Research on Image Super-Resolution Reconstruction Method Based on Deep Convolutional Neural Network

Abstract

Image super-resolution reconstruction is an ill-posed inverse problem, so a good model of the natural image prior needs to be used to solve it. Deep Convolutional Neural Networks can be used to solve the problems of traditional hand-crafted features in this paper. Systematically study the theoretical foundation of this way, analyze the degradation linear model and the ill-posedness of the inverse problem at the level of mathematical foundation, clarify the transformation characteristics of convolution mapping and the constraints of receptive fields, explore the gradient preservation of residual paths, the feature reuse of dense connections, and the recalibration mechanism driven by channel attention at the level of feature reuse, and elaborate the Laplacian pyramid decomposition, the sub-pixel convolution shuffling operation, and the recursive back-projection error correction method at the level of reconstruction strategy. The three-dimensional framework of the theoretical basis for deep convolutional neural networks in super-resolution mentioned above consists of network structure, information flow and multi-scale reconstruction.