Aiming at the problem that melanoma and non-melanoma have high visual similarity, diverse colors, blurred edges, and foreign body occlusion, which leads to poor segmentation of skin lesions, this study proposes a U-shaped structure-based context encoding and decoding network. This study uses an efficient dual-channel attention mechanism and Atrous spatial pyramid pooling modules to capture more semantic and spatial information to improve the accuracy of skin lesion segmentation. Training and testing were performed on the ISIC 2017 Dermatoscopy Image Dataset. The experimental results show that the similarity coefficient (Dice_Coefficient) of the segmentation results of the proposed algorithm is as high as 88.74%, which is 3.15 percentage points higher than the existing mainstream semantic segmentation network model DeepLab V3 Plus and is 9.93 percentage points higher than the classic U-Net network in the medical field. It has a fast running speed and good stability, and can effectively segment melanoma. The segmented image has continuous edges and clear outlines. It has good effects on quantitative analysis and recognition.