Single Image Super-resolution Reconstruction Based On Improved Convolutional Neural Network

In recent years,with the continuous development of information technology,various industries have become more and more demanding for ways to obtain more effective information as much as possible.As the most direct way to obtain information,image has important applications in many fields such as remote sensing,medical,military,public security,computer vision,etc.Therefore,the continuous improvement of the amount of information carried by images has become the unremitting pursuit of human beings.However,in many real-world situations,images can be distorted,blurred,and degraded to varying degrees,and it is often difficult to obtain high-resolution images.Common methods for improving image resolution generally fall into two categories: one is to improve the performance of the imaging device to improve the resolution of the image;the other is to use software and algorithms to improve the resolution of the image.Although the improvement of imaging equipment is the most direct method,due to the difficulties of hardware technology and the difficulty of reducing the cost,the researchers use software technology to overcome this problem,image super-resolution reconstruction(super Resolution,SR)technology research came into being.The technology is essentially a technique for improving resolution under the premise of one or more images.In the process of studying a plurality of images,the image resolution of image information is required to form a low-resolution image,and then image restoration is performed to improve the resolution of the image.Therefore,the premise of studying multiple images requires the development and maturity of single image super resolution(SISR).Considering the three factors of image reconstruction performance and network training time and image restoration,the algorithm in this paper is mainly for the problem of edge distortion and texture detail information in reconstructed images.Firstly,various preprocessing operations are performed in the underlying feature extraction layer by three interpolation methods and five sharpening methods,and the image which is only subjected to one interpolation operation and the image which is first sharpened after one interpolation are merged into a three-dimensional matrix.Then,the 3D feature map formed by the preprocessing is used as the multi-channel input of the deep residual network in the nonlinear mapping layer to obtain deeper texture detail information.Finally,the reconstruction layer introduces the sub-network in the network structure for the reduced image reconstruction time.Pixel convolution to complete the image reconstruction operation.Experimental results on several common datasets show that the proposed method has better restored texture detail information and high-frequency information of the reconstructed image compared with the classical method.The peak signal-to-noise ratio increases by 0.23 dB on average,and the structural similarity increases by 0.0066 on average.The proposed method improves the texture details of the reconstructed image and improves the image edge distortion under the premise of ensuring the image reconstruction time,and improves the performance of the reconstructed image.