network;

determine whether an intersection over union value of a second prediction bounding box generated by the auxiliary training network and a target bounding box is greater than an intersection over union value of a first prediction bounding box generated by the main training network and the target bounding box; and

replace the intersection over union value of the first prediction bounding box and the target bounding box with the intersection over union value of the second prediction bounding box and the target bounding box in response to the intersection over union value of the second prediction bounding box and the target bounding box being greater than the intersection over union value of the first prediction bounding box and the target bounding box, so as to conduct image instance segmentation; and

further causing the processor to:

generate a multi-scale first semantic feature map based on the main training network, generate a multi-scale second semantic feature map based on the auxiliary training network, and exchange the first semantic feature map and the second semantic feature map of the same scales with a predetermined probability.

13. The computer device according to claim 12 , wherein perform data enhancement on an original image to obtain an enhanced image comprises:

displacing a coordinate bounding box of the original image by a preset distance.

14. The computer device according to claim 12 , further causing the processor to:

add a main branch loss of the main training network and an auxiliary branch loss of the auxiliary training network to the image instance segmentation.

15. A computer-readable storage medium, storing a computer program thereon, when executed by a processor, causes the processor to:

set an auxiliary training network with an identical architecture to a main training network in the main training network, and perform data enhancement on an original image to obtain an enhanced image;

input the original image into the main training network, and input the enhanced image into the auxiliary training network;

determine whether an intersection over union value of a second prediction bounding box generated by the auxiliary training network and a target bounding box is greater than an intersection over union value of a first prediction bounding box generated by the main training network and the target bounding box; and

replace the intersection over union value of the first prediction bounding box and the target bounding box with the intersection over union value of the second prediction bounding box and the target bounding box in response to the intersection over union value of the second prediction bounding box and the target bounding box being greater than the intersection over union value of the first prediction bounding box and the target bounding box, so as to conduct image instance segmentation; and

further causing the processor to:

generate a multi-scale first semantic feature map based on the main training network, generate a multi-scale second semantic feature map based on the auxiliary training network, and exchange the first semantic feature map and the second semantic feature map of the same scales with a predetermined probability.

16. The computer-readable storage medium according to claim 15 , wherein perform data enhancement on an original image to obtain an enhanced image comprises:

displacing a coordinate bounding box of the original image by a preset distance.

17. The computer-readable storage medium according to claim 15 , further causing the processor to:

add a main branch loss of the main training network and an auxiliary branch loss of the auxiliary training network to the image instance segmentation.