Crack Detection Based on Improved YOLOv13 Model with Hybrid Feature Optimization

This paper proposes a novel crack detection approach based on an improved YOLOv13 model. First, a Multi-Scale Retinex (MSR) module is employed for feature enrichment to separate the illumination and reflection components of the images. Then, to enhance multi-scale feature representation and reinforce edge details, both Convolutional Block Attention Module (CBAM) and Multi-scale Edge Enhancement Module (MEEM) modules are integrated into the YOLOv13 backbone, forming a novel improved YOLOv13 architecture to significantly strengthen critical crack feature capture capability. Finally, the proposed crack detection method is validated using the Cracks dataset. Experimental results demonstrate that the proposed approach outperforms in crack detection.