广义贝叶斯字典学习K-SVD稀疏表示算法

摘要

Sparse dictionary learning is a powerful sparse representation method for video image which has attracted much attention in ex-ploiting sparsity in signal processing. The K-SVD algorithm has achieved success in sparse representation but suffers from the problem of underutilization of dictionary atoms. Sparse Bayesian Dictionary Learning ( SBDL) algorithm exists the drawbacks of non-sparse repre-sentation and convergence of signal atoms. But Generalized Bayesian Dictionary Learning ( GBDL) K-SVD algorithm offers a novel up-date mode of sparse represent coefficient to make the learning algorithm gradually convergent and give the training vectors a perfect op-portunity to become extremely sparsity over the overcomplete dictionary. The simulation experiment of two cases where image frames with missing pixels and ones based on compressive sensing shows that the efficiency of sparsely represent by GBDL K-SVD algorithm is better than SBDL K-SVD algorithm.%稀疏字典学习是一种功能强大的视频图像稀疏表示方法,在稀疏信号处理领域引起了广泛关注.K-SVD算法在稀疏表示技术上取得了巨大成功,但遇到了字典原子未充分利用的问题,而稀疏贝叶斯字典学习(Sparse Bayesian Dictiona-ry Learning,SBDL)算法存在稀疏表示后信号原子不稀疏和不收敛的缺点.广义贝叶斯字典学习(Generalized Bayesian Dic-tionary Learning,GBDL)K-SVD算法提供了一种新型稀疏表示系数更新模式,使得过完备字典稀疏学习算法逐步收敛的同时训练向量足够稀疏.仿真结果表明,对有损像素和压缩传感这两种视频图像帧进行稀疏化,GBDL K-SVD算法表示的视频图像帧的重构效果与SBDL K-SVD算法相比有明显的提高.