This paper introduces a new model-based soft decoding techniqt, e to restore the widely used joint photographic expert group （JPEG） streams. The image is modeled as a two dimensional （2D） piecewise stationary autoregressive process, and the decoding task is formulated as a constrained optimization problem. All the constraints are given by the quantization intervals which available at the decoder freely. The autoregressive model serves as an important regularization term of the objective function of the optimization, and the model parameters are solved on the decoded image locally using a weighted total least square method. In addition, a novel bilateral dualside weighting scheme is proposed to minimize the influence of the blocking artifact on the accuracy of parameter estimation. Extensive experimental results suggest that the proposed algorithm systematically improves the quality of JPEG images and also outperforms existing JPEG postprocessing algorithms in a wide bit-rate range both in terms of peak signal-to-noise ratio （PSNR） and subjective quality
字典学习是信号稀疏分解研究的热点问题.在稀疏分解字典学习中,初始字典的选择影响字典学习的效果.为减小初始字典对学习字典的影响,在递归最小二乘(recursive least squares,RLS)字典学习方法中引入遗忘因子的概念.比较了最优方向法、K奇异值分解方法和RLS等3种方法的字典学习效果.分析了RLS字典学习中不同的遗传因子对字典学习效果的影响,以及遗忘因子为不同函数时的字典学习效果.仿真结果表明:RLS字典学习方法减小了初始字典对学习结果的影响,故学习效果较好;而在RLS字典学习中不同遗忘因子的选择会影响字典学习效果.
