The invention discloses an endoscope image gastrointestinal hemorrhage detection method and system based on deep learning. On the basis of a VGG network model, the relative structures of convolution layers and pooling layers in the VGG network model are reserved, the final full connection layer of the network is changed into the convolution layers. In addition, a BN layer is connected behind eachpooling layer, so that the defect that the size of an input image is fixed is overcome, model parameters are reduced, and the network performance and the generalization capability are better improved.An inter-level feature fusion module capable of fusing shallow features and deep features is constructed, feature information of each image is fully mined and utilized, and high detection precision is still kept for some images with low shooting quality or tiny bleeding areas. According to the invention, whether bleeding occurs or not can be automatically detected, and the position of a bleedingarea can be positioned, so that the detection result is clear at a glance, doctors can be effectively helped to make accurate judgment and effective decisions, the workload of the doctors is greatly reduced, and the working efficiency of the doctors is improved.