Image Super-Resolution Reconstruction Based On Sparse Representation

The super resolution reconstruction technology enhances the image spatial resolution by performing image restoration processing on one or more low resolution images,and has important scientific significance and application value in the field of image processing.In this paper,the requirement is the high quality images acquisition in deep space exploration,and the research background is the lunar exploration project and the Mars exploration engineering load image transmission processing project.On the basis of the image compression system completed by the research group,in order to further improve the image quality,the key research is the super resolution reconstruction technology which is suitable for the remote sensing image.The super resolution reconstruction technology improves the image spatial resolution while maintaining the geometric feature integrity.Traditional methods based on the interpolation or reconstruction-based methods cannot fully utilize prior knowledge and adapt to complex image scenarios with poor performance.With the introduction of sparse processing and deep learning methods into super resolution reconstruction,its performance is greatly improved.For example,compared with the traditional methods,the PSNR of the SCSR algorithm which is based on the sparse representation and has good performance can be increased by about 2d B.Although the performance of sparse representation-based methods is slightly lower than the existing deep learning method,the sparse representation-based methods have the lower complexity and are suitable for real-time processing in orbit.Based on the SCSR algorithm,in order to further improve the image reconstruction quality,a super resolution reconstruction algorithm combining multi residual network and multi feature extraction is proposed in this paper.Compared with the SCSR algorithm,this algorithm maintains image geometric features more effectively and has similar computational complexity.The main research contents and innovations of this paper are as follows:The SCSR algorithm only extracts the horizontal and vertical features,and results in the insufficient edge and texture information.In order to improve the quality of reconstructed image and solve the problem of the SCSR algorithm,an image super resolution reconstruction algorithm with multiple features(MFSCSR)is proposed.According to the characteristics of image blocks,the algorithm extracts the contour features of non-flat blocks,transforms the texture features of flat blocks,reconstructs them with sparse models,and finally synthesizes a high resolution image.Compared to the SCSR algorithm,the proposed method is capable of restoring a more complete image structure.To enrich the internal details of the reconstructed image,the proposed method improves the deep network,and designs a multi residual network structure(MR).Only four layers of convolution are used in the network,which greatly reduces the complexity of the algorithm,introduces feature fusion,and enrichs the high-frequency detail residuals by using three residual blocks to fuse the extracted feature information of different levels.The image super resolution reconstruction algorithm which combines residual network with multi feature extraction fuses the MFSCSR algorithm and the MR network into one model.In this algorithm,the fusion of the two structures is not rude to add the respective results.The image reconstructed by the MFSCSR algorithm is used as the input of the training MR network model,and the MR network continues to correct the high-frequency detail residuals of the reconstructed image.By effectively combining the sparse representation with the residual network,the performance of the algorithm is significantly improved.Compared with the SCSR and VDSR algorithms,the PSNR value is improved by about 2d B and 0.6d B,respectively.