特征融合与约简的纹理分类方法研究

Texture Classification Study Based on Feature Fusion and Selection

提出了一种新的基于多特征融合的纹理分类算法.首先,通过灰度共生矩阵(GLCM)、高斯马尔科夫随机场(GMRF)和二进制小波(Wavelet)抽取纹理特征,采用基于K近邻域(K-NN)分类器的留一法交叉验证错误率作为顺序前向搜索算法(SFS)的评估函数进行特征约简,从而有效地将多特征进行融合,最后利用K-NN分类器对融合后的特征进行分类.对Brodaz纹理库的测试结果证实:(1)GCLM,GMRF和Wavelet方法提取的纹理特征具有互补性与协同性;(2)与单独的纹理特征提取方法相比,多特征融合与约简的方法取得了更高的识别精度;(3)与简单的特征联合方法相比,文中提出的方法识别率可提高约4%;(4)与经典特征降维方法(主成分变换(PCA)、Fish-er判别(LDA)法)相比,文中提出的方法在识别精度和识别效率方面更具有优势,是一种实用的纹理分类方法.

In this paper,a novel texture classification method is proposed based on multi-features combination.Firstly,GLCM,GMRF and wavelet decomposition are used for texture feature extraction. Then the number of features is deduced by sequential forward search（SFS）algorithm and K-NN classifier＂leave-one-out＂error rate as the evaluation function.Systematic experimental comparison using Brodatz texture set shows that（1）texture features from GCLM,GMRF and dyadic wavelet are complementary and collaborative;（2）classification accuracy based on fused features outperforms each method individually;（3）compared with the simple combination of features from each method,the accuracy of proposed method increases by 4%;（4）compared with such feature extraction and dimensional reduction methods as principle component analysis and Fisher linear discriminant analysis,the proposed method is more accurate and efficient,and,consequently,apractical texture classification method.