Objective As the core pillar industry in China, strip steel is very important for automobile manufacturing, construction and other industries. Its surface quality directly affects the product reliability and market value. At present, surface defect detection faces three major challenges: first, the defect scale does not match the network receptive field, and small-size defects are easy to miss detection. Second, continuous down sampling leads to the loss of texture information, and low contrast defects are difficult to identify in complex background. Third, the traditional deep learning model has a large number of parameters and complex calculation, which is difficult to realize real-time processing in the resource constrained industrial line. These deficiencies lead to the decline of detection accuracy and efficiency, and the increase of cost. Therefore, there is an urgent need for a high-performance solution that can take into account the detection accuracy, reasoning speed and lightweight design.

Methods The H-RAMi module is introduced into the backbone network to get through the detail and semantic feature channels by cross level feature calibration, accurately capture the micro defect information, and lay a high-quality feature foundation for subsequent detection; In view of the fact that downsampling is easy to lose key details such as defect edges and textures, which leads to missed detection, DySample dynamic upsampling component is used to reconstruct the feature space association through adaptive weights to complete the short board of previous features. Facing the dual requirements of high precision and low delay for the limited computing power at the industrial end, the detection head is designed with a lightweight structure of Detect_MBconv, which uses deep separable convolution and channel attention coordination to compress the computational overhead and parameters while maintaining the receptive field.

Results and Discussions The experimental results on NEU-DET dataset show that the average accuracy of HDM-YOLO is 81.2%, which is superior to the mainstream lightweight detector in the same parameter range. The ablation experiment results are shown in Table 2, which verifies the independent contribution of each module: H-RAMi improves the mAP by 0.7% through cross-scale feature interaction, dysample improves the small target detection accuracy by 1.4% through fine detail retention, and detect_mbconv reduces the model complexity by 22.3%. Table 3 model comparison experiment results show that compared with the baseline model YOLOv13n, the accuracy of HDM-YOLO is improved by 2.7%, the number of parameters is reduced by 37.7%, the computational complexity is reduced by 38.8%, and the reasoning speed is improved from 63.6 frame/s to 101.8 frame/s, meeting the requirements of industrial real-time detection. The comparison of visualization results is shown in Figure 9, which intuitively shows the advantages of HDM-YOLO in small defect detection. The thermodynamic diagram comparison in Figure 10 further verifies the model's ability to accurately locate the defect area. The synergy of H-RAMi and DySample significantly improves the performance of small defect detection. Through the synergy of deep separable convolution and channel attention, it reduces a lot of computational overhead and parameters while maintaining the receptive field. In addition, table 5 GC10-DET comparative experimental results show that the HDM-YOLO mAP value has increased by 3.2%, and the detection speed has increased by 46.9 frames, showing good generalization ability and engineering application value.

Conclusions HDM-YOLO provides an efficient and lightweight solution for strip surface defect detection, and effectively solves the core challenges of existing methods, such as low detection accuracy of small targets, loss of feature information and lack of real-time. By integrating hierarchical complementary feature mixing module (H-RAMi), dynamic sampling module (DySample) and lightweight detection head (Detect_MBConv), the model achieves an excellent balance between detection accuracy (NEU-DET dataset 81.2% mAP), reasoning speed (101.8 frame/s) and model compactness (1.5 M parameters). Visualization and thermal diagram analysis have confirmed the model's ability to accurately identify and locate the defect area. Its strong generalization ability and modular design on GC10-DET dataset enable it to flexibly adapt to various industrial detection scenarios, including high-speed strip production lines. In addition to steel manufacturing, the lightweight architecture of HDM-YOLO provides a useful reference for the development of computer vision models in other resource constrained industrial scenes. In general, the model has significant industrial deployment potential, which can improve the efficiency of quality control, reduce production costs, and promote the development of lightweight defect detection technology.