Concrete Highway Crack Detection Based on Visible Light and Infrared Silicate Spectrum Image Fusion

Area of Science:

• Civil Engineering
• Computer Vision
• Materials Science

Background:

• Asphalt pavement deterioration is often first indicated by cracks.
• Current crack detection relies on visible light and convolutional neural networks, limiting use to daylight and good conditions.
• Infrared spectrum data offers a complementary approach to visible light for crack detection.

Purpose of the Study:

• To propose a novel cross-modal feature alignment technique for asphalt pavement crack detection using visible and infrared images.
• To enhance crack detection capabilities under varying illumination conditions.
• To improve the accuracy and stability of crack detection systems.

Main Methods:

• Developed a YOLOV5-based cross-modal feature alignment technique integrating visible and infrared imagery.
• Introduced an adaptive illumination-aware weight generation module for training the fusion network.
• Proposed the FA-BIFPN feature pyramid module to address multi-scale feature map alignment issues.
• Utilized a parallel dual backbone network structure for efficient training.

Main Results:

• The fused image approach demonstrated more stable performance than visible-light-only images across multiple datasets (FLIR, LLVIP, VEDAI).
• The proposed detector outperformed existing YOLOV5 unimodal detectors and CFT cross-modal fusion modules.
• Achieved 98.3% accuracy in detecting 5-pixel cracks under weak illumination on a public dataset.

Conclusions:

• Cross-modal fusion of visible and infrared images significantly enhances asphalt pavement crack detection reliability.
• The proposed adaptive illumination-aware module and FA-BIFPN effectively address challenges in feature alignment and varying light.
• This method offers a robust solution for early structural deterioration detection, even in low-light environments.