Mask-Guided Gaussian Occlusion Segmentation via Multi-EM Inference: An ECE Pipeline with Azure-Orchestrated MLOps

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This paper presents a probabilistic approach to binary image segmentation that models foreground-background occlusion with pixel-wise Gaussian emissions and mask-conditioned independence, comparing iterated conditional modes, exact expectation maximization, Gibbs-sampled EM, and variational mean-field EM for learning and inference under a free-energy objective grounded in KL minimization. The dataset is synthesized by compositing masked foreground exemplars over background textures and injecting Gaussian sensor noise, enabling controlled evaluation of convergence behavior and parameter recovery for class priors, per-pixel mask probabilities, and emission statistics. An engineering-focused pipeline is outlined for operationalizing the segmentation workflow with Azure services and DevOps: infrastructure-as-code for GPU compute and storage provisioning, CI/CD for training artifacts, reproducible runs and model registry integration, and experiment tracking for algorithm variants; this demonstrates how classical probabilistic graphical models can be productionized alongside modern MLOps while preserving explainability and modular update rules. Results highlight trade-offs between hard assignments and uncertainty-aware posteriors, noting local optima sensitivity in ICM and improved robustness from stochastic sampling and variational factorization, and motivate deployment patterns that separate data synthesis, inference kernels, and monitoring to support maintainable computer engineering systems.