Research On Image Super-resolution Reconstruction Technology Based On Deep Convolutional Neural Network

Image super-resolution is a very important problem in the field of computer vision and image processing.Due to the influence of the real environment or hardware equipment,the resolution of the image is too low,so we cannot get more detailed information from the pictures.The image super-resolution reconstruction technology can reconstruct highresolution images with richer details and clearness by using the original low-resolution images.With the development of deep learning in image super-resolution reconstruction,the effect of image super-resolution reconstruction is getting better and better.This paper focuses on the research and improvement of existing convolutional neural network super-resolution reconstruction algorithms in the aspects of generalization and network feature reuse.The main contents of the research are as follows:(1)Considering that most image super-resolution reconstruction algorithms that based on convolutional neural network(CNN)are based on the premise that low-resolution images are obtained by downsampled from high-resolution images through bicubic interpolation.However,when this assumption does not hold,the objective evaluation index and subjective visual effect of the reconstruction will poor.To solve this problem,an image super-resolution reconstruction algorithm based on Gaussian blur CNN is proposed.By convolving the original low-resolution image with the Gaussian blur kernel before the image is input into the network,the generalization ability of the generalization is enhanced.And low-frequency information fusion is performed,that is,low-resolution pictures are fused with highfrequency information extracted by neural networks before up-sampling in the form of global residuals;Finally,the sub-pixel convolution is used to up-sample the image to the size of the target image.thereby achieving the goal of reducing the number of network parameters and increasing the speed of calculation.(2)Considering that most existing algorithms cannot make full use of the features extracted by the convolutional layer,although a deeper network model is constructed,the reconstruction effect does not show a corresponding improvement.For this problem,combining the advantages of residual networks and densely connected networks,an image super-resolution reconstruction algorithm based on multi-scale feature fusion is proposed.This algorithm combines local dense feature fusion and local residual information fusion in the feature extraction process to construct feature-skipping fusion blocks to enhance the use of front-end network feature information;perform global feature fusion on multiple featureskipping fusion blocks to further improve feature reusability,reduce the number of network layers while enhancing network learning capabilities;After obtaining the global fusion feature,considering the problem that the image area information is relevant,the global fusion features are upsampled at various levels through convolutions of different scales to improve the final reconstruction effect.