An Interpretable Image Classification Approach Using Prototype-Based Deep Belief Rules

Belief rule-based systems, despite their interpretability in classification tasks, face two critical limitations in image analysis: the curse of dimensionality when handling highdimension data and incompatibility with non-tabular formats, which restrict their applicability to interpretable image classification applications. To address these challenges, we propose a prototype-based deep belief rule reasoning methodology for interpretable image classification. The core methodology involves replacing decision layer in deep neural networks with a deep belief rule base, enabling joint optimization of classification performance and decision transparency. The deep belief rule base construction comprises two critical phases: anchoring initial class boundaries through δ-B e l decision graph, and generating comprehensive deep belief rule base, in which the basic principle for constructing the deep belief rule base is to rely on the distance between prototypes, instead of distinguishing the actual categories of prototypes. We conduct extensive experiments on multiple benchmark datasets, demonstrating that our approach achieves superior interpretability while maintaining high classification performance.