MSFlow:用于无监督异常检测的多尺度基于流的框架
IEEE transactions on neural networks and learning systems
|January 9, 2024
概括
本研究介绍了MSFlow,这是一种用于无监督异常检测和定位的新型多尺度框架. 它有效地处理不同的异常大小,在基准指标上取得最先进的结果.
科学领域:
- 计算机视觉 计算机视觉
- 机器学习 机器学习
- 统计建模 统计建模
背景情况:
- 无监督异常检测 (UAD) 对于只有正常数据可用于培训的应用至关重要.
- 现有的方法在不同的异常大小中扎,影响检测和定位精度.
- 规范化流提供了一个概率方法来检测异常,但面临的挑战是尺度变化.
研究的目的:
- 开发一个强大的无监督异常检测和本地化框架,以应对不同异常大小的挑战.
- 提高基于流量模型在没有先前异常信息的情况下区分和确定异常的性能.
- 引入一种新的多尺度方法,以提高异常检测和定位精度.
主要方法:
- 提出了一种基于多尺度流的新型框架 (MSFlow),利用不对称的并行流和融合流.
- 实施了针对图像智能异常检测和像素智能定位的多尺度聚合策略.
- 利用规范化流量的概率性来为异常数据点赋予低概率.
主要成果:
- 在三个异常检测数据集中,MSFlow显著优于现有方法.
- 在MVTec AD基准指标上实现了最先进的 (SOTA) 性能.
- 在MVTec AD.上达到99.7%的检测AUORC,98.8%的本地化AUCROC和97.1%的PRO得分.
结论:
- 拟议的MSFlow框架有效地将不同异常大小的异常进行一般化,用于无监督的异常检测和定位.
- MSFlow代表了基于流量的异常检测的重大进步,并设定了新的性能基准.
- 量身定制的多尺度聚合策略有助于在检测和本地化任务中实现框架的卓越性能.
相关概念视频
01:23Flow Cytometry
13.0K
The development of flow cytometry techniques began in 1934 with initial attempts by Andrew Moldavan, a bacteriologist who counted the cells in a flowing capillary system. Moldavan pumped cells through a capillary tube focused under a microscope for visualization. The invention of photometry allowed the measurement of differentially-stained cells, and Louis Kamentsky developed the first multiparameter flow cytometer in 1965 to identify and count the cancer cells in cervical tissue specimens.
In...
In...
13.0K
01:23Introduction to Types of Flows
1.2K
Fluid flows are categorized by dimensionality and behavior, with one-dimensional flow being the simplest form, where properties like velocity and pressure change only along a single axis. Water moving through straight pipes exemplifies this flow type, as variations in other directions are minimal. One-dimensional analysis helps simplify understanding such flows, focusing solely on changes along the pipe's length.
Two-dimensional flow involves changes in both length and height, as seen in...
Two-dimensional flow involves changes in both length and height, as seen in...
1.2K