Single Image Super Resolution Reconstruction Based On Multi-Feature Learning And Artifact Constraint

High resolution images not only provide people with better visual enjoyment,but also provide more detailed information in many applications such as identification tracking,remote sensing,and medical imaging diagnosis,etc.However,due to the restriction of the detector technology and the degrading factors in the imaging process,the resolusion of acquired images mostly cannot meet the application requirements.Since increasing the imaging resolution by changing the structure of the optical system or using higher resolution detectors will lead to a substantial increase in system complexity and costs,image super-resolution reconstruction technology that has the characteristics of strong applicability and high performance-price ratio only uses one or several frames of low resolution input images to recover high-resolution images.Thus,it has gradually become a trend of research and application.Classic learning-based image super-resolution reconstruction algorithms extract the edge of natural images as a priori information to reconstruct high-resolution images,which can effectively improve the image resolution.Therefore,this article mainly studies on this type of algorithms.The main work includes the following aspects.(1)Study and actualize the image super-resolution reconstruction algorithm based on sparse representation.By building a sample library of nature images,the high-frequency information is extracted to train dictionaries,which are used for image reconstruction for different low-resolution images.The effectiveness of the method was verified by comparative experiments.(2)Design a multi-feature joint learning program.Traditional reconstruction algorithm extracts the high-frequency information of different images in a single way,which may lead to the problem of low accuracy and comprehensiveness of the extracted features.Aiming to solve the problem,a multi-feature joint learning method is proposed to extract different features of different image patches to obtain multi-feature dictionaries used to reconstruct a high-resolution image.The method uses gradient operator and Gabor transform to extract edge features and texture features respectively.Apart from that,it selects the corresponding dictionary for reconstruction according to the type of image patches.(3)Propose a constraint method of artificial artifact.Possible causes of the artifacts have been analyzed in combination with the reconstruction results.According to the self-similarity theory of image,it is proposed to use non-local weighted constraint terms to strengthen the constraints of the image reconstruction when solving ill-conditioning problems.The constraint helps further improve the quality of reconstruction by suppressing reconstruction errors and noise during image reconstruction.(4)Perform and analyze single-frame image super-resolution reconstruction experiments.In order to verify the superiority of the multi-feature joint learning method and the artifact constraint,multi-group reconstruction experiments are designed.The results show that the proposed method has more advantages in both subjective perception and objective indexes.In addition,for the reconstruction of color images,we also design experiments to compare the reconstruction results of different dimensions in multiple color coordinate systems.The reconstruction results are evaluated by visual observation and numerical calculations,and the recommendatory processing scheme for color image reconstruction is determined.