YOLO-AMC: An Improved YOLO Architecture with Attention Mechanisms for Building Crack Detection

Researchers have proposed YOLO-AMC, a modified YOLO architecture that integrates attention mechanisms to improve automated detection of building cracks, a task complicated by thin, low-contrast structures and background noise [1]. The model builds on YOLOv11 by removing the original C2PSA module and inserting multiple attention mechanisms — Global Attention Mechanism (GAM), Residual Convolutional Block Attention Module (Res-CBAM), and Shuffle Attention (SA) — into the multi-scale feature fusion layers of the Neck [1]. The goal is to strengthen cross-scale feature integration for infrastructure inspection and Structural Health Monitoring [1]. In testing, YOLO-AMC consistently outperformed baseline models YOLOv11n and YOLOv8n [1]. The GAM variant delivered the highest scores, recording a [email protected] of 0.9917 and [email protected]:0.95 of 0.9506, compared with 0.9833 and 0.9112 for YOLOv11, and 0.9707 and 0.8921 for YOLOv8 [1]. The architecture maintains a computational complexity of 7.6 GFLOPs and reaches 110.95 frames per second on an NVIDIA RTX 4090 platform, while running at approximately 5 FPS on a Raspberry Pi 5 edge device [1]. Attention mechanisms have become central to modern neural architectures, powering systems from large language models to autonomous vehicles [4]. However, a recent empirical study of 555 real-world faults across 96 projects found that over half of attention-based neural network failures stem from mechanisms unique to attention architectures, highlighting the need for diagnostic heuristics that current fault taxonomies do not provide [4]. The YOLO-AMC code has been released on GitHub, continuing a pattern of open-source dissemination common in computer-vision research [1]. Open-source AI software spans libraries, frameworks, and tools for deep learning, computer vision, and natural language processing, with projects such as Stable Diffusion making model weights publicly available to run on consumer hardware [7][8]. IBM similarly opened the source code for some of its Granite foundation models, which were trained on curated datasets including internet content, academic publications, and legal and finance documents [9].