Hydraulic concrete defect detection based on deep learning

The timely identification of defects is critical to ensuring operational safety in hydraulic concrete structures. To address limitations of traditional manual inspection methods—which are labor-intensive, subjective, and prone to human error—we propose a deep learning-based computer vision framework for automated defect detection. This approach eliminates the need for complex handcrafted feature engineering by leveraging convolutional neural networks to autonomously learn discriminative features from remote imagery. Specifically, we enhance the ResNeXt50 architecture through the integration of a channel attention mechanism, enabling adaptive recalibration of feature responses to emphasize defect-relevant information and improve feature representation capabilities. Experimental evaluations demonstrate that our method achieves an F1 score of 88.0%, demonstrating effective classification performance across common concrete defect categories.