An Embedded Marked Point Process Framework for Three-Level Object Population Analysis.

Area of Science:

• Computer Vision
• Machine Learning
• Probabilistic Modeling

Background:

• Extracting complex object structures from images is crucial for various vision applications.
• Conventional methods often lack the flexibility to handle hierarchical relationships and object grouping.
• Existing generalized models for marked point processes (MPP) are frequently domain-specific.

Purpose of the Study:

• To introduce a generalized probabilistic framework for extracting complex hierarchical object structures from digital images.
• To extend conventional marked point process (MPP) models to accommodate object-subobject ensembles and coherent object groups.
• To provide an abstract-level model with clear interfaces for diverse application domains.

Main Methods:

• Extension of conventional marked point process (MPP) models to include parent-child relationships (object-subobject ensembles).
• Bayesian segmentation of object populations to form coherent object groups.
• Development of a global optimization process for multi-layer frameworks considering data, prior knowledge, and object interactions.

Main Results:

• Demonstration of the proposed method in three distinct application areas: remote sensing (built-up area analysis), traffic monitoring (airborne and Lidar data), and optical circuit inspection.
• Publication of a new benchmark database for the evaluated test cases.
• Quantitative evaluation of the model's performance across the different applications.

Conclusions:

• The proposed probabilistic framework offers a generalized and abstract approach for hierarchical object structure extraction.
• The method effectively handles object-subobject relationships and group coherence, outperforming domain-specific approaches.
• The framework's adaptability and performance are validated across diverse real-world computer vision tasks.