Abstract: A lightweight YOLOv8n method addresses high missed and false detection rates of complex hot-rolled steel strip defects and edge deployment difficulties. An E-IDC module integrates Inception depthwise convolution and channel attention to enhance subtle defect perception while controlling computational complexity. A C2f-DCNv2 module employs deformable convolution to adaptively adjust sampling distributions, improving irregular contour fitting accuracy. The WIoUv3 loss function suppresses low-quality sample interference via a dynamic focusing mechanism to stabilize training. NEU-DET experiments show a 4.4% mAP@0.5 increase, with 7% fewer parameters and 16% lower computational costs. Compared with mainstream methods, the model effectively enhances complex defect detection accuracy under lower computational overhead.