Drone-Guided Cognitive Feature Aggregation for High-Rise Building Surface Defect Inspection

While deep learning techniques like CNNs are well-established for defect detection, applying them to massive, high-resolution drone imagery remains computationally challenging due to the scarcity of defects amidst vast background noise. To address this effficiency gap, this paper proposes an Integrated Drone-based Inspection System that utilizes saliency-inspired mechanisms to prioritize processing. The system integrates BING objectness for rapid candidate proposal, Low-Rank Active Learning (LAL) for sample-effficient optimization, and an Attention-guided Inspection Path (AIP) to dynamically focus on salient regions. Feature aggregation is performed via manifold learning and classiffied using a kernelized SVM. We introduce a large-scale benchmark dataset of 650,000 images collected over three years. Experimental results demonstrate that the proposed method addresses the trade-off between accuracy and speed effectively. Quantitatively, the system achieves an average classiffication accuracy of 78.3%, surpassing state-of-the-art baselines by 1.1%-4.0%. More signifficantly, it reduces training time to 6.15 hours (a 56% reduction compared to ML-CRNN) and achieves a testing speed of 0.20 seconds per image (3.5× faster than competitive methods). The system is capable of resolving fine-grained structural defects as small as 1.5 mm under standard flight protocols, offering a robust solution for real-time, high-precision infrastructure health monitoring.