Image recognition model training method and apparatus, and image recognition method, apparatus, and system

1. An image recognition method performed by a computing device, the method comprising:

obtaining a to-be-recognized image;

extracting image feature information of the to-be-recognized image; and

obtaining a lesion category recognition result of the to-be-recognized image by using the image feature information of the to-be-recognized image as an input parameter of a preset image recognition model, the image recognition model being trained by using a training image sample set comprising at least one strong-label training image sample and at least one weak-label training image sample, to determine the lesion category recognition result; and the strong-label training image sample representing an image sample having strong-label information, comprising at least annotation information of a lesion category and a lesion position in the strong-label training image sample, and the weak-label training image sample representing an image sample having weak-label information comprising only annotation information of a lesion category.

2. The method according to claim 1 , wherein the obtaining a lesion category recognition result of the to-be-recognized image comprises:

determining, by using the image recognition model, and based on a relationship between a lesion category and image block feature information of a lesion position that is determined from the strong-label training image sample and the strong-label information, whether each image block in the image feature information of the to-be-recognized image belongs to the lesion category; and

determining whether the to-be-recognized image belongs to the lesion category according to whether the image block belongs to the lesion category.

3. The method according to claim 2 , wherein the obtaining a lesion category recognition result of the to-be-recognized image further comprises:

determining, by using the image recognition model, and based on a relationship between a lesion category and overall image feature information that is determined from the weak-label training image sample and the weak-label information, whether the image feature information of the to-be-recognized image belongs to the lesion category, the weak-label information comprising only annotation information of the lesion category; and

determining whether the to-be-recognized image belongs to the lesion category according to whether the image block belongs to the lesion category and whether the image feature information of the to-be-recognized image belongs to the lesion category.

4. The method according to claim 1 , wherein the preset image recognition model is generated by:

obtaining a training image sample set, the training image sample set comprising at least one strong-label training image sample, the strong-label training image sample being an image sample having strong-label information, and the strong-label information comprising at least annotation information of a lesion category and a lesion position in the image sample;

extracting image feature information of each image sample in the training image sample set;

marking image feature information belonging to each preset lesion category based on the image feature information of the image sample and corresponding strong-label information; and

training an image recognition model according to a mark result until a strong supervision objective function of the image recognition model converges, to obtain the preset image recognition model, the strong supervision objective function being a loss function of a recognized lesion category and a lesion category in the strong-label information.

5. The method according to claim 4 , the method further comprises:

marking the image feature information belonging to the preset lesion category according to the image feature information of the image sample and corresponding strong-label information or weak-label information; and

training the image recognition model according to the mark result until a total objective function of the image recognition model converges, to obtain the preset image recognition model, the total objective function being a total loss function of the strong supervision objective function and a weak supervision objective function, and the weak supervision objective function being a loss function of the recognized lesion category and a lesion category in the weak-label information.

6. The method according to claim 5 , wherein the extracting image feature information of the image sample in the training image sample set comprises:

inputting the each image sample in the training image sample set to a neural network; and

obtaining image feature information of a specified dimension that is outputted after feature extraction is performed on the each image sample based on the neural network.

7. The method according to claim 6 , wherein the image feature information is of a P*P*C dimension, P is a specified value, P*P represents that the image sample is equally divided into P*P image blocks horizontally and vertically, and C is a quantity of preset lesion categories.

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

marking the image feature information belonging to the preset lesion category according to the image feature information of the image sample and the corresponding strong-label information;

determining the strong supervision objective function according to the mark result; and

optimizing the strong supervision objective function until the strong supervision objective function converges.

9. The method according to claim 5 , further comprising:

marking the image feature information belonging to the preset lesion category according to the image feature information of the image sample and the corresponding strong-label information, and determining the strong supervision objective function according to the mark result in a case that the image sample is the strong-label training image sample; or

marking the image feature information belonging to the preset lesion category according to the image feature information of the image sample and the corresponding weak-label information, and determining the weak supervision objective function according to the mark result in a case that the image sample is the weak-label training image sample;

determining the total objective function according to the strong supervision objective function and the weak supervision objective function; and

optimizing the total objective function until the total objective function converges.

10. A computing device comprising a memory, a processor, and a plurality of computer programs stored in the memory and executable on the processor to perform a plurality of operations including:

obtaining a to-be-recognized image;

extracting image feature information of the to-be-recognized image; and

obtaining a lesion category recognition result of the to-be-recognized image by using the image feature information of the to-be-recognized image as an input parameter of a preset image recognition model, the image recognition model being trained by using a training image sample set comprising at least one strong-label training image sample and at least one weak-label training image sample, to determine the lesion category recognition result; and the strong-label training image sample representing an image sample having strong-label information, comprising at least annotation information of a lesion category and a lesion position in the strong-label training image sample, and the weak-label training image sample representing an image sample having weak-label information comprising only annotation information of a lesion category.

