Research On Moire Removal Method Of Screen Image Based On Multi-scale Information Fusion

With the rapid development of smartphones and the popularity of electronic office,people are increasingly using mobile phones to quickly capture electronic screens and save useful information.However,due to hardware limitations,users inevitably encounter moiré noise interference on the image durin g the shooting process.The color and form of the interference arc varied and difficult to distinguish from the image content,severely affecting the image quality.Therefore,removing moiré from screen-captured images has important research and application value.Existing methods for removing moiré from images mainly include traditional moiréremoval methods and deep learning-based moiré removal methods.However,these methods often cannot perceive large-scale moiré patterns as a whole,resulting in poor image quality after removal.In addition,with the high-speed development of cameras,the resolution of moiréimages is getting higher and higher,and existing methods cannot adapt to the increasing computing costs and the task of restoring high-definition details in the image.Based on these problems,this thesis conducts the following research:(1)Firstly,the characteristics of moiré images are analyzed.Starting from the screen-capturing process,the physical principles of moiré production are clarified,and the inherent characteristics of such images are obtained through comparative analysis of the frequency spectrum of moiré images and the moiré intensity of different color channels.This lays the foundation for subsequent deep learning-based moiré removal of screen-captured images.(2)For the problem of large-scale moiré patterns,a moiré removal method based on multi-scale non-local attention is proposed.By introducing a non-local attention module as a mask branch for information supplementation,th e module can effectively capture long-distance dependencies in the image,thereby increasing the network’s receptive field size without increasing the model’s depth.In addition,the YUV color model’s UV channel color loss is introduced and combined with the L1 loss as the total loss of the network to solve the problem of color deviation caused by moiré interference.Experimental results show that the proposed method performs well in restoring moiré images and has the characteristic of a small parameter model,which is suitable for deployment on mobile devices for real-time moiré removal of mobile phone images.(3)For high-resolution moiré images,a moiré removal method based on an attention dynamic fusion mechanism is proposed.This method designs an attention dynamic fusion module to adaptively adjust the channel and branch weight according to different moiré images for dynamic feature extraction and fusion.The weight-sharing strategy is used to promote information interaction between branches,making the network model have fewer parameters while having better performance.The deep supervision strategy is used to promote model optimization during network training.Experimental results show that the proposed method achieves good experimental results on moiré datasets of different sizes,which verifies the effectiveness and applicability of the proposed method.