Automatic classification of weld defects using simulated data and an MLP neural network

The radiography technique (RT) used in non-destructive testing (NDT) of welds has evolved rapidly in recent decades and has become an established technology in the field of weld assessment. RT is widely used as an inspection tool for detecting flaws inside welded structures, pressure vessels, structural members and pipelines(1). Most radiographic exposures and film interpretations in RT are still carried out manually(2). Human interpretation of weld defects, however, is tedious, subjective and is dependent upon the experience and knowledge of the inspector. Much effort has been made recently to converge RT into an automated process of radiographic interpretation. With the advancement of digital image processing and computer architecture, the automated interpretation of weld radiographs is made possible with a system consisting of film digitisation, pre-processing, defect detection and classification. The automatic interpretation of radiographs using digital image processing reduces human involvement, thus making the inspection more reliable and faster. Basically, automatic defect interpretation consists of five stages: image digitisation, pre-processing, weld extraction, defect segmentation and defect classification. A digitised radiographic image is often corrupted by non-uniform illumination, noise and poor contrast(3). Due to the poor quality of a radiographic image, image pre-processing is normally carried out as an initial stage of defect detection. This includes noise removal and contrast enhancement. Poor quality radiographic images have led to the development of various automatic defect detection algorithms that focus on extracting defects using various image segmentation methods(4-12). In addition, rapid growth of artificial intelligence methods such as artificial neural networks (ANNs) and fuzzy techniques has become an important component to assist and enhance the task of defect detection in poor quality images. Weld extraction is usually the first step in the development of automated inspection of weld radiographs. Several techniques of weld extraction are available in the literature, such as weld extraction based on the observation that the intensity plot of a weld profile looks more like Gaussian than the other objects in the image(4), extraction of features from line image of welds for use in training a multi-layer perceptron (MLP) neural network(5) and extraction of multiple curved welds from a radiographic image using fuzzy K-NN and C-means methods(6). Problems relating to poor quality images have been addressed by several authors. For instance, Kehoe et al(7) introduced the sigmanorm contrast enhancement and mean gradient edge detector to improve the quality of the image. Bonser and Lawson(8) developed filtering and ‘window’ based variance operator for segmentation of suspected defect areas inside the weld region of partially competed welds. Murakami(9) proposed a local arithmetic operation to a limited region and was followed by thresholding methods. Various image processing techniques were used by other authors to process and segment the weld images(10-12). Research in weld defect detection has also been carried out actively in the past. For instance, Lashkia(3) proposed a detection algorithm based on fuzzy reasoning to detect low-contrast defects using local image characteristics, such as spatial contrast, spatial variance and distance between two contrast regions. Liao and Li(13) developed welding flaw detection based on the fitted line profiles of a weld image and successfully detected 93.33% of various defects from linear welds. Jacobsen et al(15) extracted parameters from the intensity profile for the network with five techniques, namely morphological filter, derivatives of Gaussian filter, Gaussian Weighted Image Moment Vector-Operator (GWIMV) filter, Fast Fourier Transform (FFT) filter and wavelet transform. Past research in automatic weld radiography focused mainly on enhancing poor quality weld images and detecting defects from the radiographs, whilst the development of automatic defect classification systems is limited. Some of the research studies on automatic weld defect classification are reviewed briefly. Jacobsen et al(15) used several features to differentiate between crack, undercut and absence of defect. Perner et al(16) introduced a framework for distinguishing classification methods, namely neural nets and decision tree. Their system only classified crack, undercut and absence of defects. Feher(17) presented a simple classification of three types of defects, namely porosity, undercut and incomplete penetration by simple rules. However, the system was unable to detect small defects that did not differ much from the background. Aoki and Suga(18) used 10 parameters to characterise the defects and classified them into five possible classes using a MLP neural network. Their investigation on 27 defects showed that 25 defects were classified correctly with a success rate of 92.6%. In their work, 35 training samples and 27 test samples were used. These are, however, too small for practical applications. Wang and Liao(19) used a set of parameters to classify six possible defects and obtained the highest accuracy of 92%. In their work, 108 data sets were used for training while 12 were used for testing. However, the 12 test samples used for classifying six types of defect are considered RADIOGRAPHY