Granted Patent B2

US 12,406,476 · App. 18/058,878 · Granted Sep 2, 2025

Inventor: Jun-Dong Chang (Taipei, TW)

Assignee: INSTITUTE FOR INFORMATION INDUSTRY

G06V10/774G06V10/82

Abstract

Claims (60)

1. An image recognition edge device, comprising:

a memory, configured for storing a plurality of instructions; and

a processor, connected to the memory, and configured for executing an object recognition model, wherein the object recognition model comprises a plurality of convolutional layers and a combined fully connection layer, the plurality of convolutional layers are sequentially connected in sequence, and the combined fully connection layer comprises a first fully connection output layer and a second fully connection output layer, wherein the processor accesses the plurality of instructions to perform following operations:

downsampling an input image to generate a downsampled image;

inputting the downsampled image into the object recognition model thereby sequentially generating a plurality of feature tensors through the plurality of convolutional layers;

selecting a part of the plurality of feature tensors to form a first feature tensor pyramid;

selecting another part of the plurality of feature tensors to form a second feature tensor pyramid; and

inputting the first feature tensor pyramid and the second feature tensor pyramid into the combined fully connection layer, generating a first image detection label by the first fully connection output layer based on the first feature tensor pyramid, and generating a second image detection label by the second fully connection output layer based on the second feature tensor pyramid, wherein the first image detection label and the second image detection label respectively include an object identification result corresponding to the input image.

2. The image recognition edge device of claim 1 , wherein the processor is further configured for:

performing feature pyramid processing on the part of the plurality of feature tensors to generate the first feature tensor pyramid; and

performing feature pyramid processing on the other part of the plurality of feature tensors to generate the second feature tensor pyramid.

3. The image recognition edge device of claim 1 , wherein the processor is further configured for:

performing feature pyramid processing on the part of the plurality of feature tensors to generate a plurality of first pyramid feature tensors, and generating the first feature tensor pyramid from the plurality of first pyramid feature tensors according to a plurality of first loss rates generated by the plurality of first pyramid feature tensors; and

performing feature pyramid processing on the other part of the feature tensors to generate a plurality of second pyramid feature tensors, and generating the second feature tensor pyramid from the plurality of second pyramid feature tensors according to a plurality of second loss rates generated by the plurality of second pyramid feature tensors.

4. The image recognition edge device of claim 3 , wherein the part of the plurality of feature tensors comprises a first high-level tensor and a first low-level tensor, an level of a convolutional layer corresponding to the first high-level tensor is less than an level of a convolutional layer corresponding to the first low-level tensor, wherein the processor is further configured for:

performing deconvolution processing on the first high-level tensor to generate a first upsampled tensor;

performing convolution processing on the first low-level tensor to generate a first convolution tensor; and

generating one of the plurality of first pyramid feature tensors according to the first upsampled tensor and the first convolution tensor.

5. The image recognition edge device of claim 4 , wherein the processor is further configured for:

performing element-wise addition operation on the first upsampled tensor and the first convolution tensor to generate the one of the plurality of first pyramid feature tensors.

6. The image recognition edge device of claim 3 , wherein the other part of the plurality of feature tensors comprises a second high-level tensor and a second low-level tensor, a level of a convolutional layer corresponding to the second high-level tensor is less than an level of a convolutional layer corresponding to the second low-level tensor, wherein the processor is further configured for:

performing deconvolution processing on the second high-level tensor to generate a second upsampled tensor;

performing convolution processing on the second low-level tensor to generate a second convolution tensor; and

generating the one of the plurality of second pyramid feature tensors according to the second upsampled tensor and the second convolution tensor.

7. The image recognition edge device of claim 6 , wherein the processor is further configured for:

performing element-wise addition operation on the second upsampled tensor and the second convolution tensor to generate the one of the plurality of second pyramid feature tensors.

8. The image recognition edge device of claim 1 , wherein the processor is further configured for:

selecting the part of the plurality of feature tensors through experiments of a first identification task.

9. The image recognition edge device of claim 8 , wherein the processor is further configured for:

selecting the other part of the plurality of feature tensors through experiments of a second identification task, wherein the second identification task is different from the first identification task.

10. The image recognition edge device of claim 1 , wherein the first image detection label and the second image detection label respectively correspond to different identification tasks.

