Skin melanoma, which accounts for more than 75% of deaths from skin lesions worldwide, is one of the most severe medical problems. Melanoma is typically identified by dermatologists visually examining lesions. The method can increase the effectiveness for identifying clinically unknown lesions against ordinarily indistinguishable lesions, which will ultimately increase the diagnostic accuracy. Traditional machine learning techniques are still unable to fully address the issue of information loss or determine the precise boundary area division. To efficiently learn feature information and successfully separate melanoma images, we employ an enhanced semantic segmentation frame work that is reported in this paper. The experiments in this article demonstrate that our enhanced neural network design produces higher segmentation accuracy of 96.6% for melanoma images than conventional methods.