Abstract：At present, morphological methods for detecting diabetic retinopathy have high complexity.The traditional deep learning methods avoid the exclusion of physiological structure, manual design features.However, they have to do large calculation and the speed is relatively slow. In order to solve these problems, this paper presents a cascade detection framework based on deep learning. Firstly, the fundus images are divided into blocks to detect whether there are lesions, and then pixels in these lesions are classified into four categories：microaneurysms, haemorrhages, hard exudates and soft exudates. The experimental results show that on the public DIARETDB1 fundus image database, the detection of four kinds lesions with sensitivity are 88.62%, 94.91%, 98.91% and 92.91%, respectively.Compared to morphological methods, the accuracy has improved 17.39% in microaneurysms and 15.18% in haemorrhages. And the detection time is only a quarter of the traditional deep learning methods.