Abstract <p>Hot rolled steel strips play an important role in the fields of construction, automobile manufacturing, energy, shipbuilding and petrochemicals, etc. Their high strength, corrosion resistance and plasticity make them an indispensable material in industrial manufacturing. Surface defect detection is an indispensable process in hot rolling production line, which is of great significance to improve the quality of hot rolled steel strip. The current detection accuracy of small target defects on the surface of hot rolled steel strips is low and cannot meet the real-time detection needs of enterprises. To solve this problem, we propose a steel strip surface defect detection method based on YOLOv8, named TBD-YOLO. First, the downsampling mechanism from YOLOv7 (V7downsample) is referenced to replace the downsampling modules in the backbone and neck networks to enhance detection accuracy. Second, a modified bidirectional feature pyramid network (mod_BiFPN) is designed for the neck to perform weighted fusion of multi-scale feature maps. Finally, a novel task-aligned detection head (TDH) is developed to improve the classification and localization performance of the detection head. Extensive experimental results demonstrate that, compared to the original YOLOv8 model, the detection method proposed in this paper has achieved a 7.8% increase in mean Average Precision at Intersection over Union 0.5 (mAP@0.5) value, effectively enhancing the detection capability for small target defects on hot-rolled steel strips surface. Moreover, the frames per second (FPS) has reached 79.8, meeting the real-time detection requirements of industrial sites.</p>