An FCN-based segmentation network for fine linear crack detection and measurement in metals

Purpose: This paper aims to enhance the detection accuracy of fine linear fatigue cracks in aero-engine metallic components by developing an image enhancement and segmentation technique. Design/methodology/approach: A dataset of fine linear fatigue cracks was constructed using image enhancement techniques. A U-shaped semantic segmentation network based on a Fully Convolutional Network (FCN) and pre-trained VGG-16 model was developed. Dice Loss, Focal Loss, and CBAM were employed to enhance crack segmentation precision. Findings: The proposed method achieved a segmentation accuracy of 92.30% and successfully measured 95.91% of cracks with over 90% accuracy, demonstrating the effectiveness of the approach in improving crack detection and measurement. Research limitations/implications: The study focuses on small fatigue cracks, and future research may explore broader applications or the detection of other flaw types to enhance generalizability. Originality/value: This work integrates advanced image processing techniques, attention mechanisms, and loss functions to improve fine crack segmentation, providing a novel method for accurately detecting small cracks in critical aerospace components.