BLAST: Bayesian online change-point detection with structured image data

Abstract

The prompt online detection of abrupt changes in image data is essential for timely decision-making in broad applications, from video surveillance to manufacturing quality control. Existing methods, however, face three key challenges. First, the high-dimensional nature of image data introduces computational bottlenecks for efficient real-time monitoring. Second, changes often involve structural image features, e.g., edges, blurs and/or shapes, and ignoring such structure can lead to delayed change detection. Third, existing methods are largely non-Bayesian and thus do not provide a quantification of monitoring uncertainty for confident detection. We address this via a novel Bayesian onLine Structure-Aware change deTection (BLAST) method. BLAST first leverages a deep Gaussian Markov random field prior to elicit desirable image structure from offline reference data. With this prior elicited, BLAST employs a new Bayesian online change-point procedure for image monitoring via its so-called posterior run length distribution. This posterior run length distribution can be computed in an online fashion using $\mathcal{O}(p^2)$ work at each time-step, where $p$ is the number of image pixels; this facilitates scalable Bayesian online monitoring of large images. We demonstrate the effectiveness of BLAST over existing methods in a suite of numerical experiments and in two applications, the first on street scene monitoring and the second on real-time process monitoring for metal additive manufacturing.