Dunhuang murals are artistic cultural heritage with Chinese characteristics, which have rich reference value in visual communication design. In this paper, a fresco segmentation algorithm with enhanced edges is proposed by combining Sobel-Canny edge enhancement and improved GrabCut algorithm. A visual semantic segmentation model based on region suggestion network and full convolutional segmentation network is constructed to realize the high-precision extraction of visual elements of frescoes. The advantages of the proposed model in Dunhuang fresco visual element extraction are analyzed through comparative experiments, and the applicability of the model is explored from the perspective of user experience. The experiments show that the improved GrabCut model outperforms the mainstream segmentation algorithms in terms of PSNR (22.35dB) and SSIM (0.735), and the average running time of the visual semantic segmentation network is only 5.25 seconds. The user survey shows that the proposed model is highly recognized in the three dimensions of functionality, usability, and culture, and the study provides a feasible solution for the digital preservation of cultural heritage and crossmedia innovative design.
