基于纹理频谱子集的纹理识别

Texture Recognition Based on Texture Spectrum Subset

纹理频谱法是一种有效的纹理特征提取方法,但其所提取的特征高达6 561(38)维,导致很大的存储和计算复杂性.局部二值模式通过简化纹理频谱法的定义,虽然减小了计算的复杂性,却削弱了纹理的刻划能力.为了在保持纹理频谱法纹理刻划能力的同时,又减少其存储和计算复杂性,提出了基于子集的纹理频谱方法.新方法建立在统一模式的概念上,仅提取纹理频谱的1个子集,特征维数仅为原方法的12%,大大减小了空间和时间代价.实验结果表明,新方法比纹理频谱法和局部二值模式具有更好的纹理识别性能.

Texture spectrum method is very effective in extracting texture features, and has received a wide range of applications. However, it contains as large as 6 561 （38） bins, which leads to large storage and computational costs. To address this problem, Local Binary Pattern （LBP） approach simplifies the definition of the texture spectrum but at the price of weakening its capability for texture description. The paper proposes alternative texture spectrum descriptors with the merits of high discrimination power and low computational costs. Inspired by the uniform pattern in LBP, a novel encoding method is proposed which allows us to use only a portion （about 12%） of the whole texture spectrum for texture description. The experimental results on the Brodatz database indicate that the proposed method yields better performance than the original TS features with significantly lower computational costs.