纹理图像中存在的光照、旋转、尺度变化使纹理分类成为一个极具挑战性的问题.针对传统的纹理分类算法在同时解决光照、旋转、尺度变化问题和实时性方面存在的不足,提出一种高效的光照、旋转、尺度不变纹理分类算法.首先利用原始图像及其2次高斯滤波图像构造尺度空间,采用带邻域主导方向的完备局部二值模式算法在不同尺度上提取光照、旋转不变的纹理特征;然后利用跨尺度取模式最大值的方法获得尺度不变的纹理特征;最后利用最近子空间分类器进行分类.在5个有代表性的纹理库上进行实验的结果表明,该算法不需要预先学习,能较好地解决纹理分类中的光照、旋转、尺度变化问题,并具有较高的实时性.%The variation of illumination, rotation and scale in textures makes texture classification a chal-lenging problem. Traditional texture classification algorithms have weaknesses in terms of handling illumi-nation, rotation, scale changes, and providing real-time feedback. Therefore, we presented an efficient illu-mination, rotation and scale invariant texture classification algorithm. First, a scale space was constructed by the original image and its two Gauss filtered images. Second, the completed local binary pattern with domi-nant direction in neighborhood (DDN-CLBP) algorithm was used to extract the illumination and rotation invariant features in the images with different scales in the scale space. Third, scale invariant features were obtained by taking the maximum value in each pattern across different scales. Finally, the nearest subspace classifier was used to perform classification. The experimental results on five representative texture data-bases show that the proposed algorithm can handle illumination, rotation and scale variation well without pre-learning, and it is highly efficient.
展开▼
