Research On Single Image Super-Resolution Based On Deep Convolutional Neural Network

Image is an important information carrier of the network age,and its resolution implies the richness of information.High-resolution images contain significant visual features such as edges and textures,which can express critical information.However,due to the influence of environment or hardware,people often get low-resolution images with incomplete contents.Image super resolution technology can reconstruct high-resolution estimated images from one or more low-resolution images.At present,super resolution reconstruction has become a research hotspot in the field of image processing,and has broad application scenarios and practical significance.Aiming at reconstructing single low-resolution image,this thesis proposes two single image super resolution algorithms based on deep convolutional neural network.Combining the advantages of densely connected neural network and classical residual network,a super resolution algorithm based on Residual Reconstructed Dense Network(RRDN)is proposed.(1)The network cascades multiple Residual Reconstructed Dense Blocks(RRDBlocks)to explore higher level abstract features,and introduces global skip connection to learn the residual coefficients between low-level features and high-level features.(2)The bottle layer of RRDBlock,increases the diversity of low-resolution features and ensures a smaller growth rate.The contiguous memory mechanism of RRDBlock,realizes the efficient and fast flow of feature information,avoids the degradation problem in deep network training.The local skip connection with local dense concatenation of RRDBlock,mines the related information between local low-level features and local highlevel features,and further boosts the reconstruction performance.The validity of RRDBlock is verified by ablation experiments and convergence analysis.(3)Compared with the current mainstream algorithms,the objective evaluation criteria of RRDN has greatly improved on the three up-sampling factors,and RRDN has the characteristics of fast execution.According to the structural characteristics of dual path network,a super resolution algorithm based on Compact Dual Path Network(CDPN)is proposed.(1)Based on the "micro-block" structure of dual path network,an improved dual path unit is designedto study the dual path construction method suitable for super resolution tasks.(2)The Dual Path Block(DPB)which is cascaded by multiple dual path units,is designed to control the rapid growth of network width.(3)Stacking multiple DPBs forms the backbone network of CDPN,and combines deconvolution operations to complete mapping from coarse resolution features space to fine resolution feature space,thus reducing the computational complexity of the whole network.In addition,cascading multiple DPBs can gradually improve the high-level features,expand the receptive field of network in the input image,and accelerate the flow of information between different levels features.(4)Through the analysis of hyperparameter experiment results,we propose a stacking strategy based on dual path block and dual path unit to help build a more compact dual path network.CDPN has achieved competitive results on four benchmark datasets.