Multi-stage Feature Fusion Network for Edge Detection of Dunhuang Murals

Dunhuang murals have generally suffered serious damage due to natural disasters and human factors. As a result, the conservation and restoration of Dunhuang murals are becoming increasingly urgent, in which the line drawings of Dunhuang murals are an important work in the restoration of Dunhuang murals. The line drawings of the Dunhuang murals can be regarded as an image edge detection task. Although many CNN-based edge detection methods have made significant progress, they still produce thick edges and non-edges in the Dunhuang murals dataset. To solve this problem, we propose a novel edge detection method for generating line drawings of Dunhuang murals. Firstly, a novel loss function is proposed for edge detection, which uses focal Tversky loss to effectively suppress the background pixels near the edge pixels and allows CNN to produce sharp edges. Moreover, the proposed network introduces the dilated convolution modules and the spatial attention modules in each side output features to fully learn hierarchical features and obtain richer scale features. Finally, the results of experiments conducted on the BIPED dataset and the Dunhuang murals dataset show that the proposed method can generate richer and sharper edge maps.