Study On Image-based Extraction And Recognition Of Welding Defects

Radiographic testing is a widely used method of welding quality control. With the radiographic testing weld image is acquired, it is manually analyzed to evaluate the welding quality. Extensive research has been developed on automatic extraction and recognition of welding defects, in order to ensure the reliability and stability of evaluation, and reduce the artificial differences of evaluation. However, the research results can not meet the practical application. The main problems include the accurate extraction of small welding defects, effective feature description and accurate classification of welding defects. Considering the generalized application of radiography testing, as well as the similarity among images obtained using all kinds of radiographic testing methods, the methods are studied in depth to extract and recognize the welding defects from digital radiograph images. First of all, a complete and practical extraction process of welding defects was designed, and then we did some research on the location of weld zone, extraction of non-crack welding defects, as well as extraction of crack welding defects; Secondly, we studied on the feature description and selection of welding defects; last but not least, we studied on the classification of welding defects. The chief work and innovation of present study can be drawn as follows.(1) To determine the weld zone on the radiograph image, a step-refining location method of weld zone was proposed. That is to say, firstly, locate the rough weld zone through the characteristic detection of strip region; then rapidly locate the weld boundaries based on points calculating of adjacent intervals of line gray curves. It is suggested that the present method can accurately locate the weld boundaries, besides its better adaption and practicability, which is less affected by the non-uniform image brightness and the change of weld position and distribution.(2) To extract non-crack welding defects, based on the PDE theory of image segmentation, the accelerated IAC model was obtained by introducing accelerated contraction item into IAC model, and then an extraction method of non-crack welding defects using the accelerated IAC model was proposed. At first, morphological transform was applied to remove the background of weld image; secondly, the range of image gray was narrowed through brightness transform and an image was transformed into a set of vector images by multi-threshold way; thirdly, the probabilistic relaxation method was adopted to reduce the local ambiguity; finally, the accelerated IAC model was applied to block-by-block extract non-crack welding defects, with combining proper post-processing. It is shown that the present method can effectively extract the non-crack welding defects, such as porosity, slag inclusion, incomplete fusion and incomplete penetration, and the application of the accelerated IAC model can effectively reduce the needed number of iterative loops, which is quite practicable.(3) To extract crack welding defects, based on the Beamlet theory of multi-scale geometric analysis, an extraction method of crack welding defects lying on the Beamlet analysis was put forward. For one thing, morphological transform was brought to remove the background of weld image and shield the non-crack welding defects; for another, brightness transform was processed block-by-block, and Beamlet analysis and optimal BD-RDP were applied to determine several proper beamlets as the detection result of crack welding defects; in the end, appropriate post-processing was adopted by using morphological dilation, thinning and other operations. It is shown that the present method can effectively extract crack welding defects, such as longitudinal crack, transversal crack and radiate crack. The result with wide practicability can reflect the shape and extension of crack welding defects.(4) As for the feature description and selection of welding defects, the paper combining the common description method and expert experience, preliminarily determined nine features for describing welding defect. In order to select the most effective features, a feature selection algorithm was proposed and applied which combine within-class variance with correlation measure; subsequently, the most effective five features were selected from the nine to describe welding defect. In the four removed features, the flatness (FLT), symmetry (SYM) and unsharpness (USP) were presented by KATOH Y according to experience and commonly used by others. While the result of feature selection gives explanation that the removed features are invalid or redundant, with unobvious effect on classification of welding defects.(5) In the classification of welding defects, based on the support vector machine theory, the binary tree SVM algorithm embedded WVCMFS was adopted for classification of welding defects, and then the most effective structure was determined experimentally. In order to further improve the classification performance, the paper improves from two aspects, which are including the proposition and use of an uncorrelated feature transform algorithm (LUFT and NLUFT), as well as a kind of fuzzy support vector machine for imbalanced data classification (IC-FSVM). The result shown that the binary tree IC-FSVM algorithm embedded WVCMFS has better classification performance when adopts the features determined in this paper to describe welding defects. If more new features are applied to describe welding defects, the binary tree IC-FSVM algorithm embedded WVCMFS and NLUFT could be considered.