Research On Self-supervised Denoising Algorithm For Ancient Inscription Images

Nowadays,the rapid development of digital technology has prompted more information to be transmitted in the form of digital images.Therefore,more and more stone inscription protection work adopts computer technology and image processing technology for digital transformation.Because the inscriptions have suffered from natural erosion and man-made destruction for a long time,there are various scratches and wear noises in the inscription images besides the inscription information.It is necessary to use image denoising algorithm to obtain more accurate and effective inscription information.In the current related research,the denoising algorithm for inscription images based on supervised learning has the problem of insufficient generalization ability due to the lack of marking data and the unknown noise model.Therefore,this paper proposes a self-supervised learning-based denoising algorithm for inscription images,which only uses a single noisy image for training.In this paper,the image prior information is used to model the noise distribution,and the loss function is designed for the noise.And the training process is monitored according to the complexity of the text in the im age,so as to optimize the training times of the self-supervised denoising network.Since the noise of the inscription image does not have the characteristics of independent and identical distribution and the mean value is zero,it cannot be directly removed by the self-supervised denoising algorithm.To this end,In this paper,the prior information of the noise mask outside the text area of the inscription image is obtained by means of layout segmentation.And proposes the Lmask loss function guided by the noise mask,so as to constrain the noise of the inscription image in the training of the self-supervised denoising network.The design of the self-supervised learning loss function is based on pseudo-labels,and the pseudo-labels are not completely consistent with the real labels,resulting in an incomplete coverage of the noise of the inscription image by the loss function guided by the noise mask.The specific performance is that the noise in the text area will reproduce after the network is trained to a certain extent.In response to this problem,this paper analyzes the relationship between the value of Lmask and the output of the network under different iterations.According to the pixel connectivity of text strokes in inscription images,this paper proposes an adaptive early stopping mechanism based on connected domain statistics.Setting it in the network output stage to reduce residual noise in the denoising result.The decomposition experiments and the comparative experiments with mainstream methods on the self-collected inscription image data set verify the effectiveness and the advantage of the algorithm in this paper.The experimental results show that the algorithm in this paper can completely and effectively retain the strokes of Chinese characters in the inscription image while removing the most of the noise.