The water column signal at the impact point on the sea is an important basis for evaluating the firing effect, so it is of great significance to obtain the situation of water column signal detection quickly and accurately in training and exercise.In view of the characteristics of multiple sizes and shapes of marine water column signals, YOLOv5 algorithm is improved and CAs-YOLOv5s algorithm is proposed.The mixup data augmentation strategy is added to the input to build new training samples and labels by using linear interpolation, which can enrich the sample information without occupying too much storage space.The Coordinate Attention (CA) mechanism is introduced to embed location information into the channels attention, so as to enhance the feature extraction capability of the model.At the same time, the pooling method of Maxpool in Spatial Pyramid Pooling (SPP) module of the original YOLOv5s algorithm is replaced with Softpool to retain more fine-grained feature information and magnify more intense feature activation.Experimental results on the target dataset show that the mAP of CAs-YOLOv5s algorithm is increased by 4.54% to 94.75%, and the speed is up to 23.51 frames per second, which can better complete the task of water column signal detection at the impact point on the sea under the real-time requirements.