Automatic Microstructure Defect Detection of Ti-6Al-4V Titanium Alloy by Regions-Based Graph

Abstract

In this paper, we propose a simple and efficient approach for microstructure defect detection of Ti-6Al-4V titanium alloy based on image analysis. The proposed approach mimics the way that domain experts identify the defect area, by segmenting material grains via image preprocessing and detecting defects using region-based graph. The preprocessing step is a sequence of image processing techniques to produce potential defect regions. Next, a graph is constructed by considering the regions as nodes with connectivity determined by the pairwise distances. The connected components of this graph are the final detection result. An experiment involving 103 training and 517 testing microstructure images is carried out. The proposed method outperforms three benchmark methods with 0.919 G-mean score for the classification task. As to the performance of defect localization, the proposed approach largely outperforms two benchmark methods. In addition, the proposed method effectively detects the defect regions for 91 out of 96 defect images. Moreover, the implementation results also show that the proposed method has low computational cost. The processing time is on average 2.02 s per image, and 67.7 s for 517 images using parallel computation on a 32-core workstation.