FLAW DETECTION AND CLASSIFICATION IN RADIOGRAPHIC IMAGES

Abstract

Detection and inspection of flaws in the weldments are desirable, because human inspectors are not always consistent evaluators. The efficient and reliable detection of defect is one of the important task in Non-destructive testing and evaluation (NDT&E). An incorrect classification may disapprove a piece in good condition or approve a piece with discontinuities exceeding the limits established by the existing standards. The progress in computer science and artificial intelligent techniques have allowed the welded joint quality inspection to be carried out by using feature extraction tools, making the system of weld inspection more reliable reproducible and faster and can be implemented as per the desired requirements in NDT and NDE. The aim of the present work has been to find out the probable type of flaw created during welding process in a weldment by carrying out image pre-processing and image segmentation operations on the radiographic images of weldment. Segmentation constitutes one of the most significant problems in image processing, because the result obtained at the end of this stage strongly governs the final quality of interpretation. The welding defect detection system is to be realized using three major techniques: pre-processing the radiographic images, feature extraction and pattern classification. Pre-processing is carried out since the radiographic images have low contrast and high noise and the image background is not uniform. Feature extraction is centered principally on the measurement of properties of the regions. Feature extraction involves the extraction of morphological features associated with the shape and size of the defect. Morphological features are used to build a set of pattern classifier data inputs which are fed as inputs to the designed SVM Classifier. Finally, classification divides segmented regions into specific regions according to extracted features, assigning each region to one of a number of pre-established groups, which represents all possible types of flaws expected in an image. SVM Classifier is preferred since it packs all the relevant information in the training set into a small number of support vectors. Since the hypersurface is computed with the information from the supports vectors only, the computational efficiency of the classification of a test case is increased by the ratio of the number of data set points over the number of support vectors.