基于统计方法和视觉感知机制,提出一种新型的纹理特征——独立纹元矩.首先通过独立成分分析方法从图像集中学习得到独立纹元滤波器,该滤波器表现出与图像集的良好相关性;然后利用独立纹元滤波器对纹理图像进行滤波分解,抽取滤波系数的一阶矩和二阶矩组成图像的纹理特征向量.在Brodatz纹理库中的实验结果表明,独立纹元矩特征的查准率和查全率比Gabor提高了6%,达到79.27%,而且在特征向量维度相同的条件下,独立纹元矩抽取的速度比Gabor快2个数量级.
Based on statistic method and visual perception mechanism, we put forward a novel texture feature, called independent texton moment (ITEM). ITEM learns the independent texton filters from an image set by independent component analysis. After filtering texture images with such independent texton filters, ITEM features are obtained by extracting the first moment and second moment from the filtering coefficients. Experiments based on the Brodatz benchmark show that ITEM has reached to 79.27% for precision and recall, an increase of 6% in comparison with Gabor. Moreover, the feature extracting speed for ITEM is hundreds times faster than that for Gabor in the condition of equal dimension for the feature vector.