Research On Convolutional Neural Network Based On Wavelet Domain For Image Super-Resolution

With the development of technology,the demand for high-resolution images in a large number of application scenarios is becoming more and more urgent.The spatial resolution of the image is always low because of the limitation of the imaging sensor or the imaging acquisition device.So the super-resolution technology becomes the mainstream method for obtaining high-resolution images.In this paper,the superresolution algorithm based on convolutional neural network is deeply studied from three aspects: feature decomposition,feature redundancy and feature enhancement.The specific research is as follows:Firstly,this paper proposes a Deep Residual Wavelet Super Resolution(DRWSR)algorithm based on wavelet domain to solve the problem that the most of spatial superresolution algorithms tend to produce fuzzy and over-smooth high-resolution images.In DRWSR,the task of learning high-resolution images is transformed into a series of decomposed wavelet images,which largely overcomes the shortcomings of traditional deep learning methods that cannot reconstruct many details.What is more,the idea of combining global residual learning and local residual learning is used to design a more flexible and efficient network for capturing more information.On that,in order to avoid image artifacts as much as possible,the method utilizes the constraint of spatial domain and wavelet domain to train the network,which achieves more similar reconstruction to the target high-resolution image.Whether from visual effects or objective indicators,the experimental results show that the method is better than most of other image super-resolution algorithm and can reconstruct more details.Secondly,the super-resolution problem is analyzed from the perspective of feature redundancy.To solve the information redundancy problem in the original residual block used in super-resolution methods,this paper proposes an improved algorithm based on DRWSR,Self-Adaptive Residual Wavelet Super Resolution(SARWSR).The original residual block directly adds the input to the output of the convolutional layer,while the adaptive residual learning proposed in this paper obtains more effective information by weighting the input information added to the output.The weight is adaptively acquired according to the difference between the input and the output.Experimental results show that the method greatly reduces the redundancy of information and further improves the quality of reconstructed images.Finally,from the perspective of feature enhancement,the super-resolution reconstruction method should pay more attention to the recovery of detailed information.Therefore,based on DRWSR,Mixed-Attention Residual Wavelet Superresolution network is proposed,in which the residual block in the inference network is replaced with residual attention block.The residual attention block is composed of the residual block and attention block,where the mixed attention is obtained with a typical Bottom-up Top-down structure.The experimental results show that the algorithm has good effects on subjective visual and objective indicators compared with DRWSR and SARWSR.