Feature Extraction and Optimisation for X-ray Weld Image Classification

Computer aided image analysis systems for radiographic inspection (X-ray or gamma ray) are among the most commonly used Non-destructive Evaluation (NDE) methods. The accuracy of these systems is very much depending on the selected features which are extracted from weld defect images. In this paper, we firstly introduce a computer aided image analysis system for X-ray image inspection and evaluation. Then, a feature optimisation approach including features extracted based on geometrical shape, edge chain code (ECC) and geometric moment invariants (GMI), feature comparison and feature selection is proposed to get the best features for classification. 7 shape geometry features, 7 ECC features and 4 GMI features are extracted and tested separately. After feature optimisation, 13 features are selected and kept. Finally, a feed-forward back-propagation neural network is implemented for the purpose of defect classification. The experimental results have proved that the new feature extraction and optimisation approach successfully improves the weld defect identification accuracy.