IPC2006-10429 AUTOMATIC WELD BEAD RECOGNITION AND DEFECT DETECTION IN PIPELINE RADIOGRAPHS Marcelo K Felisberto, PhD Student, CPGEI/UTF-PR Curitiba, Brazil Guilherme A. Schneider, Professor, UTF-PR Curitiba, Brazil Tania M. Centeno, Professor, CPGEI/UTF-PR Curitiba, Brazil Lucia V. R. de Arruda, Professor, CPGEI/UTF-PR Curitiba, Brazil ABSTRACT The current work contributes to the research in the area of pipelines non-destructive testing by presenting new methodologies for the automatic analysis of welds radiographs. Object recognition techniques based on genetic algorithms were used for the automatic weld bead detection. In addiction, we developed an image digital filter for the detection of defects in the weld bead zone. These methodologies were tested for 120 digital radiographs from carbon steel pipeline welded joints. These images were acquired by a storage phosphor system using double-wall radiographic exposing technique with single-wall radiographic viewing, according to the ASME V code. As a result, even defects that are hard to be detected by human vision are automatically highlighted and extracted from the whole image to be classified in the further stages of the weld inspection process. INTRODUCTION For decades, radiographic inspection has been widely used as a non-destructive technique for detection of internal defects in welded structures in the industry. However the reliability of the results is directly affected by human errors during radiographs interpretation. Consequently, many efforts have been made towards the design and construction of computer tools, aiming at supporting the weld joint radiographs interpretation and, therefore, improving robustness, accuracy and speed of the inspection process (Silva et al., 2004a). According to Liao (2003), a weld radiographs interpretation system generally has three major functions: (1) segmentation of the welds from the background, (2) detection of welding defects in the weld, and (3) classification of the defects types. Our paper is concern about the first and second function, i.e., segmentation of the welds from the background and detection of welding defects. After a briefly review on related works, we explained the methods we proposed, firstly, for the automatic weld bead detection and, secondly, for welding defect segmentation. After describing the procedures for the image acquisition, we present the tests we conducted and the respective results. The article ends with our main conclusions. RELATED WORKS Liao and Ni (1996) emphasized that, since only items within a weld are of interest, it is preferred to extract the weld bead from each image before applying defect detection algorithms. Although some solutions have been proposed for this task (Lawson & Parker, 1994; Liao & Ni, 1996; Liao & Li, 1998; Liao et-al, 2000), they are very restrictive, being difficult to adapt them for a more general case where the weld bead characteristics (size, shape, position and angle of inclination) can highly vary from one radiograph to another. Concerning about automatic defect detection, different methods have been investigated by many authors (Lawson & Parker, 1994; Kazantsev et-al, 2004; Rajagopalan et-al, 2004; Lashkia, 2000; Mery & Filbert, 2002a, 2002b; Mery & Berti, 2003; Liao & Li, 1998; Liao et-al, 1999; Liao et-al, 2003; Perner et-al, 2001; Wang & Liao, 2002). The method we use in this work is closer related to what was proposed by Wang e Liao (2002), where an image background is generated and subtracted from the original weld bead image. Based on the results from a previous work (Schneider, 2004), we developed a new technique for the image background estimation. Then, similarly to Wang and Liao (2002), we subtract the estimated background from the original image to obtain a new image where the supposed defects appear highlighted. Copyright © 2006 by ASME Proceedings of IPC 2006 6th International Pipeline Conference September 25 - 29, 2006 Calgary, Alberta, Canada