Multiclass Detection of Radiographic Defects in Welded Joints usingTwo-stage Image Enhancement and Support Vector Machine

Speaker:
Faridfashin, Mohammadjavad; Shiraz University School of Mechanical Engineering; Iran

Authors:
Faridafshin, J.; Shiraz University; Iran
Movafeghi, A.; Nuclear Science and Technology Research Institute, Atomic Energy Organization of Iran; Iran
Faghihi, R.; Shiraz University; Iran

ID: ECNDT-0104-2018
Download: PDF

Detection of defects in radiographic images of welded joints and casting pieces is the most important section of radiographic testing (RT) in the quality assessment procedure of industrial parts. Insufficient resolution of the human eye, biased interpretation and the high cost of training experienced interpreters have made the radiographic film interpretation prone to error. Therefore, automatic detection of defects in radiographic images has been proposed as an alternative. This work studies the application of a two-stage image enhancement method in segmenting the radiographic images of welded joints to train a multiclass Support Vector Machine (SVM) for classification of defects. Firstly, the potential defects are isolated from each radiographic image by segmentation using image processing techniques. The two-stage noise reduction and edge detection constitute the proposed image enhancement process. In this process, the following stages are applied sequentially: anisotropic diffusion Gaussian filter as the first stage of noise reduction and edge detection; dilation and erosion tools of morphological image processing as the second stage of edge detection and low-pass Gaussian filter for noise reduction as the second stage. After selecting a threshold and segmenting the enhanced images, 9 geometrical features are calculated for each segment and a dataset consisting 87 training vectors categorized in 3 classes: Crack, Porosity and Lack of Penetration (LOP) obtained from 13 radiographic images is formed. This dataset is then used to design a multiclass classifier for the classification of the three mentioned classes. The One-Versus-All multiclass generalization of nonlinear binary SVM equipped with a Gaussian kernel as the classification system. Correct setting of the image processing and multiclass SVM parameters lead to high training accuracy. Also, a comparison is made between the different results obtained by different classifier parameter settings.