Bonfring International Journal of Advances in Image Processing
Online ISSN: 2277-503X | Print ISSN: 2250-1053 | Frequency: 4 Issues/Year
Impact Factor: 0.245 | International Scientific Indexing(ISI) calculate based on International Citation Report(ICR)
A Hierarchical Context Calibration Framework for Multi-Objective Image Segmentation Through Dynamic Neighborhood Consensus and Progressive Feature Interaction
Abstract:
Multi-objective image segmentation (MOCS), which aims at the optimization of competing objectives such as boundary approximation, homogeneity, semantic consistency and computation efficiency of the process, represents a key challenge. Current segmentation methods usually optimize a single scalarized objective, which caused a bias towards the more important criteria in the loss function, and yielded miscalibrated representations at different spatial level. present the Hierarchical Context Calibration (HCC) framework which explicitly models multi-objective trade-offs in terms of three coupled modules: (1) a Dynamic Neighborhood Consensus (DNC), that combines the prediction of multiple, adaptively sized local neighborhoods, and resolves disagreement by a learned consensus weighting, (2) a Progressive Feature Interaction (PFI), that passes information across scale-specific feature streams in a coarse-to-fine schedule, and lets the high-level context calibration the low-level boundary estimate and vice versa, and (3) a hierarchical calibration objective that aligns the confidence of per-pixel predictions against their empirical accuracy across scales while mitigating overconfident errors around ambiguous boundaries. measure and report the performance of HCC on four benchmarks (Cityscapes, ADE20K, COCO-Stuff and PASCAL-Context) following a multi-objective evaluation protocol, where various metrics are reported: mean Intersection over Union (mIoU), Boundary IoU, Expected Calibration Error (ECE), and inference latency. Compared to the other methods, HCC, with a slight latency increase, achieves the best Pareto-front trade-off of balance between calibration error and Boundary IoU that sets a gold standard for the methods considered. The results of the ablations confirm that the Dynamic Neighborhood Consensus term is the most significant for improving the quality of boundary, while the Progressive Feature Interaction term is the most significant for calibration and semantic consistency, and that the two are complementary and not overlapping.
Keywords: Hierarchical Calibration, Multi-Objective Segmentation, Neighborhood Consensus, Progressive Feature Interaction, and Uncertainty-Aware Prediction
Volume: 15 | Issue: 2
Pages: 7-11
Issue Date: December , 2025
DOI: 10.9756/BIJAIP/V15I2/BIJ25016
 |
|