储粮害虫图像识别中的特征压缩研究

为了降低储粮害虫特征空间的维数,并去除粮虫特征之间的信息冗余,需要对特征选择后的特征进行压缩处理。运用基于总体类内离散度矩阵K-L变换的特征压缩和基于距离可分性准则的特征压缩2种压缩方法,分别在累积贡献率为88.11%和99.13%的情况下,将粮虫的10维特征压缩为5维。应用压缩后的5维特征,由基于模糊决策的模糊分类器对粮仓中常见的9类粮虫进行识别分类,识别率分别为93.33%和95.56%。结果证实了基于距离可分性准则的特征压缩更适合于粮虫的特征压缩。

Research on Feature Compression in the Image Recognition of the Stored-grain Pests

The feature compression is necessary to reduce feature dimensions and remove the information redundancies among features of the stored-grain pests.The K-L translations based on the global in-class discrete matrix and the distance separable rule are proposed,and 10 dimension features are compressed into 5 dimensions,and the accumulative contribution ratio is 88.11% and 99.13% respectively.The nine categories of the stored-grain pests in grain-depot were automatically recognized by the classifier based on fuzzy decision,making use of the compressed features,the correct identification ratio is 93.33% and 95.56% respectively.The experiment shows that the compression based on the distance separable rule is more practical to compress the features of the stored-grain pests.