Abstract: An image enhancement method was constructed based on Faster R-CNN secondary transfer learning and Multi-Scale Retinex with Color Restoration(MSRCR) in order to overcome the insufficient sample size of underwater fish images for rapid detection of fish in underwater images. In the method, the Open Images high-definition fish data set was transferred to learn the initial training network via pre-training model, and then the convolutional network parameters of the lower three layers of the model were fixed and detected. The small-scale fish data set photographed under water was transferred to the fine-tune network for the second time, and MSRCR with color restoration was used to process underwater images to enhance their similarity with high division fish images, dealing with the problem of image degradation, and to enable the secondary transfer learning to be carried out efficiently. The tests showed that the method had 98.12% of network accuracy using only small-scale underwater shooting fish dataset training, with better detection ability and subsequent improvement ability than traditional machine learning methods, achieving rapid detection of fish targets in images at low underwater resolution. The finding provides certain reference value with deep sea exploration operations and engineering applications such as monitoring, protection and sustainable development of biological resources including demersal fish.