Visual inspection method based on structured light has the advantages of simple structure, low-cost, and high-precision, so it is widely used in weld seam tracking and quality inspection. This method can be used to realize the automatic detection of weld surface quality, in order to improve the welding quality and realize the intelligent production of the welding industry. Currently, the research on weld quality detection based on structured light mainly focuses on single frame image detection. It is still a difficult problem to realize weld quality detection and automatic classification of defects in a three-dimensional state. Compared with the two-dimensional image information, the three-dimensional model information is more sufficient and intuitive, so the quality test results based on it have the characteristics of credibility, higher accuracy, and better stability. Additionally, the weld toe points of a two-dimensional image have the defects of low extraction accuracy and are prone to deviation. Given the above problems, this paper proposes a weld surface quality detection method that combines a K-means clustering algorithm and an illumination model. The proposed method can enhance welding automation production.
To eliminate the noise interference in the image, Gaussian filtering, threshold segmentation, and region-of-interest extraction were used to preprocess the weld image (Fig. 2), and the centroid method was used to extract the centerline of the structured light weld. The template line was fitted by the iterative least square method, and the height feature points were calculated to correct the installation error of the base metal. Considering the change of pixels in the weld centerline as the sample of the K-means algorithm, three class sets were set, and the distance between the sample and the clustering center was used to cluster. The center of gravity of the class set was calculated as a new clustering center, which was iterative until the clustering center remained unchanged. The coordinate points on both sides of the middle-class set of the final result were used as the preset weld toe feature points, and then, the sample density and absolute value of the slope at the feature point were calculated to verify the weld toe feature point pair results. The weld width and height were calculated according to the obtained feature points and the parent plate straight line. The three-dimensional model of the weld was constructed based on the centerline data of the continuous weld, and then, the light intensity of the weld model was calculated according to the illumination algorithm to obtain the illumination model of the weld. The characteristics of the dark area plaque and the related eigenvalues were established according to the intensity and distribution of the brightness value of the weld under directional illumination. The eigenvalues were used to classify whether the weld had defects and specific defects (undercut and porosity).
During the extraction of weld toe feature points, the accuracy of feature points extraction is improved by the extraction method of weld toe feature points based on K-means clustering. The width error is less than 0.2 mm, the height error is less than 0.1 mm, and repeated measurement errors of the width and height are less than 0.02 mm (Table 1). Compared with the gradient method, the weld toe feature points extracted based on K-means clustering are more accurate and stable (Fig. 6), which can meet the accuracy error in practical application. When the defects are classified based on the illumination model, the following properties are set: the number characteristic value N of the dark block, the roundness characteristic value Ra of the dark block, and the undercut characteristic value S. The characteristic value N of the defect-free weld is equal to 1, the characteristic value N of the defect weld is greater than or equal to 2, the roundness characteristic value Rc of the porosity weld is greater than 0.7 (area without defects) or between 0.3 and 0.5 (area with porosity defect), the undercut characteristic value S is between 0.5 and 0.8 (area with undercut defect) or 0 (area without defects). For the classification of defects, the accuracy of the weld classification is 1. When classifying specific defects, the classification accuracy of undercut weld and porosity weld is above 0.95 (Table 2). Our method has a good classification effect.
The K-means clustering algorithm was used to extract the weld toe feature points, which improved the extraction accuracy of weld toe feature points and extracted the welding parameters, including height and width of the weld. Furthermore, a three-dimensional illumination model was established according to the centerline data of the weld. Additionally, the characteristics and eigenvalues of the dark block were calibrated, and the automatic classification of the weld with or without defects and specific defect types (undercut and porosity) was realized according to the illumination brightness data of the illumination model.
