基于双树复小波域统计建模及局部二值模式的纹理特征提取

纹理特征提取是图像处理领域中的经典问题，如何在较低计算复杂度的前提下获取更丰富的纹理信息一直都是一个极具挑战性的难点。对此，本项目尝试在双树复小波域上通过互补性的特征融合来突破这一瓶颈。这样考虑的原因在于：图像经复小波分解后的高频信息是变化迅速的部分，使用全局性的统计模型参数可以很容易地表征其纹理特性；而图像经复小波分解后的低频信息是变化缓慢的部分，使用局部二值模式有利于提取一些细微的纹理变化。主要研究内容包括：分别使用广义伽马分布和广义冯•米塞斯分布为双树复小波系数幅值与相对相位进行统计建模，利用快速收敛的参数估计以达到模型复杂性与拟合准确性的折中，进而提取实用而有效的纹理特征；实现基于自适应阈值的局部二值模式特征提取，通过调节权值参数动态地计算二值化阈值，进而提高纹理特征的抗噪性和表达能力。预期成果将为特征提取理论发展开拓新的途径。

Texture feature extraction is a classical issue in the field of image processing. How to acquire more abundant texture information with low computational complexity has always been challenging. Therefore, this project tries to break through the bottleneck through complementary feature fusion in dual-tree complex wavelet domain. On one hand, since the high frequency information of image after wavelet decomposition varies quickly, the texture characteristics can be easily described by using global statistical model parameters. On the other hand, since the low frequency information of image after wavelet decomposition varies slowly, some slight texture changes can be accurately captured by using local binary pattern. The main research contents include: (a) Perform statistical modeling for amplitude and relative phase of dual-tree complex wavelet coefficients by using generalized Gamma distribution and generalized Von Mises distribution respectively. The compromise between model complexity and fitting accuracy should be achieved so as to extract practical and effective texture features through a parameter estimation algorithm with rapid convergence. (b) Implement feature extraction based on adaptive thresholding local binary pattern by adjusting the weights of parameter for dynamically determining the binarization threshold value, which helps to extract texture characteristics with noise resistance and high discrimination. The expected research outcome would provide a new approach for developing feature extraction theory.