基于稀疏表示的图像分类字典学习

Image classification dictionary learning based on sparse representation

针对K-SVD算法学习得到的字典结构性不强的问题,利用图像的非局部自相似性,提出了基于稀疏表示的图像分类字典学习方法(NLC-DL).该方法利用K-means对图像块进行聚类并对每个子类进行字典学习,增强字典的有效性.根据正交匹配追踪算法(OMP)求得稀疏系数,迭代优化字典,最终利用优化后字典和稀疏系数矩阵重构图像.实验结果表明:生成的学习字典对训练样本的表达误差更小,能够有效地保持图像的结构信息,重构后的图像在峰值信噪比和视觉效果方面均优于传统方法.

In order to deal with the weak structure of dictionary in the K-SVD algorithm,an nonlocal classification dictionary learning method( NLC-DL) based on sparse representation was proposed by taking advantage of image nonlocal self-similarity. The method clustered image patches with structural similarity by the K-means algorithm,then the dictionaries for each class were learned to reinforce the effectiveness. The sparse coefficients obtained by the Orthogonal Matching Pursuit algorithm( OMP) were used to optimize all the dictionaries alternately. Both the sparse coefficients and the optimized dictionaries were used for reconstructing the true image. Experimental results showed that the obtained dictionaries achieved a better effect with less error on representing the training sample and maintained the structural information effectively. Furthermore,the proposed method for reconstructing images performed better than the traditional ones in terms of PSNR and visual effect.