Real-time multi-object detection model for cracks and deformations based on deep learning

Abstract

The paper proposes a deep learning-based multi-object real-time detection model for concrete cracks and structural deformations. The model improves the single-stage object detection framework, You Only Look Once version 7 (YOLOv7), by incorporating convolutional block attention mechanisms and global attention mechanisms into its backbone and neck networks, respectively. It also establishes dual output branches for cracks and deformations within the output module to enable multi-object detection capabilities. Utilizing transfer learning strategies, the model effectively detects concrete cracks and structural deformations with a limited dataset. The results demonstrate that the improved YOLOv7 model significantly improves the detection of non-continuous cracks and reduces noise in complex environments, indicating strong generalization and robustness. The improved model exhibits a 4.53% increase in crack detection accuracy over the original and achieves low peak relative errors for deflection deformation in concrete beams and in-plane deformation in concrete slabs at just 0.22% and 3.05%, respectively.