Abstract: The extraction of beach waterline is an important research task in the field of marine science and coastal engineering. In order to alleviate the problems of noise sensitivity, unstable thresholds, and manual parameter adjustment that exist in traditional methods when dealing with this task, this paper performs extreme value preprocessing on image data and proposes an improved SOLOv2 waterline extraction algorithm based on deformable convolution. This paper explores the effects of feature extraction network depth and the number of deformable convolution (DCN) application layers on the accuracy and speed of the algorithm. As the depth of the feature extraction network increases, the accuracy of the waterline extraction algorithm gradually improves, but the speed decreases significantly. The actual depth of the feature extraction network is recommended to be 50; As the number of DCN application layers increases, the accuracy of the waterline extraction algorithm shows a trend of first increasing and then decreasing. The actual number of the application layers is recommended to be 2; Compared with the original SOLOv2 algorithm, the improved SOLOv2 algorithm has improved performance by 1.7% and 0.1% in average precision (AP) and average Intersection over Union (IoU), respectively. In addition, by converting the algorithm weight from single precision to half precision, the inference speed of the algorithm has been further improved.