Research On Image Super-Resolution Algorithm Based On Deep Learning

Single image super-resolution is that a high-resolution image with good visual effect reconstructed from degraded single low-resolution image.It is widely used and studied in remote sensing,security monitoring,medical image,high-definition display,robotic vision and other fields.Since the deep learning method was introduced into the research of super-resolution,it has become a popular research method of super-resolution because of its superior performance.Based on the above background,we studied the super-resolution based on deep learning.The main contents of this paper are as follows:Firstly,the development of super-resolution at domestic and abroad are summarized,and the learning-based super-resolution method is described and analyzed in detail.The theoretical basis and advanced network structure of super-resolution based on deep learning are introduced and analyzed theoretically.Secondly,aiming at the fact that the deep learning method does not consider the possible impact of the down-sampling process on the reconstruction accuracy,a basic iterate module is proposed,which can divided into the up-sampling part and the down-sampling part.Each iterate module can independently complete the process of super-resolution and degradation,generate intermediate predicted images equal number of iteration times,and get the reconstruction results by the weighted sum of intermediate predicted images.In the training,the network convergence is adjusted by weights of the intermediate prediction loss function and the final prediction loss function.The experimental results show that the proposed method has a certain improvement in objective evaluation criteria and visual effects compared with the methods without considering the downsampling process.Thirdly,aiming at the problem that hierarchical feature maps in deep networks can not be effectively utilized,a feature extraction network based on hierarchical feature maps is proposed.All feature maps are fused by cross-channel fusion,and then the dimensionality of feature maps are reduced.Subpixel convolutional neural network is used as the up-sampling operator to reconstruct the fused feature maps.The experimental results show that the reconstructed texture details on different scales can be well restored and the reconstructed results are satisfactory.Then,aiming at the insufficiency of network depth and the poor visual perception effect of reconstruction results,a basic feature extraction module and a peceptual loss function are introduced.With feature extraction module as the main component,the whole network model is cascaded to achieve deeper depth,and the network is trained with perceptual loss function as the objective,so that the reconstructed high-resolution image has a more natural and real visual effect.The experimental results show that the reconstructed results have better visual perception effect.Finally,three super-resolution algorithms proposed in this paper are summarized and compared,the differences and advantages of network design are analyzed,and further research is prospected.