Adaptive Local-Global Synergistic Perception Network for Hydraulic Concrete Surface Defect Detection.

Area of Science:

• Civil Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Surface defects in hydraulic concrete structures are topologically heterogeneous and obscured by environmental noise.
• Conventional detection models struggle with irregular geometries and background artifacts due to fixed-grid convolutions.

Purpose of the Study:

• To propose a novel deep learning network, ALGSP-Net, for accurate and robust detection of surface defects in hydraulic concrete.
• To overcome the limitations of conventional models in capturing heterogeneous defect morphologies and suppressing noise.

Main Methods:

• Introduced Defect-aware Receptive Field Aggregation and Adaptive Dynamic Receptive Field modules to adapt receptive fields to defect shapes.
• Employed a dual-stream gating fusion strategy to integrate global context with local features, reducing semantic ambiguity and background interference.
• Developed and utilized the self-constructed SDD-HCS dataset for training and evaluation.

Main Results:

• ALGSP-Net achieved an average Precision of 77.46% and an mAP50 of 72.78% across six defect categories on the SDD-HCS dataset.
• The proposed method demonstrated superior performance compared to state-of-the-art benchmarks in both accuracy and robustness.
• Effective suppression of background noise and precise detection of complex defect geometries like slender cracks and interlaced spalling were observed.

Conclusions:

• ALGSP-Net provides a reliable and advanced solution for the intelligent maintenance of hydraulic infrastructure by accurately detecting surface defects.
• The adaptive receptive field and dual-stream fusion mechanisms are key innovations enabling superior performance in challenging conditions.
• The study highlights the potential of tailored deep learning architectures for addressing complex challenges in structural health monitoring.