稀疏表示中字典学习的影响因子研究

Influence Factors of Dictionary Learning in Sparse Representation

研究了稀疏表示中影响字典矩阵构建质量的关键因素,并实现了关键因子定量化表示.分别对图像数量、取块大小、字典列数和取块步长等因子进行参数调整并生成字典矩阵,结合系数矩阵对原始图像重构,以峰值信噪比和结构相似性索引测量这两种质量评价指标作为字典质量的评估依据.实验以CMU_PIE_Face数据库为数据源,结果表明当图像数量为500张、取块大小为4个像素点、字典列数为512维、取块步长为2个像素点时,所得到的字典具备对原始图像的最佳表示能力.因此,稀疏表示中关键因子的定量化表示可加速字典学习过程且简化模型复杂度,提高字典抽象层质量,具备更强的图像表现力.

We studied the key factors influencing the construction quality of dictionary matrix in sparse representation, and represented them quantitatively. The factors such as the number of images, patch size, dictionary columns and patch step were adjusted as parameters and the dictionary matrix was generated. Combined with the coefficient matrix, the original image was reconstructed, and the dictionary quality was evaluated by using the image quality assessment indices of peak signal to noise ratio and structural similarity index metric. Experiments on CMU_PIE_Face database demonstrate that the resulting dictionary has the best ability to represent the original image at image numbers of 500, patch size of 4 px, dictionary columns of 512 and patch step of 2 px. We found that the quantitative representation of key factors in sparse representation can accelerate the dictionary learning process, simplify the complexity of the model, improve the quality of the dictionary abstraction layer, and show stronger image expression.