Research On New Methods Of Image Inpainting

Image inpainting is a method of estimating the process of image degradation and compensating for the distortion caused by the degradation process,so as to obtain the original image without disturbance and degradation.There are many factors that can cause image degradation in the imaging process,such as various blurs,noise,moire,fog,rain,etc..Due to the limitations of cost and technical process,it is almost impossible to continuously improve the imaging quality by improving the accuracy of imaging equipment,so it is particularly important to use software technology to achieve image restoration.As the main tool for image restoration,super-resolution reconstruction can cope with various degradation modes and improve the subjective effect and objective indicators of degraded images.It has a wide range of applications in the medical field,criminal investigation,high-definition broadcasting,and unmanned driving.Existing deep learning-based methods have achieved promising results compared with traditional methods,but in the process of reconstruction of complex scenes,since deep convolution cannot well balance low-frequency content and high-frequency details,it is difficult to achieve subjective and objective results.There is still room for improvement.To solve this problem,this paper proposes a V-transform-based image superresolution model combined with convolution.The new model is mainly composed of three parts:V-transform module,feature fusion and upsampling module.In the Vtransform module,in order to better complete the task of primary feature extraction,we changed the previous practice of directly increasing the model channel to 64,using the V-transform as a transition to slowly increase the model channel,while letting the spatial and frequency domain information in parallel,providing richer information for subsequent deep convolutional neural networks.At the same time,the V-transform is introduced into the super-resolution task,and the ability of the V-system to process complex signals and the good characteristics of the wavelet transform are used to obtain rich frequency domain information also reduces the pressure of the network to learn low-frequency information;finally,we design a loss function mainly for frequency domain information,which further promotes the effect of the model,stimulates the Vtransformation ability,and improves the model expression ability.The training of deep convolution networks is more stable.Test experiments on four standard datasets show that our proposed V-transformbased image super-resolution model combined with convolution can achieve better results than most methods at all scales.And the three innovations we propose have positive effects on super-resolution tasks to varying degrees.