Study of defect feature dimension reduction based on principal component analysis

Feature extraction is an important link of visual defects detection, for it can transform high dimension space of image data into low dimension space of feature. But for the pattern classifier, high dimension input will lead the increasing of identification complexity. Therefore, it is necessary to select one group of features that can most express the defect essential characteristics. Principal component analysis makes use of the thought of statistical variance, which can remove the correlation between the statistical variables and keep all or most of the information. With the example of steel plate surface defects detection, this paper studies the feature dimension reduction based on principal component analysis. Select 7 types steel plate surface defects, acquire 20 sample images from each defect and extract 128 eigenvalues from each sample image. The experiment results show that the principal component analysis can effectively remove the correlation between the feature e data, and keep the necessary information effectively.