Steel surface defect detection is essential for ensuring product quality and production efficiency. Traditional manual inspection methods are inefficient and inaccurate while existing deep learning-based approaches often have high computational costs. This study proposes a lightweight semantic segmentation model for steel surface defect detection, integrating optimized image preprocessing (CLAHE contrast enhancement, noise filtering) and an improved ULite network (axial depthwise convolutions, skip connections, adaptive upsampling). On the NEU-Seg dataset, our model achieves 87.77% mIoU while maintaining only 0.9112M parameters and 211.59 FPS, outperforming state-of-the-art models in both accuracy and efficiency. This approach provides a practical and deployable solution for real-time defect detection in industrial settings. Future work will focus on enhancing domain adaptability and optimizing deployment on edge devices.