Method and terminal for recognizing object node in image, and computer-readable storage medium

1. A method for recognizing object nodes in an image, executed by a computer device, comprising:

performing image detection on an original image, to determine a target image region that comprises a target image object, the original image being a two-dimensional image, by:

invoking a preset second region detection model to perform convolution calculation on the original image, to obtain a third image feature;

performing R times of downsampling processing and S sets of convolution calculation processing on the third image feature to obtain a fourth image feature, the R times of downsampling processing comprising downsampling the third image feature and downsampling an image feature obtained after each set of convolution calculation processing, and R and S each being a positive integer;

performing convolution calculation processing on the fourth image feature to obtain an intermediate image feature;

performing dimension conversion on the fourth image feature to obtain a first feature vector, and performing dimension conversion on the intermediate image feature to obtain a second feature vector;

performing feature fusion on the first feature vector and the second feature vector to obtain quadruple data; and

determining a region position in the original image according to a first offset and a second offset in the quadruple data, determining a region size according to a width value and a height value in the quadruple data, and obtaining the target image region according to the region position and the region size;

performing image cropping on the original image according to the target image region, to obtain a subimage comprising the target image object; and

performing node recognition on the subimage based on a preset node recognition model, to obtain image position information of each node of the target image object in the original image.

2. The method according to claim 1 , wherein the performing image detection on an original image, to determine a target image region comprises:

invoking a preset first region detection model to perform image feature recognition on the original image, to obtain a region probability matrix;

and

determining the target image region further according to the region probability matrix.

3. The method according to claim 2 , wherein the invoking a preset first region detection model to perform image feature recognition on the original image, to obtain a region probability matrix comprises:

invoking the preset first region detection model to perform convolution calculation on the original image, to obtain a first image feature;

performing downsampling processing on the first image feature, to obtain a second image feature;

performing upsampling processing on the second image feature, to obtain a target image feature; and

performing deconvolution calculation on the target image feature, to obtain the region probability matrix.

4. The method according to claim 3 , wherein:

performing the downsampling processing comprises N times of downsampling processing and M sets of convolution calculation processing, N and M both being positive integers; and

performing the upsampling processing comprises N times of upsampling processing and M sets of deconvolution calculation processing.

5. The method according to claim 1 , wherein the processing the fourth image feature further comprises:

performing downsampling processing on the fourth image feature, and performing convolution calculation processing on an image feature obtained after the downsampling processing.

6. The method according to claim 1 , wherein the performing node recognition on the subimage based on a preset node recognition model comprises:

invoking the preset node recognition model to perform node recognition on the subimage, to obtain node probability matrices of P channels; and

obtaining the image position information of each node of the target image object according to the node probability matrices of the P channels.

7. The method according to claim 6 , wherein the performing node recognition on the subimage, to obtain node probability matrices of P channels comprises:

performing convolution calculation on the subimage, to obtain a convolutional image feature;

performing node recognition processing on the convolutional image feature, to obtain a first node recognition feature; and

obtaining the node probability matrices of the P channels according to the first node recognition feature.

8. The method according to claim 7 , wherein the obtaining the node probability matrices of the P channels according to the first node recognition feature comprises:

obtaining a second node recognition feature according to the first node recognition feature; and

performing node information adjustment processing on the first node recognition feature and the second node recognition feature, to obtain the node probability matrices of the P channels.

9. The method according to claim 1 , wherein the performing image cropping on the original image according to the target image region, to obtain a subimage comprising the target image object comprises:

performing image cropping on the original image according to the target image region, to obtain a cropped image, the cropped image comprising the target image object; and

adjusting a resolution of the cropped image to obtain the subimage.

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

obtaining a node connection image of the target image object according to the image position information of each node of the target image object; and

superposing the node connection image onto the original image, to obtain a superposed image.

11. The method according to claim 1 , wherein the performing dimension conversion on the fourth image feature, to obtain a first feature vector comprises:

performing scaling processing on the fourth image feature, to obtain a scaled fourth image feature, a resolution of the scaled fourth image feature being the same as a resolution of the intermediate image feature; and

performing dimension conversion on the scaled fourth image feature, to obtain the first feature vector.

