基于离散余弦变换过完备字典的机织物纹理稀疏表征

Sparse representation of woven fabric texture based on discrete cosine transform over-complete dictionary

为进一步研究基于字典学习的机织物纹理表征算法的稳定性与可比性,提出用离散余弦变换(DCT)过完备字典稀疏表征算法来重构织物纹理图像。重点探讨了稀疏度、子窗口大小、字典个数对纹理表征效果的影响,利用均方根误差和峰值信噪比指标对机织物原图与重构图像之间的近似程度进行量化,并确定最终优选的稀疏度为10,子窗口大小为8像素×8像素,字典个数为256。实验结果表明,所提方法不仅方便快捷,还可得到较好的表征效果。此外,其DCT过完备字典峰值信噪比值仅次于基于训练的自适应学习字典,且优于主成分分析和非稀疏表征算法约4 d B。

In order to investigate the stationary and comparability of the algorithm for woven fabric texture representation based on dictionary learning, the sparse representation with over-complete discrete cosine transform （DCT） dictionary was used to characterize the woven fabric texture. Firstly, the influence of sparsity on woven fabric texture reconstruction was investigated. Two indexes with root mean square error and peak signal to noise ratio were calculated to quantify the approximation of original image and reconstructed image. And then the final chosen sparsity value is 10, the image patch size is 8 pixel × 8 pixel, and the number of dictionary atom is 256. Experiments demonstrate that the proposed algorithm is quick, has simple calculation and can achieve rather good effect. In addition, the method not only can achieve stable results, but also its peak signal to noise ratio is better about 4 dB than pincipal component analysis and non-sparse representation algorithm on average, which is only inferior to the K singular value decomposition learned dictionary.