In view of the problems in diabetic retinopathy(DR)such as small differences between classes and difficulty in identi-fying lesion areas,a cross-layer bilinear strategy retinopathy grading algorithm was proposed.Gaussian filtering and other pre-processing methods were applied to the input image to enhance the difference of image features,and the Res2Net-50 dual model was used to extract features and perform feature enhancement.The multi-branch attention module was used to focus on the lesion area to reduce the attention of irrelevant information.The pixel attention was used to guide the fusion module to fuse fea-ture information of different scales.Experiments show that the secondary weighting coefficient is 91.82％,the accuracy is 80.58％,the sensitivity is 97.10％,and the specificity is 97.05％on the IDRID data set.The accuracy is 84.28％and the secondary weighting coefficient is 90.05％on the APTOS 2019 data set.Results show that the algorithm has certain application value.
维普
万方数据