Scholarlink

<p>Crack detection is crucial for infrastructure maintenance, yet it poses significant visual challenges due to the subtle and irregular nature of cracks. Traditional deep learning models, while effective, are often too complex for deployment on resource-constrained devices. This paper introduces a Memory-Aware Dynamic Distillation (MDD) framework to address this gap. By leveraging an unsupervised difficulty assessment strategy, a Gradient-Aware Memory Bank, and adaptive review scheduling, MDD effectively transfers knowledge from a large teacher model to a lightweight student model. Experimental results on the Crack Segmentation dataset demonstrate that MDD achieves a 5% improvement in mAP50 over standard knowledge distillation, with a 95.1% reduction in parameters and a 1.9-fold increase in inference speed. Here, we show that MDD significantly mitigates catastrophic forgetting, enhancing the applicability of lightweight detection models in real-world scenarios. The code and dataset are available at <a href="https://github.com/Jiaoyut/MDD-Crack-Detection.git">https://github.com/Jiaoyut/MDD-Crack-Detection.git</a> and <a href="https://doi.org/10.5281/zenodo.18059674">https://doi.org/10.5281/zenodo.18059674</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing crack detection via memory-aware dynamic knowledge distillation

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • Zhao Liang,
  • Jiao Yutong,
  • Chen Dengfeng,
  • Liu Shipeng

摘要

Crack detection is crucial for infrastructure maintenance, yet it poses significant visual challenges due to the subtle and irregular nature of cracks. Traditional deep learning models, while effective, are often too complex for deployment on resource-constrained devices. This paper introduces a Memory-Aware Dynamic Distillation (MDD) framework to address this gap. By leveraging an unsupervised difficulty assessment strategy, a Gradient-Aware Memory Bank, and adaptive review scheduling, MDD effectively transfers knowledge from a large teacher model to a lightweight student model. Experimental results on the Crack Segmentation dataset demonstrate that MDD achieves a 5% improvement in mAP50 over standard knowledge distillation, with a 95.1% reduction in parameters and a 1.9-fold increase in inference speed. Here, we show that MDD significantly mitigates catastrophic forgetting, enhancing the applicability of lightweight detection models in real-world scenarios. The code and dataset are available at https://github.com/Jiaoyut/MDD-Crack-Detection.git and https://doi.org/10.5281/zenodo.18059674.