Pareto Optimal Weight Learning and Gradient Anisotropic Supervoxel Segmentation for Thermo-Geometric Point Clouds.

Area of Science:

• Materials Science and Engineering
• Computer Vision and Image Analysis
• Industrial Process Monitoring

Background:

• Accurate analysis of geometric and thermal data is crucial for high-temperature industrial inspection.
• Conventional supervoxel segmentation methods face challenges with heterogeneous sensor data (spatial vs. thermal).
• Existing fixed-weighting schemes fail to adequately address the discrepancies between different sensor modalities.

Purpose of the Study:

• To develop an advanced segmentation framework for thermo-geometric point clouds.
• To enhance the analysis of heterogeneous sensing modalities in industrial inspection.
• To improve the accuracy of defect detection by integrating geometric and thermal information.

Main Methods:

• Implemented a Pareto-optimal weight learning approach using multi-objective evolutionary optimization.
• Developed a gradient-anisotropic supervoxel generation algorithm with a local saliency factor.
• Utilized a gradient damping mechanism for improved thermal boundary adherence.
• Employed a region-growing method with optimized multi-sensor fusion weights for supervoxel merging.

Main Results:

• Achieved high-fidelity thermal segmentation and superior multi-modal boundary preservation.
• Demonstrated outperformance compared to traditional segmentation baselines.
• Successfully accommodated spatial-thermal inconsistencies with a controlled compromise in geometric compactness.

Conclusions:

• The proposed framework effectively integrates geometric and thermal data for industrial inspection.
• Pareto-optimal weight learning and gradient anisotropy enable adaptive and fine-grained segmentation.
• The method offers a robust solution for analyzing complex thermo-geometric data in demanding environments.