A Lightweight Dual-Stream Enhanced Network With Multilevel Feature Guidance and Hybrid Knowledge Distillation for Steel Surface Defect Detection

The advancement of deep neural networks has significantly improved the precision of strip surface defect detection (SDD). However, current methods typically face difficulties in meeting real-time performance and computational efficiency requirements, particularly in high-speed industrial manufacturing environments with limited computational resources. To address these challenges, a dual-stream enhancement strategy is proposed, resulting in a lightweight dual-stream enhanced network (LDENet) that effectively balances detection accuracy, speed, and resource efficiency. Specifically, a lightweight partial convolution (PConv)-stacked backbone network is introduced to minimize model complexity and improve inference speed. The multilayer coordinated adaptive feature fusion (MCAF) module reduces feature redundancy by adaptively integrating important information across multiple layers. Moreover, a hybrid knowledge distillation (HKD) method improves the lightweight model’s capability by transferring detailed information from a pretrained teacher model. On the Northeastern University defect detection dataset (NEU-DET) dataset, LDENet achieves 81.1 mean average precision (mAP) and 181 frames per second(FPS) with only 1.41 M parameters, demonstrating its suitability for real time, resource-constrained industrial applications.