Distribution-free online change detection for low-rank images

Tingnan Gong, Seong‐Hee Kim, Yao XieSequential 2025 · DOI

TL;DR

This paper introduces DFLIM, a novel distribution-free method for detecting changes in sequences of low-rank images. It addresses the challenges of high dimensionality, temporal dependence, and non-normality in image data, offering a robust solution for online monitoring in industrial applications.

What Problem Does This Paper Address?

Modern manufacturing processes generate vast amounts of image data at high rates. Monitoring this data for anomalies or shifts in quality presents several challenges:

  • High Dimensionality: Images are inherently high-dimensional, making traditional statistical methods computationally expensive and difficult to apply.
  • Temporal Dependence: Consecutive images often exhibit weak or even strong autocorrelation, violating the independence assumptions of many monitoring techniques.
  • Spatial Correlation: Pixels within an image are not independent; their relationships can be complex and change over time, especially with anomalies.
  • Non-Normality: The underlying data distributions may not be Gaussian, which can lead to poor performance of methods that assume normality.
  • Low-Rank Structure: While many image datasets exhibit a low-rank structure (meaning they can be represented by a smaller set of underlying components), this property needs to be effectively leveraged.

Existing methods often struggle with one or more of these issues, leading to missed detections or false alarms. This paper aims to develop a method that can handle these complexities simultaneously for online change detection in image sequences.

Methods

The core of the DFLIM procedure is a distribution-free CUSUM (Cumulative Sum) chart adapted for low-rank image data. Here's a breakdown of the key techniques:

  • Low-Rank Matrix Decomposition: The method leverages the low-rank property of the in-control mean image matrix ($M_0$). Singular Value Decomposition (SVD) is used to extract the principal components (singular vectors) of $M_0$.
  • Two Types of Projection Statistics: DFLIM constructs monitoring statistics ($y_t$) by combining two types of projections of the observed image matrix ($X_t$):
  • Static Projections ($\beta_{i,t}$): These are projections of $X_t$ onto the static directions (left and right singular vectors) of the in-control mean matrix $M_0$. These capture shifts that are similar in structure to the in-control mean.
  • Dynamic Projections ($Y_{i,t}$): These are derived from the singular values of the residual matrix ($R_t = X_t - M_0$). These capture shifts that might be orthogonal to the in-control mean's structure and are particularly useful for detecting anomalies in high-dimensional settings.
  • T²-type Statistics: The combined projection statistics ($y_t$) are used to compute a Hotelling's T²-type statistic ($T_t$), which is a standard multivariate control statistic.
  • CUSUM Recursion: The $T_t$ statistics are then fed into a CUSUM recursion to create a monitoring statistic ($S_t$). CUSUM charts are known for their sensitivity to small, persistent shifts.
  • Distribution-Free Property: The procedure is designed to be distribution-free, meaning it does not assume a specific parametric distribution for the image data. This is achieved by relying on the Functional Central Limit Theorem (FCLT) for the CUSUM statistics, which allows for an approximate analysis of the Average Run Length (ARL) regardless of the underlying data distribution.
  • Analytical Control Limit Determination: The control limit ($H$) is determined analytically by solving an equation derived from the approximate ARL expressions, ensuring a target false alarm rate.

Datasets

The paper uses the following for evaluation:

  • Simulated Data: Extensive simulations are conducted using various settings for:
  • Distributions: Normal and non-normal (exponential) distributions for noise.
  • Ranks: Low ranks (2 and 5) for the in-control mean matrix ($M_0$).
  • Covariance Structures: Tri-diagonal and exponential covariance structures to model spatial correlations.
  • Autocorrelation: Moving average models with lag orders 5 and 20 for temporal dependence.
  • Shift Patterns: Four distinct patterns for the mean shift matrix ($A$): sparse, ring, sine, and chessboard.
  • Real Data:
    • Solar Flare Images: A sequence of 300 images (232x292 pixels) capturing solar flare outbursts. The first 150 images are used for training.
  • 3D Printing Textured Surface Images: Images from a 3D printing process (250x250 pixels) with different bead orientations (45° and 135°). A smaller subset of images (around 40) is used for training.

Key Insights and Ideas

  • Synergy of Static and Dynamic Features: The paper convincingly argues that combining projections based on the in-control mean's structure (static) with projections from the residual matrix (dynamic) is crucial. Static features excel when the shift is similar to the in-control mean (e.g., chessboard shift), while dynamic features are vital for detecting shifts that are structurally different or in high-dimensional settings.
  • Robustness to Non-Normality and Temporal Dependence: The distribution-free nature, enabled by the FCLT, allows DFLIM to maintain good performance even when data is non-normal or exhibits temporal correlations, which are common in real-world industrial data.
  • Analytical Control Limit Setting: The ability to analytically determine control limits based on target ARLs is a significant practical advantage, reducing the need for extensive trial-and-error calibration.
  • Leveraging Low-Rank Structure: Effectively utilizing the low-rank property of the in-control mean matrix is key to dimensionality reduction and creating sensitive monitoring statistics.

Key Results

  • Superior Performance: DFLIM consistently outperforms baseline methods (MEWMA, MGLR, ST-SSD) in terms of detecting shifts, especially for sparse and ring shifts, and generally achieves lower ARL₁ values (faster detection).
  • Accurate In-Control Performance: The analytically determined control limits for DFLIM result in estimated ARL₀ values very close to the target (e.g., 200), demonstrating the effectiveness of the theoretical framework.
  • Effective on Real Data: DFLIM successfully detects solar flare outbursts and defects in 3D printing processes, often with earlier detection and fewer false alarms compared to ST-SSD. For instance, in the solar flare data, DFLIM detects outbursts around t=190 and t=220 with statistics well away from the control limit, while ST-SSD's statistics are closer to its limit. In 3D printing, DFLIM detects defects without false alarms at build-to-build transitions, unlike ST-SSD.

Limitations and Future Work

  • Assumption of Low-Rank: The current method assumes the in-control mean matrix is low-rank. The authors acknowledge that in practice, the estimate of the in-control mean might deviate from this structure due to noise. This could lead to overestimating the rank and introducing spurious components, potentially increasing false alarms or missed detections.
  • Future Work:
    • Approximately Low-Rank Matrices: Extending the method to handle matrices that are only approximately low-rank is a key future direction. This would involve analyzing the impact of deviations from perfect low-rank structure and potentially guiding a more appropriate selection of the rank parameter $r$.
  • Extension to Other Matrix Structures: While the paper focuses on low-rank matrices, exploring extensions for other matrix structures could broaden its applicability.

Summary updated

BibTeX
@inproceedings{gong2025distribution,
  author = {Tingnan Gong and Seong‐Hee Kim and Yao Xie},
  title = {{Distribution-free online change detection for low-rank images}},
  booktitle = {Sequential},
  year = {2025},
  doi = {10.1080/07474946.2025.2481558},
}
Figure