FD²-YOLO: A Frequency-Domain Dual-Stream Network Based on YOLO for Crack Detection.

Area of Science:

• Civil Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Crack detection in cement infrastructure is crucial for structural integrity and public safety.
• Existing single-backbone methods struggle with detecting slender or variable cracks in complex environments.

Purpose of the Study:

• To propose a novel network, FD 2 -YOLO, for accurate and efficient cement crack detection.
• To enhance feature extraction by integrating frequency and spatial domain information.
• To improve crack detection performance in challenging scenarios.

Main Methods:

• Developed a dual backbone architecture integrating frequency-domain (edge, texture) and spatial-domain (semantic) features.
• Introduced the Dynamic Inter-Domain Feature Fusion (DIFF) module for adaptive feature fusion using large-kernel convolutions and Hadamard operations.
• Proposed the DIA-Head module with a Deformable Interactive Attention (DIA) Module to focus on crack texture and geometric deformation features.

Main Results:

• FD 2 -YOLO achieved state-of-the-art performance on the RDD2022 dataset, outperforming existing YOLO models.
• Improvements include +1.3% mAP50, +1.1% mAP50-95, +1.8% recall, and +0.5% precision.
• On the UAV-PDD2023 dataset, FD 2 -YOLO achieved 67.9% mAP50 and 75.9% precision, outperforming lightweight and Transformer-based detectors.

Conclusions:

• FD 2 -YOLO demonstrates superior effectiveness and robustness for cement crack detection in real-world and aerial imaging scenarios.
• The proposed dual-stream architecture and fusion modules significantly enhance crack feature extraction and detection accuracy.
• The approach offers a promising solution for infrastructure health monitoring and safety.