For the difficult problem of small water floating object recognition, a deep learning-based target recognition method is improved. In this paper, an improved YOLOv5s target recognition algorithm is used to recognize water floating objects. First, according to the characteristics of the shape of floating objects on the water surface, the improved K-means algorithm is used to re-cluster the anchor boxes, then, Squeeze-and-Excitation networks module is added and then α-IOU is applied to the YOLOv5s network. The experimental results show that, compared with the standard YOLOv5s algorithm, the precision and average precision of the improved YOLOv5s algorithm are increased by 2% and 4% respectively, which verifies the effectiveness of the algorithm. This method can overcome the influence of water surface environment and effectively identify floating objects on the water surface.