Multi-scale Generation and Discrimination for Sketch Extraction of Painted Cultural Relics

Sketches of painted cultural relics can accurately represent the main content and outline structure of painted images, providing convenience for researchers studying ancient culture and the painting styles of artists. Most existing sketch extraction methods rely on edge detection technology, which may result in incomplete details, blurring, and artifacts. Additionally, for severely damaged murals, these methods may not yield ideal extraction results. In recent years, generative adversarial networks have made remarkable progress in image processing. Cycle-consistent generative adversarial networks, in particular, provide a new approach for the extraction of painted cultural relics. Therefore, this paper proposes a multi-scale generation and discrimination method for the extraction of painted cultural relics based on cycle-consistent generative adversarial networks. The method utilizes multi-scale gradient technology, allowing the discriminator to view the multi-scale output of the generator. Furthermore, a feature fusion optimization module replaces the traditional skip connection to achieve multi-scale feature fusion. Additionally, an attention residual module is designed to retain the original features, strengthen the constraints on the features of the sketch, and reduce the influence of background noise on the generated results. Experimental results demonstrate that compared with existing methods, this approach yields higher-quality sketches and performs better in objective indicators.