Image analysis and CNN-based crack depth estimation using eddy current data

Abstract

This study presents a comprehensive approach for crack depth estimation utilizing advanced image analysis techniques and a Convolutional Neural Network (CNN) model. The aim is to enhance accuracy and reliability in predicting crack depths, particularly for sub-millimeter cracks. The research addresses challenges arising from noise in images by employing a pre-processing technique and augmentation methods. The proposed method's effectiveness is showcased through its application to experimental crack data from diverse specimens. The outcomes exhibit a Mean Relative Error (MRE) of around 6%, indicating a high level of precision. These results affirm the potential of the methodology for real-world industrial applications. Additionally, the study explores the integration of eddy current image processing with CNN for Non-Destructive Evaluation (NDE) problems, offering a new approach for tiny surface-crack detection and characterization.