The invention provides a diabetic retinopathy recognition system based on fundus images, comprising: an image acquisition device for acquiring fundus images that include a target fundus image and a reference fundus image from a same person; an automatic recognition device that employs a deep learning method to process the fundus images from the image acquisition device, and automatically judges whether the fundus images present diseases and outputs judging results. In the system, the deep learning method is used to automatically judge the fundus images and output judging results; therefore, the system can assist doctors in diagnosing the fundus images for disease recognition, recognition efficiency is greatly improved, and the system is particularly applicable to wide-range screening of fundus diseases.