12. An intelligent terminal, comprising a processor and a memory coupled to the processor, the processor being configured to:

perform image detection on an original image, to determine a target image region that comprises a target image object, the original image being a two-dimensional image, by:

invoking a preset second region detection model to perform convolution calculation on the original image, to obtain a third image feature;

performing R times of downsampling processing and S sets of convolution calculation processing on the third image feature to obtain a fourth image feature, the R times of downsampling processing comprising downsampling the third image feature and downsampling an image feature obtained after each set of convolution calculation processing, and R and S each being a positive integer;

performing convolution calculation processing on the fourth image feature to obtain an intermediate image feature;

performing dimension conversion on the fourth image feature to obtain a first feature vector, and performing dimension conversion on the intermediate image feature to obtain a second feature vector;

performing feature fusion on the first feature vector and the second feature vector to obtain quadruple data; and

determining a region position in the original image according to a first offset and a second offset in the quadruple data, determining a region size according to a width value and a height value in the quadruple data, and obtaining the target image region according to the region position and the region size;

perform image cropping on the original image according to the target image region, to obtain a subimage comprising the target image object; and

perform node recognition on the subimage based on a preset node recognition model, to obtain image position information of each node of the target image object in the original image.

13. The intelligent terminal according to claim 12 , wherein the processor is further configured to:

invoke a preset first region detection model to perform image feature recognition on the original image, to obtain a region probability matrix;

and

determine the target image region further according to the region probability matrix.

14. The intelligent terminal according to claim 13 , wherein the processor is further configured to:

invoke the preset first region detection model to perform convolution calculation on the original image, to obtain a first image feature;

perform downsampling processing on the first image feature, to obtain a second image feature;

perform upsampling processing on the second image feature, to obtain a target image feature; and

perform deconvolution calculation on the target image feature to obtain a region probability matrix.

15. The intelligent terminal according to claim 14 , wherein the processor is further configured to:

perform the downsampling processing comprises N times of downsampling processing and M sets of convolution calculation processing; and

perform the upsampling and fusion processing comprises N times of upsampling processing.

16. The intelligent terminal according to claim 12 , wherein the processor is further configured to:

obtain a node connection image of the target image object according to the image position information of each node of the target image object; and

superpose the node connection image onto the original image, to obtain a superposed image.

17. The intelligent terminal according to claim 12 , wherein the processing the fourth image feature further comprises:

performing downsampling processing on the fourth image feature, and performing convolution calculation processing on an image feature obtained after the downsampling processing.

18. The intelligent terminal according to claim 12 , wherein the performing node recognition on the subimage based on a preset node recognition model comprises:

invoking the preset node recognition model to perform node recognition on the subimage, to obtain node probability matrices of P channels; and

obtaining the image position information of each node of the target image object according to the node probability matrices of the P channels.

19. The intelligent terminal according to claim 12 , wherein the performing image cropping on the original image according to the target image region, to obtain a subimage comprising the target image object comprises:

performing image cropping on the original image according to the target image region, to obtain a cropped image, the cropped image comprising the target image object; and

adjusting a resolution of the cropped image to obtain the subimage.

20. A non-transitory computer-readable storage medium, storing a computer-readable instruction, the computer-readable instruction, when executed by one or more processors, causing the one or more processors to perform:

performing image detection on an original image, to determine a target image region that comprises a target image object, the original image being a two-dimensional image, by:

invoking a preset second region detection model to perform convolution calculation on the original image, to obtain a third image feature;

performing R times of downsampling processing and S sets of convolution calculation processing on the third image feature to obtain a fourth image feature, the R times of downsampling processing comprising downsampling the third image feature and downsampling an image feature obtained after each set of convolution calculation processing, and R and S each being a positive integer;

performing convolution calculation processing on the fourth image feature to obtain an intermediate image feature;

performing dimension conversion on the fourth image feature to obtain a first feature vector, and performing dimension conversion on the intermediate image feature to obtain a second feature vector;

performing feature fusion on the first feature vector and the second feature vector to obtain quadruple data; and

determining a region position in the original image according to a first offset and a second offset in the quadruple data, determining a region size according to a width value and a height value in the quadruple data, and obtaining the target image region according to the region position and the region size;

performing image cropping on the original image according to the target image region, to obtain a subimage comprising the target image object; and

performing node recognition on the subimage based on a preset node recognition model, to obtain image position information of each node of the target image object in the original image.