This paper proposes an improved YOLOv8 algorithm for steel surface defect detection. The algorithm enhances the model's ability to detect complex textures and irregular shapes by integrating deformable convolution (DCNv2) into the backbone network. It also replaces the original feature fusion structure with BiFPN to improve multi-scale feature integration and accuracy. The BiFormer attention mechanism is introduced to adaptively allocate attention to target features, and the loss function is replaced with WIoUv3 to enhance bounding box regression and prevent overfitting. Experimental results show that the improved model achieves a mean Average Precision (mAP) of 84.8%, a 6.9% improvement over the original YOLOv8 model. The model can quickly and accurately detect and classify various steel surface defects, meeting industrial production needs. The algorithm is evaluated on the NEU-DET dataset, demonstrating superior performance compared to other detection models like SSD, Fast R-CNN, DETR, YOLOv5s, YOLOv7, and YOLOv8n. The improved model shows significant improvements in detecting small and complex defects, such as cracks and rolled-in scale. Future work includes applying semi-supervised learning to reduce reliance on labeled data and improve generalization for less common defect types.This paper proposes an improved YOLOv8 algorithm for steel surface defect detection. The algorithm enhances the model's ability to detect complex textures and irregular shapes by integrating deformable convolution (DCNv2) into the backbone network. It also replaces the original feature fusion structure with BiFPN to improve multi-scale feature integration and accuracy. The BiFormer attention mechanism is introduced to adaptively allocate attention to target features, and the loss function is replaced with WIoUv3 to enhance bounding box regression and prevent overfitting. Experimental results show that the improved model achieves a mean Average Precision (mAP) of 84.8%, a 6.9% improvement over the original YOLOv8 model. The model can quickly and accurately detect and classify various steel surface defects, meeting industrial production needs. The algorithm is evaluated on the NEU-DET dataset, demonstrating superior performance compared to other detection models like SSD, Fast R-CNN, DETR, YOLOv5s, YOLOv7, and YOLOv8n. The improved model shows significant improvements in detecting small and complex defects, such as cracks and rolled-in scale. Future work includes applying semi-supervised learning to reduce reliance on labeled data and improve generalization for less common defect types.