The K-L（Karhunen-Loeve） transform is applied to reduce dimensions of the characteristics extracted from the contaminated insulator infrared image. Three independent principal components including information of every original characteristic are then obtained by the K-L transform. An orthonormal matrix composed of three eigenvectors which correspond to three maximal eigenvalues of the covariance matrix of the original characteristic data is selected by the principal component cumulative contribution proportion and thus used to extract the principal components. The data among the orthonormal matrix denote the proportion of every original characteristic in the principal components. In order to compare the changes of the space distances between classes of the original characteristic data classes and the principal component characteristic data classes， the three-dimensional figures imaging the characteristic data distribution before and after the K-L transform are illustrated. Finally， experimental results indicate that the K-L transform can reduce the dimensions of the original contaminated data while to a great extent preserving the information of the original characteristic data， also， lessen the calculation， increase the distances between classes， and improve the accuracy of data classification.