Investigating the effects of data and image enhancement techniques on crack detection accuracy in FMPI

Fluorescent magnetic particle inspection (FMPI) is a vital non-destructive testing technique for detecting surface defects in ferromagnetic materials. However, existing research on FMPI crack detection using deep learning models has been hindered by the limited availability of high-quality and diverse training data. This study addresses this challenge by proposing an approach to synthesize and enhance FMPI crack images, enabling comprehensive exploration of data augmentation strategies and their impact on model performance. A largescale dataset of high-quality FMPI crack images is generated through a stepwise image synthesis method combining a diffusion model and Poisson image blending. Leveraging the synthesized dataset, the effects of various spatial and pixel-level transformations on crack detection accuracy are systematically investigated, leading to the identification of optimal data augmentation strategies tailored to the unique characteristics of FMPI crack images. A ToneCurve mapping method is developed for image enhancement, enhancing the contrast between crack indications and backgrounds, further improving model performance. The proposed image synthesis and enhancement methods significantly boost crack detection precision on a small-sample FMPI dataset, achieving a 35.2% and 17.6% improvement in mean Average Precision (mAP@0.5, YOLOv5s), and a 27.6% and 8.3% improvement (mAP@0.5, YOLOv8s), compared to non-enhancement and conventional enhancement methods, respectively, demonstrating their practical applicability. The findings underscore the importance of data augmentation strategies and the effectiveness of the proposed methods in enhancing FMPI crack detection accuracy, particularly in scenarios with limited training data. The synthesized dataset is open-sourced (https://drive.google.com/drive/folders/1ES47PcW1y6CobrOVr29jGmU6kMdeECJl?usp=sharing) to facilitate further research in this field.