An improved YOLOv10-based lightweight multi-scale feature fusion model for road defect detection and its applications

Intelligent road damage detection is critical for ensuring traffic safety and extending the lifespan of roads. However, existing methods struggle to balance high accuracy and real-time performance in complex detection scenarios and resource-constrained environments. To address this issue, this study proposes a lightweight multi-scale feature fusion model based on an improved YOLOv10—GAS-YOLO. The model utilizes a novel lightweight architecture (GSF-ST) designed through a combination of feature generation, asymmetric convolution, and grouped channel shuffling optimization strategies, significantly reducing computational complexity and parameter count while enhancing both global and local feature representation. To improve multi-scale damage detection performance, GAS-YOLO incorporates an improved bidirectional feature pyramid network (BiFPN) and Swin Transformer module. A resolution halving and channel doubling strategy enhances the detection ability of small targets. Moreover, the WiOU loss function further optimizes bounding box regression accuracy, mitigating errors caused by sample imbalance. Channel pruning techniques are applied to achieve secondary lightweight compression of the model, resulting in significant resource savings. Through comparative experiments and ablation analysis with several advanced damage detection models, this study demonstrates a significant performance improvement of GAS-YOLO. Experimental results show that GAS-YOLO exhibits outstanding performance in multi-scale damage detection tasks, with 5.6 M parameters, 8.4GFLOPs of computational complexity, and a model size of only 5.8 MB. Compared to baseline models, detection accuracy improves by 10.8 %, computational complexity is reduced by 2.57 times, and parameter count is reduced by 1.29 times, with an average detection accuracy of 86.5 % and a single image processing time of 6.1 ms. Validation on both public datasets and self-constructed datasets further proves its real-time processing capability while maintaining high accuracy. The GAS-YOLO model proposed in this study not only provides a practical solution for road damage detection in resource-constrained environments but also offers new insights for intelligent management of intelligent transportation and urban infrastructure, with broad application prospects.