CNN Based Image Super-Resolution Reconstruction And Quality Assessment

Image super-resolution(SR)reconstruction is a basic research topic in the field of computer vision,and has a wide range of applications in satellite remote sensing,public safety and biomedical fields.The high-frequency information recovery of the low-quality image is performed by the image SR algorithm,so that the image obtains a better visual perception.Compared to improving hardware conditions,SR algorithms have many advantages,such as lower cost and wider application,thereby they have been closely watched by academics and industries.With the explosive development of deep learning(DL)in various fields,a large number of DL-based SR methods have been proposed and achieved remarkable results.However,these advanced algorithms still have some problems.On the one hand,the reconstruction difficulty increases as magnification increases,which makes network training more difficult.Hence,how to reduce the training difficulty of the high-magnification network is one of the worthy problems;On the other hand,current networks often have higher complexity and more parameters for the pursuit of better objective evaluation indicators,so that the practicality is greatly restricted.Meanwhile,the quality assessment for reconstructed image is always a difficult problem.Traditional objective assessment methods are often difficult to meet human subjective visual perception.In view of the above problems,this paper proposes corresponding solutions,including progressive upsampling network for reducing training difficulty,local binary convolution SR network for reducing the number of learnable parameters,and multi-layer perceptual decomposition based full reference images quality assessment.The main contents and contributions are as follows:(1)An image SR network with a progressive structure is constructed.In order to reduce the difficulty of SR network training to obtain better results,this paper proposes a progressive upsampling network to achieve end-to-end reconstruction,which decompose the difficult problem into multi-step and low-difficulty problems,and a gradually training method is proposed,which can further improve the network reconstruction performance under the same structure.(2)A lightweight SR network is constructed.Aiming at the problem that the DL-based SR network has a huge number of learnable parameters,this paper constructs a local binary convolution SR network by adopting pre-set fixed-value parameters,and the amount of network learnable parameters is greatly reduced without significantly affecting the subjective visual quality of the image.(3)An image assessment method and framework are proposed that is more consistent with human visual perception.In this paper,the method of full reference image quality evaluation is studied,and an image quality assessment framework based on DL is proposed.The method and framework can effectively enhance the classic image quality assessment method,and the framework can be applied to improve performance of many traditional algorithms.