FS-YOLO: An efficient metallic crack detection framework

Accurate and real-time detection of surface cracks in metallic components is crucial for ensuring structural integrity and reducing maintenance costs in industrial environments. However, conventional detection methods often face challenges such as poor adaptability to complex textures, sensitivity to lighting conditions, and high computational burden, which hinder their deployment in practical real-time systems. To address these limitations, we propose an efficient metallic crack detection framework named FS-You Look Only Once(YOLO) that combines lightweight image pre-processing and a dual-branch deep learning architecture. Specifically, the pre-processing stage employs Contrast Limited Adaptive Hitogram Equalization, TopHat, and RollingBall filtering to enhance contrast, highlight surface discontinuities, and suppress background noise. The detection module includes two complementary branches based on YOLOv11 : Full-YOLO captures the global context from the entire image, while Strip-YOLO focuses on fine-grained crack details through localized cropping. The outputs are merged using weighted box fusion to improve robustness and accuracy. Without relying on heavy computation, the proposed method balances detection performance and resource efficiency. In experiments, FS-YOLO achieves 0.876 mAP50–95 at 120.4 FPS with approximately 20% fewer parameters and memory than the unpruned model. making it suitable for real-time deployment in industrial inspection tasks.