Method and electronic apparatus for processing image and training image tag classification model

1. An image processing method applied to an electronic apparatus, comprising:

acquiring an image;

determining a feature map of the image based on an image tag classification model, wherein the image tag classification model comprises a plurality of classification tasks; and

determining tags corresponding to the feature map based on the classification tasks, wherein each of the tags comprises a probability value,

wherein said determining tags corresponding to the feature map based on the classification tasks comprises:

determining an intermediate feature map of the feature map by performing a dimension reducing processing on the feature map through the image tag classification model; and

determining a feature vector of the intermediate feature map under each of the classification tasks by performing an average pooling on the intermediate feature map through the image tag classification model, wherein each feature vector comprises a plurality of elements, and each of the elements comprises a probability value that a tag of the feature map is a tag of a corresponding classification task.

2. The method according to claim 1 , wherein each of the classification tasks is one of a multi-tag classification task and a single-tag classification task.

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

selecting tags of the image from tags of the classification tasks according to the probability value of each of the tags.

4. The method according to claim 3 , wherein said selecting tags of the image from tags of the classification tasks according to probability values comprises:

determining first tags, wherein each probability value of the first tags is greater than a probability threshold; and

determining the tags of the image from the tags of the classification tasks based on the first tags.

5. The method according to claim 1 , wherein the image tag classification model comprises a convolutional neural network and a classifier; wherein the convolutional neural network is used to acquire the feature map of the image, and the classifier comprises the classification tasks.

6. The method according to claim 1 , wherein the image tag classification model comprises a convolutional neural network and a plurality of classifiers; wherein the convolutional neural network is used to acquire the feature map of the image, and each of the plurality of classifiers comprises one classification task.

7. A training method for an image tag classification model, wherein the method is applied to an electronic apparatus, and the method comprises:

acquiring sample images corresponding to each of a plurality of classification tasks, wherein each of the sample images has tag data, the tag data comprises a classification task to which the sample image belongs and at least one target tag corresponding to the sample image under the classification task to which the sample image belongs, and a probability value corresponding to each of the at least one target tag is 1 and each of probability values corresponding to other tags is 0 in a classification result of the sample image under the classification task to which the sample image belongs; and

training the image tag classification model based on the sample images corresponding to each of a plurality of classification tasks,

wherein said training the image tag classification model based on the sample images corresponding to each of a plurality of classification tasks comprises:

determining a first feature map of each of the sample images through the image tag classification model, wherein each of the sample images corresponds to the at least one target tag and a probability value corresponding to each of the at least one target tag;

determining a second feature map of the first feature map by performing a dimension reducing processing on the first feature map through the image tag classification model; and

determining a first feature vector of the second feature map under each of the classification tasks by performing an average pooling on the second feature map through the image tag classification model, wherein each first feature vector comprises a plurality of elements, and each of the elements comprises a probability value that a tag of the first feature map is a tag of a corresponding classification task.

8. The method according to claim 7 , wherein said acquiring sample images corresponding to each of a plurality of classification tasks comprises:

for each sample image, determining the classification task to which the sample image belongs and the at least one target tag corresponding to the sample image under the classification task to which the sample image belongs;

setting a probability value at a position of each of the at least one target tag to 1 in a classification result of the sample image under the classification task to which the sample image belongs; and

setting a probability value at a position of each of tags other than the at least one target tag to 0 in a classification result of the sample image under each of the classification tasks, wherein each tag corresponds to a position in a classification result of a classification task to which the tag belongs.

9. The method according to claim 7 , wherein said training the image tag classification model based on the sample images corresponding to each of a plurality of classification tasks comprises:

acquiring a prediction accuracy about each of the sample images based on the tags corresponding to each of the sample images, the probability value corresponding to each of the tags, the at least one target tag corresponding to each of the sample images and the probability value corresponding to each of the at least one target tag; wherein each of the classification tasks corresponds to a prediction accuracy, and the prediction accuracy about each of the sample images is determined based on prediction accuracies obtained by each of the sample image under the classification tasks; and

updating model parameters of the image tag classification model according to the prediction accuracy.

10. A non-transitory computer-readable storage medium on which one or more programs is stored, wherein the one or more programs, when executed by a processor, causes the processor to perform an image processing method comprising:

acquiring an image;

determining a feature map of the image based on an image tag classification model, wherein the image tag classification model comprises a plurality of classification tasks; and

determining tags corresponding to the feature map based on the classification tasks, wherein each of the tags comprises a probability value,

wherein said determining tags corresponding to the feature map based on the classification tasks comprises:

determining an intermediate feature map of the feature map by performing a dimension reducing processing on the feature map through the image tag classification model; and

determining a feature vector of the intermediate feature map under each of the classification tasks by performing an average pooling on the intermediate feature map through the image tag classification model, wherein each feature vector comprises a plurality of elements, and each of the elements comprises a probability value that a tag of the feature map is a tag of a corresponding classification task.

11. The non-transitory computer-readable storage medium according to claim 10 , wherein each of the classification tasks is one of a multi-tag classification task and a single-tag classification task.

12. The non-transitory computer-readable storage medium according to claim 10 , wherein the method further comprises:

selecting tags of the image from tags of the classification tasks according to the probability value of each of the tags.

13. The non-transitory computer-readable storage medium according to claim 12 , wherein said selecting tags of the image from tags of the classification tasks according to probability values comprises:

determining first tags, wherein each probability value of the first tags is greater than a probability threshold; and

determining the tags of the image from the tags of the classification tasks based on the first tags.

14. The non-transitory computer-readable storage medium according to claim 10 , wherein the image tag classification model comprises a convolutional neural network and a classifier; wherein the convolutional neural network is used to acquire the feature map of the image, and the classifier comprises the classification tasks.

15. The non-transitory computer-readable storage medium according to claim 10 , wherein the image tag classification model comprises a convolutional neural network and a plurality of classifiers; wherein the convolutional neural network is used to acquire the feature map of the image, and each of the plurality of classifiers comprises one classification task.