<![CDATA[This study introduces an innovative approach for detecting cracks on levee surfaces by integrating an Auxiliary Classifier Generative Adversarial Network (ACGAN) for data augmentation with a Faster R-CNN model enhanced by an attention mechanism. The ACGAN-based augmentation aims to generate synthetic images that enrich data variability in the original dataset. The attention-optimized Faster R-CNN is designed to improve detection precision, particularly for small objects and fine cracks that are difficult to distinguish from the background. Experimental results demonstrate that the incorporation of ACGAN improves detection performance, increasing both the mean Average Precision (mAP) and Average Recall (AR). The model achieved an mAP of approximately 0.56 at IoU = 0.50 and 0.34 at IoU = 0.75, while the AR (maxDets = 100) reached 0.55, indicating a strong capability in identifying most crack instances. When trained on the combined dataset of original and synthetic images, the Faster R-CNN model reached a precision of 0.92 for the severe crack class, while performance for minor cracks remained lower (precision 0.78). Adjusting the confidence threshold to 0.65 improved detection reliability by reducing noise and retaining high-confidence predictions. Improved performance in detecting severe cracks supports timely maintenance and repair decisions. This study demonstrates the effectiveness of GAN-based data augmentation and attention-enhanced object detection for automated structural health monitoring (SHM) of levee infrastructure]]>
Copyrights © 2026