11. An image recognition edge method for an image recognition edge device, comprising:

downsampling an input image to generate a downsampled image;

inputting the downsampled image into an object recognition model thereby sequentially generating a plurality of feature tensors through a plurality of convolutional layers, wherein the object recognition model comprises the plurality of convolutional layers and a combined fully connection layer, the plurality of convolutional layers are sequentially connected in sequence, and the combined fully connection layer comprises a first fully connection output layer and a second fully connection output layer;

selecting a part of the plurality of feature tensors to form a first feature tensor pyramid;

selecting another part of the plurality of feature tensors to form a second feature tensor pyramid; and

inputting the first feature tensor pyramid and the second feature tensor pyramid into the combined fully connection layer, generating a first image detection label by the first fully connection output layer based on the first feature tensor pyramid, and generating a second image detection label by the second fully connection output layer based on the second feature tensor pyramid, wherein the first image detection label and the second image detection label respectively include an object identification result corresponding to the input image.

12. The image recognition edge method of claim 11 , wherein the step of selecting the part of the plurality of feature tensors to form the first feature tensor pyramid comprises:

performing feature pyramid processing on the part of the plurality of feature tensors to generate the first feature tensor pyramid; and

performing feature pyramid processing on the other part of the plurality of feature tensors to generate the second feature tensor pyramid.

13. The image recognition edge method of claim 11 , wherein the step of selecting the part of the plurality of feature tensors to form the first feature tensor pyramid comprises:

performing feature pyramid processing on the part of the plurality of feature tensors to generate a plurality of first pyramid feature tensors, and generating the first feature tensor pyramid from the plurality of first pyramid feature tensors according to a plurality of first loss rates generated by the plurality of first pyramid feature tensors; and

performing feature pyramid processing on the other part of the feature tensors to generate a plurality of second pyramid feature tensors, and generating the second feature tensor pyramid from the plurality of second pyramid feature tensors according to a plurality of second loss rates generated by the plurality of second pyramid feature tensors.

14. The image recognition edge method of claim 13 , wherein the part of the plurality of feature tensors comprises a first high-level tensor and a first low-level tensor, an level of a convolutional layer corresponding to the first high-level tensor is less than an level of a convolutional layer corresponding to the first low-level tensor, wherein the step of performing the feature pyramid processing on the part of the plurality of feature tensors to generate the plurality of first pyramid feature tensors comprises:

performing deconvolution processing on the first high-level tensor to generate a first upsampled tensor;

performing convolution processing on the first low-level tensor to generate a first convolution tensor; and

generating one of the plurality of first pyramid feature tensors according to the first upsampled tensor and the first convolution tensor.

15. The image recognition edge method of claim 14 , wherein the step of generating the one of the plurality of first pyramid feature tensors according to the first upsampled tensor and the first convolution tensor comprises:

performing element-wise addition operation on the first upsampled tensor and the first convolution tensor to generate the one of the plurality of first pyramid feature tensors.

16. The image recognition edge method of claim 13 , wherein the other part of the plurality of feature tensors comprises a second high-level tensor and a second low-level tensor, a level of a convolutional layer corresponding to the second high-level tensor is less than an level of a convolutional layer corresponding to the second low-level tensor, wherein the step of performing the feature pyramid processing on the other part of the feature tensors to generate the plurality of second pyramid feature tensors comprises:

performing deconvolution processing on the second high-level tensor to generate a second upsampled tensor;

performing convolution processing on the second low-level tensor to generate a second convolution tensor; and

generating the one of the plurality of second pyramid feature tensors according to the second upsampled tensor and the second convolution tensor.

17. The image recognition edge method of claim 16 , wherein the step of generating the one of the plurality of second pyramid feature tensors according to the second upsampled tensor and the second convolution tensor comprises:

performing element-wise addition operation on the second upsampled tensor and the second convolution tensor to generate the one of the plurality of second pyramid feature tensors.

18. The image recognition edge method of claim 11 , further comprising:

selecting the part of the plurality of feature tensors through experiments of a first identification task.

19. The image recognition edge method of claim 18 , further comprising:

selecting the other part of the plurality of feature tensors through experiments of a second identification task, wherein the second identification task is different from the first identification task.

20. The image recognition edge method of claim 11 , wherein the first image detection label and the second image detection label respectively correspond to different identification tasks.

Assignments (1)

Priority Claims (1)

Continuity (1)

References Cited (10)