ABSTRACT
Ancient documents are historical evidence of cultural inheritance, and the damage brought by natural and human factors to ancient documents is inevitable, resulting in the collected images of ancient Chinese characters containing a large amount of noise, which seriously affects the accuracy of subsequent image recognition and thus creates a great obstacle to the digitization of ancient documents. To address the complexity of ancient text structure, this paper proposes a Chinese ancient text image denoising method based on the Chinese character writing standard model. The method firstly adds four additional local branches based on the global branching, and uses the supplementary character detail information to weaken the phenomenon of strokes adhering to noise due to the lack of local details; secondly, it introduces the simulation noise of ancient documents to simulate the real ancient character image morphology, which can be used for the adversarial training of this method. In the training process, the minimum absolute value deviation, smoothing loss, structural consistency loss and the refined loss function formed by the adversarial loss are used to iteratively optimize the parameters. Finally, experiments prove that the model in this paper can increase the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) of the image by at least 23.8% and 11.4%, and the user evaluation index (UV) has also reached more than 80%.
