Super-resolution Reconstruction Based On Integrated Features And K-means Structured Sparse Representation

High-resolution images contain more high-frequency information,which can ensure the accuracy of target recognition,image retrieval,and other technologies.At present,learning-based super-resolution technology has the widest application field and the best reconstruction effect.Therefore,we use the sparse representation-based super-resolution reconstruction algorithm as the research framework.The algorithm uses a simple gradient features in feature extraction,and can not make full use of the prior information of training samples.The feature extraction of high-resolution images uses the entire image,resulting in a larger number of image blocks.Only a single high and low resolution dictionary is built in the dictionary,so that the overall data of the dictionary is too large,seriously affecting the quality of the reconstructed image.At the same time,the regularization parameter of the dictionary is a fixed value and cannot be applied to the entire test set when dealing with a large number of different test images.Therefore,this paper focuses on the related research on the dictionary of feature extraction and establishment.It starts from ensuring the quality of the reconstructed image,reducing the time complexity and enhancing the universal applicability of the algorithm.We propose super-resolution reconstruction based on integrated features and K-means structured sparse representation.By experimenting with different test sets,the analysis of experimental results can verify the effectiveness of the proposed algorithm.Feature extraction part: by combining the improved gradient operator with DoG operator through the rule of integrated feature.With the integrated feature,low-resolution images can be directly extracted to obtain low-resolution image blocks.The high resolution image is subtracted from the enlarged low resolution image to obtain a difference image.The difference image feature is extracted to obtain a high resolution image block.This step processing can reduce the number of high-resolution image blocks,reduce the time of building dictionaries,while leaving more accurate high-frequency information.Experiments show that this method can be seen clearly in subjective evaluation,and the details of the image are clearer and the edge blurring is weakened.From the objective evaluation,we get the PSNR value and SSIM value of the algorithm are all higher than the contrast algorithm.K-means structured dictionary part: combining adjacent high and low-resolution imageblocks in a dictionary into a single image block pair,and using K-means to cluster low-resolution image blocks.At this moment,the high-resolution image blocks in the corresponding image block pairs are also automatically classified in the same class.Each class obtained through clustering will be used as a sub-dictionary within the dictionary.Due to a large number of different test images,we propose adaptive regularization sparse parameters.The final regularization parameter is determined based on the sparse representation coefficients calculated when the dictionary was created.The algorithm has universal applicability.Because the algorithm establishes the dictionary for a long time,the proposed algorithm based on the integrated feature super-resolution reconstruction algorithm is applied to the structured dictionary.Experiments show that the K-means+New_InfSR algorithm not only shortens the duration of building a dictionary,but also reduces the amount of data in the dictionary,while also maintaining a high quality image after reconstruction.