Abstract: To address the issues of large parameter count, high computational overhead, and insufficient detection accuracy for small-target defects in steel surface defect detection models, a lightweight and high-precision steel surface defect detection model is proposed based on the YOLOv8 baseline. In the backbone network, the C2f module is replaced with the C2f_Ghost module based on Ghost convolution, which leverages the generation mechanism of primary features plus "ghost" features to reduce the total number of parameters by approximately 16.7%. In the neck feature fusion network, the VoVGSCSP module is introduced, integrating the GSConv channel-shuffle mechanism with CSP cross-stage partial connections, which reduces the computational cost of the neck by approximately 35%. Meanwhile, channel lightweighting operations are performed on the neck network. Additionally, Coordinate Attention modules are embedded in the P3 and P4 output layers of the neck to enhance the detection capability for directional defects such as scratches and rolled-in scale. Experimental results on the NEU-DET steel surface defect dataset from Northeastern University show that the proposed model achieves an mAP@0.5 of 77.4%, with 5.9 GFLOPs, 2.2M parameters, and a weight file size of only 4.6 MB. Real-machine tests on an NVIDIA Jetson Orin Nano edge inference device showed that the model achieved an inference speed of 79.6 frames/s and a single-frame latency of approximately 12.6 ms. In terms of real-time performance indicators, including inference latency, power consumption, and edge-device resource utilization, the model meets the requirements for edge deployment in in-line inspection on hot-rolling production lines. It should be noted that the accuracy metrics reported in this study were obtained on public benchmark datasets such as NEU-DET; the model’s detection performance on images collected from actual production lines remains to be further validated.