Abstract: Aiming at the problems of unbalanced sample distribution and difficulty identification of the lesion area in diabetic retinopathy, we propose a retinal lesions grading algorithm that integrates coordinate perception and hybrid extraction. This algorithm first processes the retinal input image and the Gaussian filtering to enhance the difference between the image lesions and the background of the noise, and then the hybrid dual models composed of the backbone network of Res2Net-50 and Densenet-121 will be enhanced. The image is extracted layer by layer to achieve the full capture of the multi-scale feature texture, then the multi-layer coordinate perception module and the attention characteristics fusion module are integrated at the mixed dual model connection to achieve the purpose of eliminating the characteristics of the lesions and the realization of different lesions. The weight of semantics is reshaped, finally uses the combined loss function to relieve the uneven distribution of samples to further supervise the training and test of the model. This article is experimented on the IDRID and Aptos 2019 data sets, with the secondary weighted coefficients of 88.76% and 90.29%, respectively. Accuracy rates were 81.55% and 84.42%, which provides a new window for the diagnosis of retinopathy grades and intelligent auxiliary diagnosis.