Feature Selection of Steel Surface Defect based on P-ReliefF Method

The complex and various defects of the steel surface bring great difficulty to the feature extraction and selection. The traditional ReliefF method adjusts the weight of each-dimension feature by measuring the discrepancy between the intra-class samples and inter-class samples, which is one of the most excellent filter-style feature selection methods. However, ReliefF method measure the Discriminative power of each-dimension feature independently and lacks the consideration of the relevance within multiple feature variables. In this paper a new feature evaluation and selection method is proposed which consider the correlations between different dimensional features, which is named as P-Relief. First, one-dimensional feature with the largest Discriminative power is selected and it is used to constitute feature pairs with all other features; then, the weights of all other features are updated by evaluating the performance of the feature pair. The final feature set is obtained by iterative joint evaluation. The experimental results show that the proposed method can extract highly Discriminative and robust image features of steel surface defect. And the recognition ratio to multiple complex defects such as scratch, fold, mountains, stain and so on is significantly enhanced.