To address the challenges posed by limited sample sizes and varying defect sizes on strip-steel surfaces in industrial applications, in this paper, we introduce a small-object detection you-only-look-once (YOLO) network (SODY-Net) specifically designed for such surfaces by machine learning technology. Initially, we build upon the YOLOv5s framework and develop a multiscale path aggregation network that incorporates an attention mechanism to improve the model’s capability to predict across multiple scales. Next, we present an adaptive coordinate-decoupled head module for resolving the conflict between the classification and regression tasks. Finally, we propose a bounding box regression loss function that integrates the Wasserstein distance to enhance detection accuracy for small defects. Experimental results indicate that our SODY-Net surpasses other small-object detection frameworks when evaluated on a few-shot dataset of strip-steel surface defects, making it particularly suitable for defect detection tasks in industrial settings.