ESAD-Net: an efficient steel adaptive detection network for real-time surface defect detection

Steel surface defect detection is a critical component of quality control in the steel industry. Traditional detection methods face technical challenges including low accuracy and poor real-time performance. This paper proposes ESAD-Net (Efficient Steel Adaptive Detection Network), which achieves an average inference latency of 8.7 ms per frame (115.3 FPS) while ensuring high precision. Through the serial integration of RegNet-X lightweight feature extraction and the ATSS adaptive training sample selection algorithm. RegNet-X constructs a lightweight backbone network based on design space optimization theory. It integrates edge enhancement module and multi-scale feature fusion mechanisms. This achieves computational efficiency optimization while ensuring feature representation capability. The ATSS algorithm adopts an adaptive IoU threshold selection strategy targeting geometric feature differences of steel defects. It dynamically adjusts the positive and negative sample selection mechanism. This effectively alleviates sample imbalance issues and significantly improves small target defect detection performance. Experimental results on the NEU-DET steel defect dataset demonstrate that ESAD-Net achieves 87.0% mAP@0.5 detection accuracy and 115.3 FPS processing speed. Compared to the baseline model YOLOv8n, it improves overall detection accuracy by 6.8 percentage points and small target defect detection by an average of 13.2 percentage points. The algorithm increases positive sample utilization from 15.2% to 24.8% and reduces false positive rate to 6.7%. Training convergence efficiency is improved by 29%. Meanwhile, model parameters are reduced by 25% and inference time is shortened by 17%. While performance benchmarking is conducted on an RTX 3070 GPU, the model is designed with high architectural efficiency aimed at potential industrial edge deployment. By minimizing parameter redundancy (2.4M) and computational complexity (2.4 GFLOPs), the framework establishes a theoretical foundation for operational feasibility on resource-constrained edge devices.