摘要： 为提高Fisher判别分析的质量，对图像中各像素本身的灰度值及其邻域平均灰度值特征进行两步聚类分析，根据聚类结果选取Fisher判别分析所需的训练样本，同时为了尽可能降低判别分析过程中有用信息的损失，将所得到的原训练样本集进行非线性变换，使其映射到高维空间中，利用映射后的训练样本求得Fisher判别规则。实验结果表明，与基于原训练样本的Fisher判别分析和基于寻找更多样本特征的Fisher判别分析方法生成结果相比，该方法能够获得更好的图像分割精度。

关键词: Fisher判别分析, 图像分割, 非线性变换, 两步聚类

Abstract: In order to promote the quality of Fisher discriminant analysis, this paper uses the two-step clustering to choose training samples and meanwhile expands the dimension of training samples through nonlinear transform in order to decrease the loss of useful information hidden in the samples during the process of Fisher discriminant analysis. The rule of discriminating is gained according to the expanded training samples to classify the unlabelled pattern in images. Experimental results show the image segmentation approach based on decision rules of the paper has a higher precision comparing to Fisher discriminant analysis based on initial training samples in low dimension and seeking much more sample features. It has better image segmentation precise.

Key words: Fisher discriminant analysis, image segmentation, nonlinear transform, two-step clustering

中图分类号: 

• TP391.72