Super Resolution Reconstruction Algorithm Based On Dictionary Learning Using Multi-feature And Multi-scale Description

Super resolution (SR) reconstruction technology brings improvement in image quality atlower cost for the signal of limited imaging, and it provides a rich-detailed signal for thesubsequent fields such as medical image processing, remote sensing, telemetering, and videomonitoring. Super resolution reconstruction algorithm based on machine learning is a recentkey research field. This algorithm was implemented by studying the mapping betweenhigh-resolution patches and low-resolution patches, in order to search for the optimalweighting coefficient of image patches in the trained mapping. In particularly, the dictionarydescription based on sparse representation can maximize the utilization of prior knowledge inimage library, thus it becomes an open research highlight.This paper focuses on new dictionary learning of super resolution reconstruction usingmulti-feature and multi-scale description, and suppression of reconstruction residual. Mainworks are as follows:1. A structural dictionary using multi-feature description (SDB-SCSR) is suggested tosolve the insufficiency of structural distinction in common dictionary and the time-consumingmatching of optimal atoms in sparse reconstruction. In stage1, a classification criterion ofimage patches is constructed according to image entropy and variance. In stage2, a classifieddictionary is established by using gradient operator to descript edge feature and using LBPoperator to descript texture feature. In final stage, BTV regularization term is employed toobtain the optimal solution in super resolution reconstruction. Experimental results show thatSDB-SCSR improves PSNR of reconstructed image by0.2877dB, and improves MSSIM andFSIM by0.0059and0.0043respectively in comparison with classical SCSR. The dimensionof dictionary is reduced by structural classification, thus the reconstruction time is reduced toapproximately20%of the time in SCSR.2. A modified scheme based on multi-scale dictionary of weighting clustered patches(MD-SCSR) is proposed, with the purpose to improve the generalizability of dictionary.Firstly, scaled patches are constructed to learn a pyramid-like scale dictionary. Then, anadaptive k-value algorithm constrained by minimizing residual is designed. The signal isapproximated to the origin by mapping the similar patches under different scales to the high resolution patch. Meanwhile, neighborhood size is adjusted by residual to obtain the optimalhigh resolution image quality in reconstructed pyramid. Experiments demonstrate that PSNRof the reconstructed result using MD-SCSR is0.5620dB higher than that using pyramid-basedsuper resolution reconstruction; while0.2985dB higher when compared with SCSR. Thereconstructed image quality using adaptive k-value algorithm is superior to that of usingfixed-size k-NN, as the PSNR using adaptive k-value algorithm increases0.0108~0.0351dB.3. The essence of weighted dictionary atoms causes the difficulty in reconstructing thediscontinuity of signal, therefore salient residual clusteredly distribute on image edge. Aregularization constraint scheme based on multi-direction latticed gradient (MLG) is putforward. Firstly, a latticed gradient operator, which performs the differential along rationaldirection, is defined according to the discrete structure characteristic of digital image. Then aregularization term based on multi-direction latticed gradient is suggested to solve theoptimization by utilizing its character of edge suppression. Experiments show that the residualof reconstructed image edge using latticed regularization term is approximately20%of thatusing BTV regularization term.