An improved version of Otsu's method for segmentation of weld defects on X-radiography images

Abstract

It is highly desirable to increase the accuracy of automatic weld defect detection in X-radiography for non-destructive testing and evaluation. For this purpose, a machine vision method is proposed for weld defect segmentation, using the Least Probability Weighted Background Group (LPWBG), an improved version of Otsu's method. This automatically selects a desired threshold value for segmenting the weld defects by adapting the Weibull distribution. The least non-zero probability value of gray-levels of the whole image has been considered as a weighted parameter of the background group of Otsu's within-class criterion. It has been determined that the resulting threshold value will always lie at the left bottom side of the unimodal distribution, especially when the defects are smaller than the background area. Existing approaches such as Otsu's threshold, valley-emphasis, neighborhood valley-emphasis, and weighted object variance have similarly been tried and compared with the proposed method. Our results establish that the proposed method provides satisfactory segmentation results over the others. The performance of the LPWBG method has been both evaluated and compared, using Misclassification Error measure (ME) with significant results.