Research On Image Restoration Algorithms Based On Wavelet Domain Joint Sub-band Learning

With the latest iterations of imaging technology,it is becoming easier to obtain highquality images.However,during the process of imaging,storage and transmission,the image quality is severely degraded due to blurring and noise.As an important and popular technique in the field of image processing,image restoration can effectively improve image quality and enhance visual perception,providing effective image information for advanced computer vision tasks such as image segmentation and target detection,thereby significantly improving the accuracy of subsequent tasks.Therefore,the study of image restoration techniques is of practical and academic value.Deep learning methods have achieved excellent performance results in the field of image restoration due to their powerful non-linear representation,and deep learning-based image restoration methods have become the dominant algorithms.Most deep learningbased methods are dedicated to finding mapping relationships on spatial pixels and do not make sufficient use of high-frequency information resulting in recovered images that lack high-frequency texture details and look very smooth.Therefore,we propose two works around specific problems in the subfields of image super-resolution and image deblurring,as follows:1.Existing wavelet transform-based super-resolution methods use a mixture of all frequency sub-bands to predict a clear image,resulting in unwanted artifacts due to the interaction between the individual frequency sub-bands.Therefore,this paper proposes an image super-resolution algorithm with a joint wavelet sub-band guided network.The method uses a multi-branch network to recover different frequency information separately and uses the high-frequency information to estimate an edge feature map to guide the recovery of different frequency features on the branch network.The method also explores the complementary relationships that exist between different frequency sub-bands to correct the high-frequency components and obtain clear high-resolution images.In quantitative comparison with current wavelet domain methods,the method outperforms the PSNR performance of current wavelet domain algorithms by about 0.2dB on each data set.In terms of subjective performance,the method can recover clear high-frequency texture information and satisfactory visual results.2.Existing wavelet-based image deblurring methods use a codec structure for recovery,which contains repeated upsampling and down-sampling resulting in loss of texture detail.Therefore,this paper proposes a simple and efficient end-to-end CNN model in the wavelet domain called a straight dilated network with wavelet transformation.The method is designed with a dilated convolution module that makes full use of the information from different receptive fields to help the network achieve better deblurring performance.In addition,a wavelet reconstruction module is designed based on the relationship between frequency sub-bands,which effectively helps the network to recover clear high-frequency texture details.Compared with existing frontier algorithms,the PSNR performance metric can be improved by about 0.5dB,while the model parameters are reduced by about half.In terms of subjective performance,the algorithm in this paper can also achieve better human eye vision than other methods.Overall,this paper proposes two joint sub-band learning image restoration networks for both image super-resolution and image deblurring problems,respectively,and achieves the best performance on multiple datasets.Through quantitative and qualitative comparisons,the proposed method is able to recover clear high-frequency texture details better,yielding higher PSNR performance and sharper texture details,as well as more pleasure human visual effect.