Real-time detection and acquisition of the health status of farmed fish is one of the key technologies for large-scale fish farming to achieve accurate, green and sustainable development, among which, real-time identification of sick and dead fish and timely collection and treatment can constrain aquaculture water contamination, prevent disease spreading and reduce breeding risks. However, in a complex shoal environment, such as changes in illumination, overlapping targets, unstable positions, and blurring caused by water fog, it is very challenging to identify and collect sick and dead golden pomfrets in time. In this paper, an improved algorithm based on YOLOv4 is proposed. A custom Super network is integrated in the PaNet module, and the input feature is encoded and decoded to reduce the interference caused by the external environment in the fine-grained feature extraction. In addition, the activation of the tanh-v1 function enhances feature propagation and ensures the maximum information flow in the network. The Resblockbody1 module is simultaneously used to improve the positioning accuracy of the target frame. In the shoal farming scene, by analyzing the images of dead golden pomfrets and comparing the test results on different models, the YOLOv4-v1 network identified an  value of dead golden pomfrets as high as 98.31%, and the real-time detection performance reached 27FPS. Through the comparison experiment with the YOLOv4 network, YOLOv4-v1 algorithm has a basically same detection speed to the original network in the offline experiment, while the value is increased by 3.36%, the R rate increased by 2.54%, and the score (the balance between precision and recall) enhanced by 0.56%. It can be seen that the YOLOv4-v1 method has a good application potential in dead fish identification.