Research On Image Super-resolution Reconstruction Based On Convolutional Neural Network

The main purpose of single image super-resolution(SISR)work is to use low-resolution images(LR)to produce high-resolution images(HR)with sharper edges through end-to-end training.Recovering from a low-resolution image to a high-resolution image is a daunting task.Especially when the scale factor becomes larger,the recovery of local features becomes more difficult.Excessive redundancies of the information will result in the inability to recover enough high-frequency details.The existing benchmark technology also has certain limitations,including insufficient use of feature information and so on.Convolution neural networks(CNN)have achieved significant performance improvements for partial image super-resolution(SR)problems.This paper intends to extend the relevant structures and concepts of convolutional neural networks to single image super-resolution work to improve the above problems.The main innovative ideas are as follows:1.The paper utilizes a newly designed fully convolutional neural network named Accurate Image Super-resolution Using Dense Connections and Dimension Reduction Network(DCDRN)to fully exploit the image features.Contextual information of image regions utilizes efficiently and accurately through uniting dense connections and cascading small filters multiple times.And such implementation can be regarded as feature extractors to fuse local and global image features.We newly introduce 1×1 CNNs parallelization structure in the image reconstruction section to reduce data dimensions of the previous layers,which alleviates the computational burden effectively while avoiding the context info losing.2.Deep Skip Connection and Multi-Deconvolution Network for Single Image Super-Resolution(SDSR)emerged.The main work focused on enhancing the expression of CNNs,resulting in a higher peak signal-to-noise ratio(PSNR)and improving texture details of images greatly.Deep skip connection(DSC)becomes integrant to make full use of low-frequency information with rich features and demands low-frequency information to be passed to the next layer as much as possible and the layer-to-layer interaction yields the most efficient low and high-frequency information.The introduction of multiple 1×1 convolutional layers(MOL)reduces the image size and reduces the computational burden caused by deepening the number of layers.3.Image Super-Resolution Using Dense Fusion Blocks and Parallel Deconvolution Network(DPSR)invented a new type of dense blocks to improve the utilization of features.The proposed dense fusion blocks(DFB)extract rich local features through dense connections.Local features concatenate in dense blocks are utilized to learn efficient features from previous and current convolutional layers and made the network more stable.The hierarchical features are merged in a holistic manner after extracted.The introduction of parallel deconvolution network(PDN)reduces the dimension of features,improves the speed of training and the output image performance.4.Combining densely connected and stimulating modules for end-to-end processing of low-resolution to high-resolution images.The network structure composed of dense blocks composed of dense connections and multiple cascaded small filters enables the context information of the image area to be effectively and accurately used.The incentive module selectively amplifies valuable global information and suppresses useless features.The 1×1 convolution layer structure in the image reconstruction part reduces the size of the previous layer,speeding up the calculation and reducing the loss of information.In general,through a large number of data experiments and comparisons with the existed models,it is shown that the single-image super-resolution model proposed in this paper has a higher accuracy for end-to-end processing of images from low resolution to high resolution.A higher peak signal-to-noise ratio is produced and the edges remain well.