AGBR-YOLO: an infrared crack detection algorithm based on improved YOLOv11

Abstract In this study, we introduce an infrared crack detection algorithm, AGBR-YOLO, developed to overcome the limitations of visible-light methods. The algorithm integrates GSConv (Group Shuffle Convolution) modules designed for AGBlock (Adaptive Group Block) integration, which enables the network to better respond to multi-dimensional cracks and different fault lines. A new neck module, containing TMFusion (Triple Multi-scale) Fusion and RepBlock (Re-parameterization Block), was also developed to jointly extract and optimize multi-dimensional features. Additionally, we use the MPCA (Multi-path Coordinate Attention) mechanism to focus on key features, while the Inner-IoU loss function improves bounding box accuracy. Experimental results show that AGBR-YOLO achieved a crack detection rate of 84.7% in infrared images, a 2.9% increase over the base model YOLOv11. The model maintains a low parameter count of 2.61M and a GFLOPs of 6.9, while achieving a high inference speed of 272 FPS.