Research On Image Super Resolution Reconstruction Technology Based On Convolutional Neural Network

Image super-resolution reconstruction techniques aim to reconstruct detail-rich highresolution images from low-resolution images.In the actual imaging process,the resolution of the imaging system itself is affected by complex factors such as optical distortion,random noise,motion transformation,down-sampling,etc.Image super-resolution reconstruction is a technology to improve the resolution of low-resolution images by means of software or hardware.In recent years,as a hot research topic,convolutional neural networks have powerful data fitting and feature characterization capabilities,and have achieved better reconstruction results than traditional methods in many fields,providing new thoughts to solve the image super-resolution reconstruction problem.In view of these,this paper focuses on the research of image super-resolution reconstruction technology based on convolutional neural network,including single super-resolution reconstruction algorithm,video super-resolution reconstruction algorithm and spatio-temporal video super-resolution reconstruction algorithm,aiming to improve the resolution of images in different application scenarios.It is known that the limited receptive will reduce the performance of single image superresolution.And the traditional non-local is difficult to reuse for its large computing expense.To improve it,a single super-resolution reconstruction algorithm based on shifted full-scale nonlocal residual network is proposed.In this algorithm,a shifted full-scale non-local residual block based on full-scale non-local method is adopted,which can effectively enhance the channel dependence and spatial correlation of feature layer,and greatly reduce the huge amount of computation caused by traditional non-local networks.The experimental results show that the peak signal-to-noise ratio(PSNR)of the reconstructed image by the proposed algorithm is improved by 0.3d B compared with the benchmark model.To address the problem that the existing video super-resolution reconstruction networks have weak feature alignment and feature fusion capability,which limit the super-resolution reconstruction performance,a video super-resolution reconstruction algorithm based on multiscale spatio-temporal feature fusion network is proposed.The algorithm alleviates the problem of blurring the reconstruction effect due to the motion of the neighbor frame scenes by feature alignment of adjacent frame feature maps through the reconstructed deformable convolution.And it mitigates the spatio-temporal feature map misalignment problem during feature fusion by constructing a multi-level spatio-temporal feature fusion module based on a dual pyramid structure.The experimental results demonstrate that the algorithm can still construct sharp edge details in the face of fast motion scenes on the video dataset.Compared with the benchmark VSR＿SOF algorithm,the peak signal-to-noise ratio(PSNR)performance is improved by 3.2%,and the image structural similarity index measure(SSIM)performance is improved by 5.1%.To address the problem that there are fast moving targets in the scene and the traditional video super-resolution reconstruction algorithm cannot increase the time-scale information richness and lead to poor viewing of the reconstruction results,a spatio-temporal video superresolution reconstruction algorithm based on cross-frame transformer-based network is proposed.The algorithm proposes a cross-frame transformer module,which complements the feature sequences by calculating the maximum similarity matrix and similarity coefficient matrix between adjacent frames and the current frame matrix.A multi-level residual reconstruction module based on the maximum similarity matrix and the similarity coefficient matrix is also proposed to reconstruct the feature sequence from coarse to fine.Experimental results demonstrated that the algorithm has a smaller model compared to existing two-stage spatio-temporal video super-resolution algorithms,and the reconstruction results are richer in high-frequency texture.Compared with the benchmark DAIN+EDVR method,the peak signalto-noise ratio(PSNR)is increased by 0.1d B,the image structural similarity index measure(SSIM)is increased by 0.17,and the size of the network model is reduced by 60.2%.A single-frame super-resolution reconstruction algorithm based on enhanced multichannel depth residual network is proposed to address the problem of limited resolution and severe noise in single image detecting through scattering media based on transmission matrix.The single image detecting through scattering media based on transmission matrix suffers from the limitation of the projection image resolution by the spatial lighting modulator.Moreover,the complex images reconstructed based on phase conjugate factor and pseudo-inverse factorbased reconstruction have severe random noise,which affects the observation of the target image.The algorithm uses a single frame of transmission matrix reconstructed low-resolution image as the network input and constructs an enhanced multi-channel depth residual multichannel residual network,which can effectively triple the spatial resolution of the phaseconjugate-based and pseudo-inverse factor-based reconstructed images and significantly remove the contained random noise.Compared with bicubic method,the peak signal-to-noise ratio(PSNR)of the reconstructed image is increased by 10.7 d B.