Research On Image Super Resolution Reconstruction Technology Based On Deep Learning

Image super-resolution is the technique of obtaining a high resolution image from one or a group of low resolution images by means of software.In recent years,the super-resolution technology based on deep learning has been developed rapidly.However,due to the complexity of super-resolution as a pathological problem and the diversity of application scenarios,the current algorithm is far from satisfying in all aspects.In this paper,the following three works are done from the aspects of computation,memory consumption and image quality reconstruction of the algorithm:1)We propose a fast single image super-resolution algorithm based on dense connections.The proposed algorithm gathered different layer's output with different receptive field by means of dense connection to solve the fast super-resolution algorithm with simple receptive field to the final layers that lead to the problem of low precision of reconstruction image.At the same time,The proposed algorithm use as little as possible convolution layers in the densely connected structure and use 1 x 1 convolution instead of others to reduce network computation cost.Experiments show that the quality of image reconstruction is greatly improved with less computation.2)We propose a single image super-resolution algorithm based on dense connections,residuals connections and attentional mechanisms.The algorithm is a hierarchical structure of inner and outer layers.In the inner layer,the problem of redundancy in the traditional dense connection structure is improved by proposing the channel separable dense connection structure.The outer structure considers the residual structure as an unweighted feature fusion method.In the outer structure,with inner layer structure is used as the basic element,the output from different components is weighted by the attention mechanism and then densely summed.Such structure will aggregates the characteristics of different layers in a fine-grained way.To improve the practicability of the network,the algorithm adopts less convolution kernel and convolution layer.Experimental results show that,compared with other algorithms,the algorithm USES less parameters and less computation to obtain higher image reconstruction quality.3)Experiments have proved the generalization ability of CNN-based super-resolution algorithm in remote sensing image processing,and propose a block processing strategy to solve the problem that the algorithm takes up too much memory.The strategy of local and global fine-tuning of network with remote sensing data set proves that the algorithm pre-training model of common image training sets can be directly applied to the processing of visible remote sensing images,which provides a basis for practical application.In order to solve the problem of insufficient memory of the running platform in the process of image processing of large-scale model,the block processing strategy is adopted to divide the input image into small image blocks for multiple processing.On the premise of not changing the algorithm model,the algorithm can be used in the platform with limited memory resources.