面向有限带宽信道的字典学习图像超分辨率重建

固定邻域回归(ANR)算法采用K层奇异值分解(K-SVD)算法进行字典训练, 在字典学习过程中存在稀疏表示系数不准确的问题, 导致重建的结果不理想. 因此, 引入一种改进的K-SVD算法对字典进行训练, 该算法对字典训练改变了传统K-SVD算法更新稀疏表示系数的方式, 使得稀疏表示系数更加准确, 而且加快了字典的收敛速度, 使得训练得到的字典具有更好的稀疏表达能力. 同时, 针对ANR算法的不足, 提出一种面向有限带宽信道基于字典学习的图像超分辨率方法, 该方法采用改进的K-SVD算法训练字典对 , 并将其应用到ANR算法中, 实现图像的超分辨率重建. 实验结果表明, 本文提出的方法不仅能够保持ANR算法快速重建的优势, 而且提高了图像的重建质量, 具有更高的峰值信噪比和结构相似度.

Image super-resolution with dictionary learning for limited bandwidth channel

ANR algorithm trains the dictionaries using the K-SVD algorithm, which makes sparse representation coefficients inaccurate and results in unfaithful reconstruction effect. Thus, an improved K-SVD algorithm is presented in this paper to train the dictionaries. Compared with the K-SVD algorithm, the improved K-SVD algorithm only updates the nonzero coefficients in the sparse representation coefficients at the iteration of dictionary learning. Doing this will not only make sparse representation coefficients more accurate but also speed up the convergence of the dictionary training, which make the dictionary more expressive. At the same time, aiming at the deficiencies of the ANR algorithm, an image super-resolution method with dictionary learning for limited bandwidth channel is proposed. This method uses the improved K-SVD algorithm to train the dictionary pair and applies it to the ANR algorithm to realize image super-resolution reconstruction. Experimental results show that the proposed algorithm not only preserves the advantages of ANR algorithm but also gets better visual effect, higher PSNR and SSIM.