Defect Detection And Classification Of Pipeline Weld In X-ray Image

X-ray image is widely used in pipeline welding defects detection. How to use the computer to automatically detect and recognize weld defect in X-ray image is a hot research topic. With the development of image processing and pattern recognition technology, the accuracy of automatic(or semi-automatic) detection and recognition system for X-ray weld image defect is increasing daily. In this paper, the defect detection and recognition of X-ray image are realized by using saliency object detection and sparse representation theory. The following are the main contents:1. A fast saliency detection algorithm is proposed for the weld defect image. Saliency object detection is to find the visual attention area of the i mage by simulating human attention mechanism, which is similar with the behavior of QA Inspector. In this paper, saliency object detection is used in the defect detection. According to the characteristics of the weld image, a fast saliency detection algori thm is proposed to improve the traditional method which is visual object detection with a computational attention system(VOCUS).2. Discriminant sparse reconstruction projections(DSRP) is proposed to extract the defect image feature, based on the method sparse neighborhood preserving embedding(SNPE). Since SNPE is failed to make full use of the sparse representation model and supervision information, we combine the fisher criterion and the supervised sparse reconstruction error to perform defect image feature extraction. Extensive experiments on the weld images compared the proposed algorithm with SNPE and PCA in terms of effectiveness and robustness.3. Defect detection in weld image is realized based on the saliency object detection and our proposed algorithm DSRP. During the training procedure, firstly the image patches are randomly sampled from the original image and the corresponding saliency image respectively, from which the texture features are extracted. Then the texture features are concatenated with the corresponding original gray value to form the feature vector for every image patch. The proposed algorithm DSRP are adopted for those feature vectors to perform dimension reduction. Finally support vector machine(SVM) is trained for classification. In the detection procedure, the patch features are extracted while the original images and saliency images are traversed by sliding window according to the defect detection performed by DSRP and SVM. The effectiveness of the defect detection framework for weld images based on saliency detection and DSRP is verified by the experimental results.4. The recognition of weld defects is realized according to the defect characteristics and evaluation criteria. Geometric features, gray features and invariant moments are extracted to form a feature vector. The SVM decision tree classifier is constructed to classify the defects into five types: pore, incomplete fusion, incomplete penetration, slag inclusion and crack. According to the spatial distribution of the image fusion defects and the experience of the QA Inspector, the merging rule of the defects is proposed. The experiments demonstrated that the optimization method could improve the recognition rate.