Image Super-resolution Reconstruction Based On Predictive Learning Of Wavelet Coefficients

In recent years,the demand for high-resolution image and video data is increasing with the rapid development of high-resolution display devices.Image super-resolution reconstruction is the process of reconstructing high-resolution image from low-resolution image by specific methods,which enables us to obtain high-resolution image data with flexibility and low cost without updating the imaging system.Recently,it has been a mainstream which applying the convolutional neural networks in the field of image super-resolution reconstruction.The main ideas for improving the super-resolution precision can be listed as follow: 1.Increasing the depth of the network structure;2.Strengthening the sparseness of learning data.This thesis bases on these two ideas,designing the network structure using improved residual block and proposing the wavelet coefficients prediction-based SR algorithm.for the problems of vanishing gradient,computational complexity when deepening network structure and the specific method for sparse data.Experiments prove that,proposed method can greatly reduce the consumption of computing resources while obtaining better image edge and detail information.Furthermore,after analyzing the improved method,existing shortcomings resulting from the advanced image interpolation,the downsampling in the orthogonal wavelet transform and the constrained network layer are found.The image is processed by the stationary wavelet transform and the wavelet coefficient residual is constructed.The effect of magnifying the image is achieved by using the scale variation generated by stationary wavelet transform and wavelet packet transform,and a deeper network structure is constructed by using the improved residual block,and the training of the network is constrained by using two loss function combinations.Experiments show the effectiveness and the advanced nature of the improved method when compared with state-of-the-art super resolution algorithms based on convolutional neural networks.