A lightweight model for steel surface defect detection based on EDGI-YOLO

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

To address the limitations of YOLOv8n in steel surface defect detection–namely, its large parameter count and computational cost, insufficient multi-scale feature modeling capability, and poor scale adaptability of the detection head– we propose a lightweight improved model named EDGI-YOLO. The model is restructured along four dimensions. First, an ECA-GC module that integrates efficient channel attention with global context modulation is embedded to enhance the semantic representation of minute defects. Second, a depthwise separable convolution module DSEConv is designed to decouple spatial filtering from channel transformation while incorporating SE attention, reducing computation while preserving critical texture information. Third, an AllGhostC2f module with full-link Ghost convolution optimization replaces the original C2f structure, generating redundant features via cheap linear transformations to significantly lower parameters and computational overhead. Fourth, an IAFHead independent adaptive fusion detection head is constructed, using DSEConv to reduce branch load, CrossScaleSE for global channel fusion across three scales, and asymmetric depth configuration to adapt to scale-specific modeling needs. Experimental results on the NEU-DET dataset show that EDGI-YOLO achieves an mAP@0.5 of 78.85%, a 3.94 percentage point improvement over YOLOv8n, while compressing parameters to 1.66M, FLOPs to 6.0G, and model weight to 3.63M. Cross-dataset validation on GC10-DET further confirms its generalization, and edge deployment on a representative platform achieves up to 159 FPS with TensorRT acceleration. The model offers a high-accuracy, lightweight solution for resource-constrained industrial edge devices. The model achieves an excellent balance between detection accuracy and computational efficiency, offering a high-accuracy, lightweight solution for resource-constrained industrial edge devices.