A device for grading sugar network features in fundus images based on attention mechanism and feature fusion

Abstract

The invention discloses a device for grading sugar net features in fundus images based on attention mechanism and feature fusion, including: a feature detection and classification network module, which is used to classify first-level sugar net features and second-level sugar net features in input sample fundus pictures Carry out extraction, and output the subdivided classification feature map extracted from the first-level sugar net feature and the second-level sugar net feature; the original image classification network module is used to perform the third-level sugar net feature and the fourth-level sugar net feature in the input sample fundus image. Extract and output the rough classification feature map of the 3-level sugar network feature and the 4-level sugar network feature extraction; the attention mechanism and feature fusion module, this module uses the attention mechanism to fine-tune the feature map and the original image output by the feature detection network module The coarse classification feature map output by the classification network module is subjected to feature fusion, and the output is the predicted probability of the sugar network feature level of the input sample image. The device can achieve 81.33% of classification evaluation index Kappa while ensuring relatively fast speed.

Description

technical field [0001] The invention belongs to the technical field of image processing, and in particular relates to a device for grading sugar net features in fundus images based on attention mechanism and feature fusion. Background technique [0002] Most of the existing deep learning methods use the original image or the image after simple data enhancement to classify the features of the sugar network (diabetic retina) in the fundus image. First, the input image in the training data is passed into the neural network composed of a series of convolutional structures and fully connected structures, then the trained neural network parameters are saved, and finally the test images in the test set are processed using the trained neural network. Prediction, to get the predicted probabilities at the sugarnet feature level for each test image. In order to improve the receptive field of the neural network to the input image, the resolution of the input image will be sacrificed, w...

AI Technical Summary

Problems solved by technology

In order to improve the receptive field of the neural network to the input image, the resolution of the input image will be sacrificed, which will cause the neural network to lose some of the smaller feature information in the original input image, and only highlight the main features.

Therefore, the convolutional neural network generally has a better classification effect on the 3 and 4 levels of sugar network features that rely on more obvious features to distinguish, but the classification effect on 1 and 2 levels of sugar network features that rely on more subtle features to distinguish

Images