基于核Fisher判别分析的粮虫特征压缩方法

针对粮仓害虫种类多、类别之间相似度比较高的特点,提出基于核Fisher判别分析的粮虫特征压缩方法.利用高斯径向基核函数,对特征选择后的10维原始数字特征进行核Fisher判别分析,即通过非线性变换将样本数据从输入空间映射到高维特征空间,然后在高维特征空间进行特征提取.从粮虫分类效果方面,将KFDA法与FDA法、PCA法和KPCA法3种方法进行了比较分析.应用KFDA法提取的前4个特征,由最近邻分类器对粮仓中常见的9类粮虫进行分类,验证集的识别率为93.33%.结果表明:KFDA法对粮虫特征的非线性比较敏感,在有效降低特征维数的同时,还提高了类别之间的可分性.

Feature compression of stored-grain insects based on kernel Fisher discrimination analysis

Due to the characteristics of stored-grain insects with multi-species and high similarity among various species,a insects feature compression method was proposed based on kernel Fisher discrimination analysis(KFDA).According to Gaussian RBF kernel function,ten selected morphological digital insect features were analyzed by KFDA.The sample data were mapped from input space to high dimensional feature space through a nonlinear mapping function to extract.nonlinear features of raw space by Fisher discrimination analysis(FDA).According to classifier recognition ratio,KFDA was compared with FDA,principle component analysis(PCA) and kernel principle component analysis(KPCA).Based on the first four features from KFDA,nine species of stored-grain insects in grain-depot were automatically recognized by the nearest neighbor classifier with correct identification ratio of 93.33% for validation set.The results show that KFDA is sensitive to nonlinear features of insects.The feature dimensions can be effectively reduced with high separability among species by KFDA.