A method of feature extraction which is composed of invariable moment functions and Principal Component Analysis (PCA) is presented in order to recognize and classify the surface defects of strips. First, a 22-dimensional eigenvector which was invariable was extracted from images when the image was translated, scaled and rotated. And then, in order to improve the efficiency of classification, PCA was applied to reduce the dimension of the eigenvector. As a result, the 4-dimensional eigenvector was obtained. Finally, using these eigenvectors as input, weights and thresholds of the BP neural network were trained for the purpose of defect classification. Experimental results show that the average efficiency of the correct identification can reach 85%, and it’s fit for the application for detection of surface defects of strips.