According to a fundus image retinal vessel segmentation method based on evolutionary neural architecture search provided by the invention, a U-shaped decoding and coding structure is used as a backbone network to search and optimize the internal structures of modules in the backbone network, so that structures and operations which are better than those of artificial design and have lower computational complexity are found for the modules; according to the invention, the neural network model can more effectively process the complex condition of the fundus image, has higher robustness for interference in the complex fundus image focus, the vascular center reflection phenomenon and the illumination imbalance phenomenon, and can more accurately extract the features of retinal vessels, therebyimproving the segmentation accuracy of the whole image. According to the method, the flexibility of the architecture is guaranteed, the searching efficiency of the architecture is improved, the crossoperation in the genetic algorithm is improved, the searching capability of the genetic algorithm in the architecture searching process is improved, and the method has more potential to be applied toclinical disease diagnosis.