ADMA-YOLO: a lightweight strip steel surface defect detection network via adaptive dual-path and multi-branch aggregation

Abstract In the manufacturing process of hot-rolled strip steel, surface defect detection is a key step in ensuring production quality and usage safety. However, there are still many challenges in achieving high-precision and real-time detection with limited computing resources. To address this issue, this paper proposes ADMA-YOLO, a lightweight defect detection model based on YOLOv11n, which integrates Adaptive Dual-path and Multi-branch Aggregation. Firstly, a lightweight adaptive dual-path dynamic interaction network is mainly used for defect feature extraction, in which the cross-stage partial (CSP) path is responsible for local detail features, and the Hierarchical Path focuses on global context modeling. In view of the spatial features extracted from different paths, the adaptive spatial attention gate mechanism is used to realize the dynamic interaction of features extracted from different paths. Secondly, a multi-scale progressive information aggregation module is proposed to progressively extract multi-scale defect features through grouped convolution. Finally, a Global Multi-Branch Feature Pyramid Network is constructed by proposing a cross-level feature fusion mechanism and combining efficient upsampling convolution blocks, CSP multi-scale shift channel mixed convolution and weighted feature fusion, efficient aggregation between high- and low-level information is achieved. The evaluation results on the NEU-DET dataset show that the proposed ADMA-YOLO achieves a mean average precision (mAp50) of 80.4%. In addition, the model has 1.14 M parameters and 3.7 G floating-point operations (FLOPs). Compared with the latest YOLOv11n, the model achieves a 56.9% reduction in parameters, and a 42.3% reduction in FLOPs. The code has been released at https://github.com/LiangbinLi050/ADMA-YOLO .