Diabetes retinopathy (DR) is a universal eye disease, which brings irreversible blindness risks to patients in severe cases. Due to the scarcity of professional ophthalmologists, it has become increasingly important to develop computer-aided diagnostic systems for DR grading diagnosis. However, the current mainstream deep learning methods face challenges in accurately classifying the severity of DR, making it difficult for them to provide a reliable reference for clinicians. To tackle this problem, we propose two novel modules to improve the accuracy of DR classification. Specifically, we design a multi-scale feature extraction module to capture tiny lesions in fundus images and differentiate similar lesions simultaneously. In addition, we develop a class attention module to alleviate the adverse impact of intra-class similarity on DR grading. Experimental results show that our proposed modules attain significant performance improvement on the APTOS2019 blind detection dataset, with accuracy and quadratic weighted Kappa metrics achieving 95.98% and 97.12%, respectively.