[Paper Review] Sequential Adaptive Detection for In-Situ Transmission Electron Microscopy (TEM)

Yang Cao, Shixiang Zhu+5 more|arXiv (Cornell University)|Oct 31, 2017

Sparse and Compressive Sensing Techniques8 references3 citations

TL;DR

This paper proposes sequential adaptive detection algorithms—adaptive CUSUM (ACM) and adaptive SRRS (ASR)—for real-time detection of sparse, transient signals in in-situ transmission electron microscopy (TEM) video sequences. By combining sequential change-point detection with online convex optimization and $ε$-regularized maximum likelihood estimation, the method enables fast, memory-efficient, and robust detection of weak, sparse mean shifts in high-dimensional Gaussian data, achieving a detection delay of only 1 sample in real TEM data, outperforming classic CUSUM and GLR procedures.

ABSTRACT

We develop new efficient online algorithms for detecting transient sparse signals in TEM video sequences, by adopting the recently developed framework for sequential detection jointly with online convex optimization [1]. We cast the problem as detecting an unknown sparse mean shift of Gaussian observations, and develop adaptive CUSUM and adaptive SSRS procedures, which are based on likelihood ratio statistics with post-change mean vector being online maximum likelihood estimators with $\ell_1$. We demonstrate the meritorious performance of our algorithms for TEM imaging using real data.

Motivation & Objective

• To address the challenge of real-time detection of transient, sparse signals in high-speed in-situ TEM video sequences.
• To overcome the limitations of classical CUSUM and GLR procedures, which suffer from poor performance under parameter misspecification or high sensitivity to noise.
• To develop an online, recursive algorithm that adapts to unknown post-change mean parameters in high-dimensional, sparse Gaussian data.
• To enable in-situ, real-time analysis in TEM by automating detection of emerging structural features without requiring offline data storage or pre-specification of signal parameters.

Proposed method

• Formulates the detection problem as identifying a sparse mean shift in high-dimensional i.i.d. Gaussian observations, with pre-change mean zero and post-change mean in a sparse set $\mathcal{A} = \{\theta : \|\theta\|_0 \leq s\}$.
• Develops adaptive CUSUM (ACM) and adaptive SRRS (ASR) procedures based on likelihood ratio statistics with online maximum likelihood estimators using $\ell_1$-regularization.
• Employs a one-sample update scheme to recursively compute post-change mean estimates $\hat{\theta}_{k,i}$ via online convex optimization, avoiding storage of raw data.
• Applies window-limited detection by maximizing or summing over a sliding window of size $w$ to control memory usage and prevent infinite memory accumulation.
• Uses polar coordinate transformation to reduce 2D diffraction images to 360-dimensional angular signals at fixed radii, enabling detection of weak, localized bright spots.
• Performs preprocessing via histogram analysis, thresholding, Hough transform, and Canny edge detection to remove background rings and shadows before signal extraction.

Experimental results

Research questions

• RQ1 Can sequential adaptive detection methods detect sparse, transient signals in high-dimensional in-situ TEM data faster and more reliably than classical CUSUM and GLR procedures?
• RQ2 How can online convex optimization with $\ell_1$-regularization improve the robustness and efficiency of change-point detection in high-dimensional, sparse signal scenarios?
• RQ3 What preprocessing pipeline enables reliable extraction of weak, localized signals (e.g., bright spots) from noisy, ring-structured diffraction images in TEM?
• RQ4 How does the performance of adaptive detection algorithms compare to classical methods in terms of detection delay and false alarm rate on real TEM data?
• RQ5 Can the proposed method achieve real-time, in-situ detection without storing raw data, while maintaining high sensitivity to sparse, transient features?

Key findings

• The proposed ACM and ASR procedures detected the true change-point at time $t=18$, corresponding to a detection delay of only 1 sample, which matches the true change at $t=17$.
• The classic CUSUM procedure with an all-one post-change vector detected at $t=24$, resulting in a 7-sample delay, indicating poor performance due to parameter misspecification.
• The GLR procedure raised a false alarm by detecting at $t=4$, demonstrating high sensitivity to noise and lack of robustness in real-world settings.
• The adaptive algorithms achieved superior performance by dynamically estimating the post-change mean using online $\ell_1$-regularized MLE, improving robustness and detection speed.
• Preprocessing using histogram analysis, thresholding, Hough transform, and Canny edge detection successfully isolated and removed background rings and shadows, enabling clear detection of weak, localized bright spots.
• The method enables real-time, in-situ detection with minimal memory usage, as raw data are not stored—only one-sample updates are used for parameter estimation.