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

Single image super-resolution(SISR)is a classic problem in computer vision.The goal of SISR is to generate a high-resolution(HR)image with fine lines and texture from the corresponding low-resolution(LR)one.In recent years,with the development of deep learning,convolutional neural networks(CNN)have been applied in more and more computer vision tasks including SISR and achieved the state-of-the-art performance.CNN-based algorithm for SISR is a learning-based one that directly learns a mapping from LR image to HR one from a large number of LR and corresponding HR image pairs.In recent years,CNN with deeper structure and more skip connections for SISR are proposed to get better quality of reconstructed image by using larger size of receptive field and making better use of multiscale features.Unfortunately,these CNNs are computationally much more expensive and require much more memory.Therefore,balancing performance and speed is an urgent problem to be solved.In this paper,the receptive field and multi-scale features in the convolutional neural network are studied.A network module termed as fast sensing block(FSB)is proposed to efficiently get a large receptive field.Furthermore,a fast-sensing super-resolution network(FSSRN)consisting of parallel FSBs is proposed.Experimental results show that FSSRN achieves better computational efficiency and good performance.In order to further reduce computational cost and memory requirement,a more compact network structure,namely fast and compact super-resolution network(FACN),is proposed.FACN uses three dilated convolutional layers with adaptive dilation factors to efficiently get a large receptive field.It also uses dense connections method between three dilated convolutional layers to make full use of multi-scale features between these layers.Experimental results show that FACN significantly reduces computational cost,number of parameters,and memory requirement while achieving the state-of-the-art quality of reconstructed image.