Real-time deep-learning-based object detection and unsupervised statistical analysis for quantitative evaluation of defect length direction on magnetooptical faraday effect

Abstract

Magnetooptical nondestructive inspection (MONDI) is widely used in various industries to detect defects in metal including ferromagnetic materials. Despite its high precision and sensitivity, its inspection accuracy is highly dependent on the direction of the magnetizer. In addition, the irregular shapes of defects make the assessment of their severity challenging. To address these limitations, we propose an efficient real-time deep learning (DL)-based object detection model called the MONDI detector. This model is designed to integrate a DL-based object detection model with statistical analysis based on regression model to classify and quantify the magnetizer direction. To comprehensively assess algorithm performance, this study investigates variations in the excitation field, defect shape, and defect depth. This analysis is expected to contribute to the state-of-the-art object detection paradigm in MONDI system.