Image Super-resolution Based On Convolutional Neural Networks

With the development of computer vision applications,the single image super-resolution reconstruction algorithm become a research hotspot.Image super-resolution reconstruction is a process that turn one or more low resolution image to a high quality image.Traditional reconstruction algorithms always have problems like high computational complexity,low robustness,the size of input image need to be fixed.But we can solve these problems using convolutional neural networks.At present the convolutional neural network reconstruction algorithm only have three layers.Relevant research has shown that shallow structure can have a good effect on problems that internal structure is complex,data constraint is not strong,but when dealing with real word data that internal structure is complex,it will appear problem of insufficient characterization ability.In order to get a more suitable model for image reconstruction,this paper has a further study of the super-resolution reconstruction algorithm based on convolution neural network.The main work and innovation points of this paper is as follows:1.Detailed deduced the relation between convolutional neural network and image texture and detail reconstruction,by means of derivation of the formula of forward propagation and back propagation,confirmed the superiority of convolutional neural network algorithm in image reconstruction,and verified in contrast experiment of the image reconstruction methods based on gradient operator and convolutional neural network.2.Fut forward a four-layer reconstruction model and take simulation experiment for the four-layer model and confirmed that it can obtain good effect which is preponderance of texture detail recover.The fine degree of super-resolution reconstruction results depends on whether the algorithm can extract the detailed features,in convolutional neural networks it depends on parameter settings.We take experiments and change convolution kernel parameters and the number of layers to get the best model,and come to a conclusion that the highest level of reconstruction model which is full convolution network without pre-training and parameter transfer is four.3.In order to meet the requirements of higher super-resolution reconstruction precision,we put forward a six-layer structure based on feature transfer.According to the research of features exracted by the middle of convolutional neural networks,we divided the network layers into two parts,the first three layers get shallow texture information by feature transfer,the last three layers implement feature enhanced.The input of the network is low resolution image and the output is the high resolution image.In the experiment,this model get good results and has an advantage of higher precision.Compared with the traditional method,it can recover image texture details better.