A Research Of Image Super-resolution And Reconstruction Algorithm Based On Deep Neural Network

Under the influence of physical conditions such as hardware devices and transmission media,the final image resolution on display devices still cannot meet people's needs,especially in the fields of military and medical where image details are required.With developing of deep learning and increasing of data scale,researchers have applied deep learning into the image super-resolution reconstruction.By offline end-toend training,the corresponding high-resolution images are reconstructed from known low-resolution images.Subsequent studies have shown that the convolutional network which can dynamically adjust the receptive field has better capability of feature transmission and representation.However,at present,majority of super-resolution models only use the convolution structure which fix its receptive field,resulting in the lack of smoothness and clarity of the reconstructed image structure,or the generation of artifacts that look not real enough.In order to solve above problems,we focus on solving the problem of image super-resolution reconstruction,and proposes two different image super-resolution reconstruction models:(1)Multi-path selective kernel super-resolution(MSKSR)network is based on dynamic selective kernel convolution.The fixed size of kernel in previous super resolution networks limited the receptive field to extract information,and the reconstructed image became too smooth without capturing the context information of different scales.MSKSR automatically adjusts the size of the convolution kernel through the merge-and-run structure,which emphasizes useful information and compresses the noise information simultaneously.Finally,with the copious combinations of receptive field,the reconstructed images have finer texture structures.(2)Octave super-resolution(OctSR)network is constructed by Octave convolution,which performs well in the field of image classification.As we all know,the information transmitted in general images has different frequencies.In this way,the feature maps through convolution layers can be analogized to the information fusion at different frequencies.OctSR utilizes Octave convolution to separate and re-merge features.During separation phase,the low-frequency feature components with relatively smooth transformation are down-sampled in spatial domain.As a result,the receptive field is enlarged relatively while ensuring the process of convolution.And the retained high-frequency information can still restore the fine details of the image after processing.The results show that the algorithm can not only restore high quality images,but also save space and computing cost.All final results show that the proposed two image super-resolution models in the level of network structure and convolutional layer can make full use of feature information to reconstruct images.The texture details of recovered images are clearer,and the models are easy to converge with pretty performances.The experimental results on DIV2 K show that MSKSR has higher PSNR and SSIM indexes,which improves the image clarity.OctSR can keep the performance of network and decrease the Flops by about 8.7%,which provides a reference for the application of Octave convolution in the field of super-resolution.