To address the difficulty in recognizing features of retinopathy images and the low efficiency of disease grading, a retinopathy grading algorithm based on feature adaptive filtering was proposed. Firstly, the algorithm used the multi-scale filtering branches (MFB) constructed by the ResNet-50 network to extract features of retinopathy images step by step. Secondly, cascade adaptive feature filter blocks (AFFB) was adopted after filtering branches of different scales to perform feature enhancement and filtering on retinopathy images. Then, the feature complementary fusion module (FCFM) was utilized to complement the multiple local enhancement features after feature filtering, and enrich the local details of the retinopathy image by aggregating the complementary information of the local enhancement features. Hierarchical models with different local feature information were trained and experiments were performed on the IDRiD dataset. The experimental results show that the accuracy of the proposed grading algorithm was 80.58%, the weighted kappa coefficient 88.70%, the specificity 94.20%, and the sensitivity 94.10%. The algorithm could effectively identify the subtle lesions in retinopathy images and improve the grading efficiency.