11. The computing device according to claim 10 , wherein the obtaining a lesion category recognition result of the to-be-recognized image comprises:

determining, by using the image recognition model, and based on a relationship between a lesion category and image block feature information of a lesion position that is determined from the strong-label training image sample and the strong-label information, whether each image block in the image feature information of the to-be-recognized image belongs to the lesion category; and

determining whether the to-be-recognized image belongs to the lesion category according to whether the image block belongs to the lesion category.

12. The computing device according to claim 11 , wherein the obtaining a lesion category recognition result of the to-be-recognized image further comprises:

determining, by using the image recognition model, and based on a relationship between a lesion category and overall image feature information that is determined from the weak-label training image sample and the weak-label information, whether the image feature information of the to-be-recognized image belongs to the lesion category, the weak-label information comprising only annotation information of the lesion category; and

determining whether the to-be-recognized image belongs to the lesion category according to whether the image block belongs to the lesion category and whether the image feature information of the to-be-recognized image belongs to the lesion category.

13. The computing device according to claim 10 , wherein the preset image recognition model is generated by:

obtaining a training image sample set, the training image sample set comprising at least one strong-label training image sample, the strong-label training image sample being an image sample having strong-label information, and the strong-label information comprising at least annotation information of a lesion category and a lesion position in the image sample;

extracting image feature information of each image sample in the training image sample set;

marking image feature information belonging to each preset lesion category based on the image feature information of the image sample and corresponding strong-label information; and

training an image recognition model according to a mark result until a strong supervision objective function of the image recognition model converges, to obtain the preset image recognition model, the strong supervision objective function being a loss function of a recognized lesion category and a lesion category in the strong-label information.

14. The computing device according to claim 13 , wherein the plurality of operations further comprise:

marking the image feature information belonging to the preset lesion category according to the image feature information of the image sample and corresponding strong-label information or weak-label information; and

training the image recognition model according to the mark result until a total objective function of the image recognition model converges, to obtain the preset image recognition model, the total objective function being a total loss function of the strong supervision objective function and a weak supervision objective function, and the weak supervision objective function being a loss function of the recognized lesion category and a lesion category in the weak-label information.

15. The computing device according to claim 14 , wherein the extracting image feature information of each image sample in the training image sample set comprises:

inputting the each image sample in the training image sample set to a neural network; and

obtaining image feature information of a specified dimension that is outputted after feature extraction is performed on the each image sample based on the neural network.

16. The computing device according to claim 13 , wherein the plurality of operations further comprise:

marking the image feature information belonging to the preset lesion category according to the image feature information of the image sample and the corresponding strong-label information;

determining the strong supervision objective function according to the mark result; and

optimizing the strong supervision objective function until the strong supervision objective function converges.

17. The computing device according to claim 14 , wherein the plurality of operations further comprise:

marking the image feature information belonging to the preset lesion category according to the image feature information of the image sample and the corresponding strong-label information, and determining the strong supervision objective function according to the mark result in a case that the image sample is the strong-label training image sample; or

marking the image feature information belonging to the preset lesion category according to the image feature information of the image sample and the corresponding weak-label information, and determining the weak supervision objective function according to the mark result in a case that the image sample is the weak-label training image sample;

determining the total objective function according to the strong supervision objective function and the weak supervision objective function; and

optimizing the total objective function until the total objective function converges.

18. A non-transitory computer-readable storage medium storing a plurality of computer programs that, when executed by a processor of a computing device, cause the computing device to perform a plurality of operations including:

obtaining a to-be-recognized image;

extracting image feature information of the to-be-recognized image; and

obtaining a lesion category recognition result of the to-be-recognized image by using the image feature information of the to-be-recognized image as an input parameter of a preset image recognition model, the image recognition model being trained by using a training image sample set comprising at least one strong-label training image sample and at least one weak-label training image sample, to determine the lesion category recognition result; and the strong-label training image sample representing an image sample having strong-label information comprising at least annotation information of a lesion category and a lesion position in the strong-label training image sample, and the weak-label training image sample representing an image sample having weak-label information comprising only annotation information of a lesion category.

19. The non-transitory computer-readable storage medium according to claim 18 , wherein the obtaining a lesion category recognition result of the to-be-recognized image comprises:

determining, by using the image recognition model, and based on a relationship between a lesion category and image block feature information of a lesion position that is determined from the strong-label training image sample and the strong-label information, whether each image block in the image feature information of the to-be-recognized image belongs to the lesion category; and

determining whether the to-be-recognized image belongs to the lesion category according to whether the image block belongs to the lesion category.

20. The non-transitory computer-readable storage medium according to claim 19 , wherein the obtaining a lesion category recognition result of the to-be-recognized image further comprises:

determining, by using the image recognition model, and based on a relationship between a lesion category and overall image feature information that is determined from the weak-label training image sample and the weak-label information, whether the image feature information of the to-be-recognized image belongs to the lesion category, the weak-label information comprising only annotation information of the lesion category; and

determining whether the to-be-recognized image belongs to the lesion category according to whether the image block belongs to the lesion category and whether the image feature information of the to-be-recognized image belongs to the lesion category.