Extraction And Generation Of Sketches Of Painted Cultural Relics Based On Deep Learning

The painted cultural relics are valuable historical research materials.As an important component of painted cultural relics,sketches reflect the painting structure of cultural relics patterns and play an important role in the restoration and protection of cultural relics.Traditional sketch extraction mainly uses manual copying,which is not only inefficient,but also the accuracy of copying varies depending on the level of drawing.The sketch extraction method based on image processing can efficiently and objectively extract the sketches,which has broad application prospects.However,due to the much disease of the painted cultural relics and the complicated image background,the extraction effect of the existing methods still has much room for improvement.Recently,deep learning has achieved unprecedented results in the field of image processing.This paper is based on deep learning,from the perspective of sketch extraction and sketch generation,respectively,based on convolutional neural networks and based on generative adversarial network,researches on the computer sketch extraction and generation methods to improve the effect of sketch extraction.The main research contents are as follows:(1)Inspired by the edge detection algorithm,a detail-aware hierarchical sketch extraction method of painted cultural relics based on CNN(Convolutional Neural Networks)is proposed.This method divides the sketch extraction into two stages: coarse extraction and fine extraction.Coarse extraction combines the flow-based difference-of-Gaussians algorithm with the bi-directional cascade network under the framework of transfer learning,which reduces the requirement for deep network training for large data sets,and uses prior knowledge to guide the network to pay attention to and learn detailed features.The MSU-Net(Multi-Scale U-Net)model designed for fine extraction combines multiple middle layer features of the network decoder to effectively remove disease noise and refine the sketch.Experiments show that this method can not only extract richer and complete sketches,but also deal with more complex backgrounds,and has better results in subjective and objective metrics.(2)Inspired by the idea of image translation,a gradient guided dual-branch sketch generation method for painted cultural relic is proposed.This method uses two independent GANs to design the Sketch Generation Branch(SGB)and the Gradient-image Generation Branch(GGB)to learn different and complementary characteristic.On this basis,a feature transmission module is designed to transmit the context information of SGB to the intermediate stage of GGB to supplement,so that the gradient features of GGB are more coherent and complete.In addition,a fusion module is designed to achieve gradient guidance from GGB to SGB,so that the generated sketch can pay more attention to shape and details,and noise suppression.Experiments show that compared with the other seven methods,this method can generate rich and clean high-quality sketch images of painted cultural relics...