［Objective ］ In order to improve the recognition accuracy of different kinds of marine fish,an improved YOLOv 5s marine fish species recognition method was proposed..［Methods ］ K-means++algorithm was used to cluster the real frames of marine fish,and more matching anchor frames were obtained with the self built data set.CIoU Loss function was replaced by SIoU Loss function as the boundary box regression algorithm to improve the accuracy and rate of convergence of the boundary box regression.Improved some C 3 modules of the backbone network,and integrated CA coordination attention mechanism into the C 3 module,which improved the recognition accuracy and detection speed of the model while reducing the number of model parameters.Finally,optimized the path aggregation network of the model to enhance the feature fusion ability of the network.［Results］ The experimental results showed that the improved Our-YOLOv 5s model had a mAP of 98.4% and a detection speed of 64 s-1 in the dataset,which was 2.4% and 6 s-1 higher than the original model,respectively.［Conclusion ］ The model can meet the real-time detection requirements of marine fish.