A Sensitivity-Aware Mixed-Precision Quantization Framework for Vision Mamba in Steel Surface Defect Detection

Deep learning foundation models such as Vision Mamba have shown remarkable performance in industrial visual inspection and Prognostics and Health Management (PHM). However, deploying these models on resource‑constrained edge devices remains challenging due to their high memory and computational demands. Traditional Post‑Training Quantization (PTQ) methods struggle to preserve accuracy—especially on rare defect samples commonly found in real‑world PHM.To address these limitations, this paper introduces Active‑Quant, a novel sensitivity‑aware mixed‑precision quantization framework specifically designed for Vision Mamba. The method features two key innovations: (۱) an entropy‑based active learning strategy that actively selects the most informative high‑entropy samples for calibration, enabling PTQ to maintain accuracy on rare and hard defects; and (۲) a new layer‑wise sensitivity analysis tailored to the architecture of Vision Mamba, allowing adaptive bit‑width assignment across Mamba blocks instead of using uniform quantization.To the best of our knowledge, this is the first comprehensive study that investigates quantization of Vision Mamba for industrial PHM and edge deployment. Experimental results on the NEU‑DET surface‑defect dataset demonstrate that Active‑Quant achieves a ۲.۶۷× compression ratio, a reconstruction error of only ۰.۰۱۱ (MSE), and less than ۱% accuracy drop compared to the full‑precision model, significantly outperforming conventional uniform PTQ baselines. These findings highlight the practical value of Active‑Quant in enabling efficient and reliable deployment of Vision Mamba models on edge devices with limited resources.

کلیدواژه ها:

Vision Mamba ، Active Learning ، Sensitivity Analysis ، Industrial PHM ، Edge AI