Research On Image Super Resolution By Deep Neural Network

There are already many mature related algorithms in the field of traditional image enhancement: traditional noise reduction algorithms such as mean filtering and median filtering,image haze removal algorithm using dark channel prior,bicubic interpolation,etc.,which have been widely used.With the advent of deep neural network,the drawbacks of traditional image enhancement algorithms are gradually revealed.The traditional algorithms are strongly related to prior acknowledgement on some fileds and are not able to adapt to a variety of scenarios.However,algorithms on image enhancement based on deep learning can adaptively imporve the effects under multiple scenarios by training models which are suit for the characteristics of those scenarios.In this paper,we conducts an in-depth discussion on super-resolution,a sub-field of image enhancement,based on deep neural networks.In addition,we alse conduct further researches for different super-resolution application scenarios.We tried to use a highperformance computing library to implement a video player that can achieve real-time super-resolution.Specifically,the main work of this article consists of the following three parts:1.We propose a new design of residual basic module based on the spatial attention mechanism to adapt to the characteristic that we need to pay attention to the image restoration from region to region according to the specified details.On the basis of preserving the original residual module structure,this module can assign different weights to the feature in the network among different image regions,according to the different frequency information in the image feature.So that the network can be more effective and more efficient on dealing with the information contained by image features.And then more information attention is focused on the midfrequency and high-frequency regions which are able to greatly improve network performance.The module can be embedded into every basic residual module in the super-resolution network to totally replace the basic module thanks to the lightweight design.In the case of adding a small amount of complexity on the network,the performance of our new super-resolution network model is greatly improved.2.We propose a brand-new video super-resolution algotithm,which enhance the feature reconstruction module of existing networks based on convolution group and dense connection.The design of traditional video super-resolution networks' feature alignment module brings a much heavier computational burden on the video super-resolution network,leading to the difficulty of effective utilization of features at the end of the super-resolution network.The new feature reconstruction module utilizes dense connections to increase the width of the network while maintaining the depth of the model,so that the information inside the model can flow more smoothly within the whole network.By flexibly setting the number of convolution layers in the convolution group and the number of dense connection groups,dense connections can be used to improve the performance of the network in the case of limited computing resources.In addition,we use the basic residual module proposed in 1 as the basic module to build a video super-resolution network for strengthening the model's ability of feature adaptation.In the practice of real scene,we apply additional optimizations on the feature alignment module.The use of weight normalization layer makes the performance of the video super-resolution network more stable when encountering noise data.3.As far as we know,there has not been a real-time super-resolution application come out until now.So this article tries to use C++ and high-performance computing libraries of Cuda to design a real-time video player for video super-resolution as an attempt to apply the super-resolution algorithm to practice.At the end of the chapter,this article compares the performance of the implemented real-time super-resolution engine with the extensively used learning frameworks,and analyzes the possible problems in the practical application of the current super-resolution algorithm.