| 摘 要: 为了提升图像异常检测模型对复杂异常的识别率以及对多尺寸异常的检测稳定性,提出了基于多组师生网络分层特征融合框架的异常检测模型(Multi-group Teacher-student Network Hierarchical Feature Fusion Framework,MTHF)。通过构建3组不同的师生检测模块,在每组学生网络的蒸馏层后引入特征融合模块,提升多尺度异常检测精度,采用两阶段融合策略实现异常区域全面覆盖的同时保证边缘清晰。在 MVTecAD数据集上进行实验,结果表明,MTHF模型的平均检测精度相比当前最先进方法提升了7.62%,充分证明了模型在图像异常检测任务中的有效性。 |
| 关键词: 异常检测 知识蒸馏 特征融合 注意力机制 |
|
中图分类号:
文献标识码: A
|
| 基金项目: 国家自然科学基金项目(32360437);甘肃省高等学校产业支撑计划项目(2021CYZC-57) |
|
| Anomaly Detection Model Based on Hierarchical Feature Fusion Framework of Multi-group Teacher-student Network |
|
DONG Haoran1, YE Ning1,ZHOU Xiaoliang2
|
(1. College of information science and technology & Artifical Intelligence, Nanjing Forestry University, Nanjing 210037, China;
2. School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing 210094, China)
1626835594@qq.com; yening@njfu.edu.cn; xiaoliangzhou@njust.edu.cn
|
| Abstract: To enhance the recognition rate of image anomaly detection models for complex anomalies and improve detection stability across multi scale anomalies, this paper proposes an anomaly detection model based on a Multi group Teache-r student Network Hierarchical Feature Fusion Framework (MTHF). By constructing three distinct teacher student detection modules and introducing feature fusion modules after the distillation layers of each student network, the model enhances multi scale anomaly detection accuracy. A two-stage fusion strategy achieves comprehensive coverage of anomalous regions while preserving sharp boundaries. Experiments on the MVTec AD dataset demonstrate that the MTHF model improves the average detection accuracy by 7.62% compared to state-o-f the-art methods, validating its effectiveness in image anomaly detection tasks. |
| Keywords: anomaly detection knowledge distillation feature fusion attention mechanism |
