IMPACT-Net: An integrated multi-scale and computation-efficient timely network for surface defect detection in industrial embedded systems

Automated defect detection is crucial in industrial production, with typical scenarios in steel, automotive, and new energy manufacturing. Although deep learning-based defect detection methods have achieved significant progress, challenges such as limited detection accuracy for low-contrast defects and difficulties in efficient inference still persist. To address these challenges, this paper proposes an integrated multi-scale and computation-efficient timely network (IMPACT-Net) for surface defect detection. Firstly, a Precision-Enhanced Feature Pyramid Network (PE-FPN) is designed to improve the detection performance for low-contrast and fine defects by enhancing multi-scale feature fusion. Secondly, an Adaptive Normalized Wasserstein Distance Loss (ANWD-Loss) is proposed to optimize bounding box localization accuracy and enhance robustness. Finally, by employing Progressive Block-Freezing Architecture Search (PBF-AS) and a ZYNQ-based acceleration platform, computational complexity is significantly reduced, and efficient inference is achieved under low-power conditions. Experimental results show that the proposed IMPACT-Net achieves an mAP50 of 78.9 % on the NEU-DET dataset with 2.3 ms inference time and 71.4 % mAP50 on the GC10-DET dataset with 2.5 ms inference time, demonstrating a good balance between detection accuracy and real-time performance that is well suited for resource-constrained embedded industrial environments.