Image classification method and device and style migration model training method and device

Abstract

The embodiment of the invention discloses an image classification method and device and a style migration model training method and device, and relates to the field of deep learning, cloud computing and computer vision in artificial intelligence. The specific implementation scheme comprises the steps of inputting a first style image into a style migration model to obtain a second style image corresponding to the first style image; and inputting the second style image into the image classification model to obtain a classification result of the second style image, wherein the style migration model is obtained by training based on a sample image of a first style and a sample image of a second style, and the image classification model is obtained based on second style sample image training. According to the embodiment of the invention, the image classification accuracy can be improved. The embodiment of the invention can be applied to fundus screening.

Description

technical field [0001] This application relates to the field of artificial intelligence, in particular to the fields of deep learning, cloud computing and computer vision in artificial intelligence. Background technique [0002] In the application of deep learning models, sometimes the sample data of the training set used in the model training process is inconsistent with the data style processed in the actual use of the model, resulting in a significant drop in the accuracy of the actual scene of the model compared with the accuracy of the laboratory . [0003] For example, in recent years, algorithms for automatic diagnosis of fundus diseases based on deep learning on fundus photos have been widely studied and applied. Such algorithms can accurately screen out various fundus diseases and glaucoma under certain conditions. However, due to the sensitivity of the medical industry, fundus photo data are usually not shared. Therefore, the fundus images used for training deep l...

AI Technical Summary

Problems solved by technology

However, due to the sensitivity of the medical industry, fundus photo data are usually not shared. Therefore, the fundus images used for training deep learning models in the laboratory often do not come from the same type of fundus camera as the fundus images acquired in the actual scene, resulting in two types of fundus images. The style of the graph is different; this difference in image style leads to a situation where the accuracy of the deep learning model in the actual use scene is greatly reduced compared with the laboratory accuracy

[0004] In view of the above-mentioned problem that the accuracy of the deep learning model is reduced due to the difference between the training set and the actual scene, it is generally used to add some actual scene samples during model training, and confuse the characteristics of the training set and the actual scene, so as to guide the depth used for classification. The learning model only learns the features and discriminative patterns that are really effective for classification (such as disease diagnosis)

The disadvantage of this method is that it is necessary to add samples from the actual scene when training the classification model (such as a disease discrimination model), making it unable to meet the needs of rapid deployment in multiple scenarios

At the same time, it usually requires the samples in the actual scene to be labeled, and the quality of the annotation in the actual scene in some fields is far inferior to that in the laboratory scene, so this kind of method does not have a good effect on improving the accuracy of the classification model.

Images