Research On Super-Resolution Image Reconstruction Based On Deep Convolution Neural Network

Super-resolution reconstruction is a technique for reconstructing a high-resolution image from a series of low-resolution images through digital image processing method.This essentially increases the high-frequency content of the reconstructed image and eliminates the degradation that is generated by during low-resolution imaging.Super-resolution reconstruction is being widely paid attention to military and economic applications.Therefore,it is necessary and urgent for super-resolution to further research,and it will be a wide range of application prospects and practical significance in various fields.This paper elaborated the current research status of image super-resolution reconstruction algorithms,then trained several mainstream super-resolution reconstruction network models,and analyzed its advantages and disadvantages.Finally combining the basic knowledge of deep learning to reconstruct single high-resolution image,this article mainly works in the following aspects:(1)Super resolution reconstruction algorithm based on improved very deep CNN.Deep learning algorithms have been widely utilized in image super-resolution.Applying Rectified Linear Units(ReLU)to deep model can help to achieve faster convergence and improve the performance of deep model.Nevertheless,networks with ReLU tend to perform poorly when the number of filter parameters is constrained to a small number.To solve this problem,an image super-resolution reconstruction algorithm is proposed in this paper,which is improved a deep convolution neural network through Max-Feature-Map,(MFM).MFM is able to compress the given feature maps by stopping the transmission of close-to-zero values in the feature maps.In doing so,the network compactness is improved by preserving needed information.Meanwhile,ReLU,Parametric ReLU and MFM activation function were utilized to validate the performance in different network structures under different iteration times.Experimental results show that this method could obtain better subjective visual evaluation and objective quantitative evaluation with fewer filters compared with the mainstream SR method.(2)Research on image super-resolution based on compact multi-path convolution neural network.In order to improve the resolution degradation problem of single image and greatly reduce the amount of network parameters,a compact multi-path convolution neural network based image super-resolution algorithm is proposed in this paper.This multi-path structure model is adopted to learn residual information between the low level and high level images to reconstruct images super-resolution.To enhance network generalization ability and make the network more compact,the activated the function named MFM is adopted here.Experimental results show that the proposed method has good reconstruction ability with image clarity and edge sharpness and has better super-resolution perform in the both objective evaluation and subjective visual effect.(3)Research on network super-resolution method based on depth-wise separable convolution cross-cascade residuals.Deep convolution neural networks have achieved remarkable results in accuracy and speed for single image super-resolution.However,when the depth of the network is increased,the flow of information is weakened and training is difficult to achieve.At the same time,most existing network structures use a single-stream structure,and context information is difficult to obtain under different receptive domains.To improve information flow to get enough details and reduce network parameters,in this paper,a Cascaded Multiscale Crossing Network(CMSC)based on a deeply separable network is proposed to super-resolution.In each cascading sub-network,multiple multi-scale crossover modules are stacked to fuse complementary multi-scale information,thereby effectively improving cross-layer information flow.At the same time,a residual learning strategy was introduced at each stage to make full use of the low-resolution feature information to further enhance the reconstruction performance.The evaluation of the benchmark datasets shows that the proposed method outperforms the most mainstream super-resolution methods.