Research On Optimization Of Image Super-resolution Reconstruction Based On Deep Learning

Since the beginning of the 21st century,information technologies such as artificial intelligence,big data,and cloud computing have emerged one after another.The development of electronic technology has reached a new level.People have developed a variety of scientific and technological equipment,which has brought great convenience to people.,And enriching people's material and cultural life,images have played a pivotal role in the collection,transmission and storage of information.However,the image resolution of many terminal equipment will be interfered by a variety of external factors,such as transmission media,hardware equipment,etc.,which makes it difficult to achieve people's expectations,especially in the military and medical fields.Requirements.The image processed by traditional image super-resolution reconstruction technology does not perform well in terms of sharpness and smoothness.At this stage,more and more scholars are beginning to pay attention to the research of deep learning,and use the deep learning technology to carry out the image super-resolution reconstruction process to get richer image details.In order to achieve the fundamental goal of improving the spatial resolution of the image,the super-resolution reconstruction technology is to find the high-frequency features lost in the high-resolution image in the input low-resolution image,and predict it to achieve high-definition results.However,the solution to the super-resolution reconstruction problem is not unique,so the reconstruction problem needs to be restricted.The constraints are mostly the spatio-temporal complementary relationship between low-resolution images,prior knowledge of high-resolution image features,or low-resolution images.Non-linear mapping between high-resolution images and so on.This paper analyzes the process of traditional image super-resolution reconstruction,and establishes two reconstruction optimization models based on deep learning:(1)Image super-resolution model based on Octave convolution(Octave Super Resolution,OctSR): In the process of transmitting natural image information,the frequency is often inconsistent,so the output of the convolution layer is the collection of each frequency information.The principle of the OctSR algorithm is as follows: with the help of Octave convolution,the image features are separated first,and then integrated,the flatness of the low-frequency features obtained by the final separation is obviously increased,and the spatial sampling process is completed on this basis.The advantage of this reconstruction model is that the quality of the network receptive field and convolution operation are guaranteed,and the high-frequency details of the image are well preserved.According to the experimental results,the above algorithm model has advantages in calculation cost and image quality.(2)Image super-resolution model based on residual dense block network(Residual Dense Network,RDN): The network model is composed of the basic structure of dense residual block(Residual Dense Block,RDB).If the resolution of the image is low,Local feature extraction can be achieved through the above-mentioned residual dense blocks.Based on the structure of these residual dense blocks,two methods of local residual learning and local feature fusion are used for processing,making the residuals dense and fast in the network training process The stability is significantly improved.In order to improve the stability of the sampling layer of the image super-resolution reconstruction model,the requirement can be met by adding a sub-pixel convolution layer.According to the experimental results,the residual dense block network can effectively improve the speed and quality of image super-resolution reconstruction.Finally,according to the experimental results,both the network structure optimization algorithm and the convolutional layer optimization algorithm can reconstruct the image based on the feature information,and have obvious advantages in the richness of image high-frequency details and model convergence.