LRConfNet: Logical Reasoning-Driven Confidence Adjustment and Regularization for Hierarchical Classification of Degraded Images

YANG CHU, Keli Deng, Xiaomeng Yang, Yuntao Qian

Keywords: Uncertainty Quantification, Dynamic Confidence Adjustment, Hierarchical Classification, Logical Reasoning Regularization, Attention Enhancement

Abstract: Hierarchical classification (HC) is widely applied in remote sensing and natural image analysis. However, real-world degradations—such as noise, blur, occlusion, and low resolution—often compromise fine-grained predictions for HC. Existing methods struggle to balance coarse- and fine-level accuracy, handle sparse hierarchies, and integrate multi-modal features, particularly under low-confidence predictions and complex semantic structures. We propose LRConfNet, a unified framework that addresses these challenges by combining Uncertainty Quantification (UQ) with Logical Reasoning Regularization (LogReg) to dynamically adjust classification paths. A Vision Transformer (ViT) backbone extracts global visual features, while a Semantic-Guided Cross-Attention module enables multi-modal fusion. When fine-grained confidence is low, LRConfNet triggers a logic-driven hierarchical fallback mechanism, guided by LogReg, to back off to coarse-level predictions and avoid over-classification. To further enhance generalization, we introduce a multi-level loss optimization strategy with adaptive weight adjustment. An attention enhancement loss and attention-gradient fusion are incorporated to refine spatial focus, especially confronting degraded conditions and data scarcity. Moreover, a position prompting mechanism reinforces feature selection in sparse hierarchies. Extensive experiments on degraded remote sensing and natural image benchmarks show that LRConfNet significantly outperforms SOTA methods, demonstrating superior robustness and adaptability.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 10307