Multi-Scale Convolutional Neural Networks Methods For Image Super-Resolution Reconstruction

At present,image super-resolution(SR)reconstruction has attracted much attention in computer vision and other fields.Traditional SR methods gradually tend to be saturated.With the increasing progress of deep learning,convolutional neural networks(CNNs)have begun to thrive in this field.However,current methods of this kind often resort to complex deep networks to pursue the improvement of reconstruction performance,while costing high calculation.This paper mainly focuses on the multiscale CNNs for image SR.Based on the existing algorithms,a series of efficient networks are designed to capture the multi-scale feature information contained in the image,thereby achieving the reconstruction performance comparable to complex deep networks.Our primary works are as follows:1.A multi-scale residual channel attention network based on image SR is constructed.This network is mainly based on the multi-scale residual channel attention blocks.For the sake of capturing multi-scale feature information,the convolution kernels of different sizes are combined.Simultaneously,channel attention mechanism is introduced,which can adaptively recalibrate the feature information across channels.Based on the multi-scale residual channel attention blocks,the low-resolution and highresolution feature extraction stages are constructed to fully capture various image features in different spatial dimensions.Finally,experiments validate that the expected reconstruction performance is achieved with a relatively shallow network structure.2.An improved multi-scale residual network based on image SR is designed.This network effectively combines the multi-scale feature extraction branch and the high-frequency attention branch,which are responsible for extracting multi-scale features and enhancing high-frequency details,respectively.The multi-scale feature extraction branch is composed of the improved multi-scale residual blocks,using an active weighted mapping strategy to adjust the information flow of different paths;and the highfrequency attention branch relies on an improved encoder-decoder model,whose main purpose is deepening the contour and texture of the image through emphasizing high-frequency information.The experimental results show that this method can well restore the image details;3.A lightweight multi-scale residual network based on image SR is established.This network is mainly dedicated to lightening the parameter burden.The network is built by alternately stacking the multi-scale feature extraction blocks and the designed residual blocks.Firstly,convolutions with multiple dilation rates are employed to construct the multi-scale feature extraction blocks,which can expand the receptive field while controlling network parameters.Secondly,a residual block based on the spatial and channel attention mechanism is proposed to properly correct the feature information of different spatial regions and channels.And the group convolution is utilized to further reduce the parameter quantity.In particular,a multi-scale feature attention module is constructed,where deep features provide pixel-wise guidance for shallow features.In addition,a dual-path upscaling module is also designed for sufficiently upsampling the feature information on high-frequency and low-frequency.Experiments show that the desired reconstruction performance of complex deep networks is attained with a lightweight network structure.