基于分维特征和反向传播神经网络的自然纹理识别

Texture Classification Based on Fractal Dimension and BP Neural Network

提出一种利用分维特征， 即自然纹理的自相似性进行纹理识别的研究。利用原始图像、高灰度图像、低灰度图像、四个方向（０°， ４５°， ９０°， １３５°）的梯度图像及二阶多分维共八个分维数作为特征值； 分维的计算采用改进的盒子计数法（ＭＢＣＭ）； 最后利用反向传播（ＢＰ）神经网络进行纹理的分类识别。实验结果与其它技术进行了比较， 并提出利用维纳滤波进一步改进分类性能。

A nature texture classification method based on fractal dimension using self similar texture characterization is presented. For this purpose, eight fractal dimensin (FD) features are used based on the original image, the high gray level image, the low gray level image, four directional (0°, 45°, 90°, 135°) gradient image and the multi fractal dimension of order two. The fractal dimension is estimated with the modified box counting method. The texture classification is completed with back propagation (BP) neural network. The results are compared with other techniques. The Wiener filter is used to improve the performance of this method.