Target detection is the key link of fish tracking, behavior recognition and abnormal behavior detection of fish body. Therefore, fish detection has important practical significance. Due to the low imaging quality of underwater surveillance video, the complicated underwater environment, and the high visual diversity of fish bodies, multi-target fish detection in complex background is still a very challenging problem. In order to solve the problem that the existing multi-target fish detection is mostly carried out in a controlled environment and the generalization ability is limited, a simple and effective multi-target fish detection model in complex background was proposed. The feature extraction method based on DRN was constructed by transfer learning. The features were extracted from the original image, and the candidate detection frame was further generated by combining RPN. A multi-target fish detection model in complex background was constructed based on Faster R-CNN. The experimental results on the ImageNet2012 data set showed that the detection accuracy of this model for goldfish in complex background reached 89.5%, which was much higher than the detection accuracy of the R-CNN+AlexNet model, Faster R-CNN+VGG16 model and Faster R-CNN+ ResNet101 model in this data set, indicating that this model can effectively and accurately realize the detection of multi-target fish in complex background.