To address the problem that the traditional UNet is ineffective for segmentation of skin malignant melanoma images with variable size and shape, the improved method is mainly implemented to fully utilize the multi-scale features through two improvements, firstly, in the encoder, the global dense network, the local dense network and the dilated convolution design, and later, in the decoder, the local residual design and the classification regularization. Compared with UNet, the improved method improves 0.82%, 0.03%, 1.99%, and 1.03% in the Dice coefficient, accuracy (ACC), sensitivity (SE), and intersection-to-merge ratio (IOU) metrics, respectively. The experimental results demonstrate that the improved method can improve the image segmentation of skin malignant melanoma and is an effective underlying network structure.