Sub-pixel Convolutional Neural Network For Image Super-Resolution

With the advent of the digital age,image-based information processing has been widely used in various scenarios.The resolution of the image can affect the degree of information acquisition,and the high resolution image can provide more data information.In recent years,the accuracy and computational performance of several network models based on deep learning have been greatly improved compared with traditional methods in the field of image super-resolution reconstruction.In the pre-processing stage,low-resolution images are usually enlarged into high-resolution images of the same size as the target image by using bicubic interpolation.This means that in the process of training network is required to operate in high resolution image space,which will increase the computational complexity of the network.Deepen the depth of the network can effectively improve the accuracy of image reconstruction,but only increasing the number of network layers will lead to gradient explosion and gradient disappearance.Aiming at this problem,this paper proposes an image super-resolution reconstruction method based on sub-pixel convolution network.The main research contents include:In this paper,the RGB space of low-resolution image is transformed into YCbCr space in image preprocessing stage,and the brightness channel of low-resolution image is used as network input to make the network training process in low-resolution image space,which can reduce the amount of network computation and effectively improve the training efficiency of the network.Two layers of feature extraction layer are constructed to make image feature acquisition more accurate;four layers of non-linear mapping layer are constructed,the deep network has more abundant semantic information,so deepening the number of layers of network can extract more image information.Using residual network layer instead of partial convolution layer,the feature information of low-level network is transmitted directly to high-level network,which realizes the reuse of feature information and avoids the situation of gradient explosion or gradient disappearance caused by the deepening of network layers.Using sub-pixel convolution layer as reconstruction layer,the information of the same position in multiple feature images is rearranged according to the pixel points to achieve the goal.At the same time,the correlation of feature information is unchanged,and the computational efficiency is improved.In order to verify the effectiveness of the algorithm,this paper compares with four classical algorithms in public data sets,traffic data sets and defect data sets,and verifies the peak signal-to-noise ratio(PSNR)and structural similarity(SSIM),visual reconstruction effect and single image reconstruction time.The experimental results show that the evaluation index of the algorithm is the highest.In terms of visual effect,the texture details of the algorithm are clearer,closest to the target image,and the reconstruction time of a single image is the shortest.