The simultaneous analysis of geometric morphology and thermodynamic states from heterogeneous sensing modalities is essential for high-temperature industrial inspection. While supervoxel segmentation is effective for extracting fine structures, conventional fixed-weighting schemes often struggle with the inherent heterogeneity between spatial sensors and thermal sensors. This paper proposes a segmentation framework for thermo-geometric point clouds based on Pareto-optimal weight learning and gradient anisotropy. A multi-objective evolutionary optimization algorithm is employed for multi-modal Pareto weight learning to adaptively balance geometric and thermal constraints. The developed gradient-anisotropic supervoxel generation algorithm introduces a local saliency factor to achieve fine-grained thermodynamic segmentation. Furthermore, a gradient damping mechanism is implemented to ensure high thermal-boundary adherence even in geometrically planar regions by imposing anisotropic penalty forces. Finally, a region-growing method based on the optimized multi-sensor fusion weights is utilized to merge similar supervoxels. Experimental results demonstrate that our approach outperforms traditional baselines by achieving high-fidelity thermal segmentation and multi-modal boundary preservation, while accepting a modest and necessary compromise in geometric compactness to accommodate spatial–thermal inconsistencies.
Xutong et al. (Wed,) studied this question.