A Perceptually Driven Moire Pattern Removal

When photographing an electronic display screen or an object with detailed texture with a digital camera,a colored water ripple pattern often appears,that is Moire pattern In real life,the Moire in the screen image seriously affects the image quality and visual effect.Nowadays,the method of avoiding Moire is mainly at the hardware level,by adding a low-pass filter in front of the lens or adjusting the frequency of the camera's photosensitive device.But these methods are usually difficult to implement in daily shooting.Due to the irregularity of Moire patterns,it is difficult to remove all kinds of Moire patterns using geometric analysis or traditional image processing methods.Due to the appearance of Moire image datasets in recent years,this thesis chooses to use a deep learning framework and proposes a Moire removal method that can maintain consistent image perception.That is,the image after purging Moire is perceived to be of the same quality as the original clean image.Due to the limitation of the dataset,this thesis mainly focuses on the Moire that occurs when shooting the screenThis thesis believes that in the restoration of Moire images,we should pay attention to the visual subjective consistency of the restored image and the clean image,that is the perceptual consistency.But perceptual consistency cannot often be measured by objective image quality evaluation criteria such as mean square error(MSE)Therefore,we introduce perceptual loss to training to maintain the perceptual consistency of the output image and the clean image.In addition,this thesis believes that if the location information of the Moire is known,it will have a further guidance on its removalThe main contributions of this thesis are as follows1.A multi-scale residual network is proposed and applied to the Moire removal problem2.Introduce perceptual loss into the loss function,so that the output image of the network and the clean image are consistent in subjective visual perception3.A method for extracting Moire masks by measuring the angle difference of the corresponding pixel color is proposed.This mask is added as additional information to the training,which further guides the removal of moire in the imageExperimental results show that the method proposed in this thesis has a good effect on removing Moire.Finally,the method of this paper is compared with other Moire removal work and common image denoising methods.Through quantitative and qualitative analysis,the effectiveness of this method is proved.