Yapısal sağlık izleme için çatlak segmentasyonu için hafif bir derin öğrenme mimarisi

Özet (abstract)

Automated crack segmentation is a key component of vision-based structural health monitoring for civil infrastructure. However, many deep learning approaches remain computationally demanding for embedded edge devices and exhibit limited generalisation across diverse materials and imaging conditions. This paper presents a lightweight deep learning architecture designed to balance segmentation accuracy, computational efficiency, and cross-domain robustness. The artificial intelligence contribution is an efficient encoder–decoder integrating depthwise separable convolutions, lightweight multi-scale context aggregation, and channel attention. The proposed approach targets crack continuity through a unified design for thin-structure segmentation under strict computational constraints. The engineering contribution is a practical inspection method for structural health monitoring systems with limited memory, latency, and power budgets. Experiments on a joint dataset comprising 30,017 crack images collected from 20 heterogeneous datasets demonstrate competitive performance across diverse conditions. The proposed Lightweight Attention Channel-Enhanced Crack Segmentation (LACE-Crack) architecture achieves a higher harmonic mean of precision and recall (F1 score) (0.569 vs. 0.566), precision (0.751 vs. 0.732), mean intersection over union (0.694 vs. 0.693), optimal image scale (0.489 vs. 0.480), and optimal dataset scale (0.425 vs. 0.411) than the comparison architecture, with comparable recall (0.459 vs. 0.462). It also requires approximately 43% fewer parameters and 64% fewer floating-point operations. These results demonstrate a strong balance between crack segmentation performance and computational efficiency, supporting the practical use of lightweight artificial intelligence for infrastructure inspection under deployment constraints.

Yazarların özeti; kaynağından alınmıştır. Engineering Applications of Artificial Intelligence, 2026 · DOI ↗