Surface defects in steel manufacturing significantly degrade product quality and safety, making accurate and efficient defect detection essential for modern industrial production. Existing deep learning-based detection models often struggle with balancing computational complexity, accuracy, and real-time performance, especially in resource-constrained industrial scenarios. To address these challenges, this paper proposes SLI-Net, an improved detection network based on YOLOv8, specifically designed for efficient and precise detection of complex steel surface defects. Firstly, we designed the Spatial Channel Redundancy (SCR) module to eliminate redundant information, enhance feature representation and reduce model parameters. Secondly, we design a Lightweight Continuous Adaptive Convolution (LCAconv) to dynamically adjust the receptive field, significantly enhancing detection accuracy for defects with diverse and elongated shapes while maintaining a lightweight model architecture. Thirdly, an Inner-WIoUv3 loss function is proposed, integrating adaptive scale auxiliary bounding boxes and dynamic non-monotonic focusing mechanisms to accelerate convergence and improve regression accuracy. Comprehensive experiments on the NEU-DET and GC10-DET datasets demonstrate that the proposed SLI-Net achieves mAP values of 82.1% and 72.4% respectively with only 2.6 million parameters. This delivers near state-of-the-art accuracy while exhibiting superior parameter efficiency and computational effectiveness. The results confirm that SLI-Net effectively balances high accuracy and computational efficiency, making it ideal for industrial applications with limited computational resources.