AI Insight
This study presents YOLO11-FR, an improved bridge crack detection system that combines frequency-domain feature processing and edge enhancement to identify cracks in concrete structures. The method addresses challenges in detecting thin, low-contrast cracks that are often obscured by concrete textures and stains by integrating a Fused Fourier Conv Mixer module for spectral analysis and a Residual Edge Enhancement Module for boundary detection. Testing on two crack datasets showed improvements of 3.4-3.7 percentage points in detection accuracy compared to the baseline YOLO11n model.
Why it matters
Accurate crack detection is critical for maintaining bridge safety and preventing structural failures. This automated detection system could improve efficiency and consistency in bridge inspection programs, potentially reducing maintenance costs and enhancing public safety through earlier identification of structural deterioration.
Understand the Science
by Yangming Zhang, Baohui Tian, Hufeng Guo
Bridge cracks are important indicators of structural deterioration, and accurate crack detection is essential for bridge operation, maintenance, and safety assessment. However, crack detection remains challenging because cracks are often slender, low-contrast, and easily confused with concrete texture, stains, and other background patterns. To address these problems, this paper proposes YOLO11-FR, an improved YOLO11-based bridge crack detector that integrates a Fused Fourier Conv Mixer (FFCM) and a Residual Edge Enhancement Module (REEM). FFCM combines local convolution with a gated residual Fourier branch to introduce full-spectrum Fourier-domain feature interaction, while lightweight gating and bounded residual scaling regulate the reconstructed response. REEM enhances crack boundary features using local and dilated depthwise branches, horizontal and vertical stripe-convolution branches, a Sobel edge prior, and channel-spatial gates. Comparative experiments were conducted on GYU-DET-Crack, a crack subset extracted from the public GYU-DET dataset. Compared with the YOLO11n baseline, YOLO11-FR improves mAP50 and mAP50-95 by 3.4 and 3.7 percentage points, respectively. Validation on the Crack500 crack dataset further shows that YOLO11-FR increases mAP50 from 56.0% to 57.7% and mAP50-95 from 32.7% to 34.8%. These results indicate that the proposed YOLO11-FR improves bridge crack detection accuracy and provides a practical detection approach for crack screening under complex concrete surface backgrounds.