Part-Aware Prototype-Aligned Interpretable Image Classification with Basic Feature Domain

In recent years, the interpretive this looks like that structure has gained significant attention. It refers to the human tendency to break down images into key parts and make classification decisions by comparing them to pre-existing concepts in their minds. However, most existing prototypical-based models assign prototypes directly to each category without considering that key parts with the same meaning may appear in images from different categories. To address this issue, we propose dividing prototypes with the same meaning into the same latent space (referred to as Basic Feature Domain) since different category parts only slightly affect the corresponding prototype vectors. This process of integrating prototypes based on the feature domain is referred to as prototype alignment. Additionally, we introduce the concept of part-aware optimization, which prioritizes prototypical parts of images over simple category labels during optimizing prototypes. Moreover, we present two feature aggregation methods, by row and by cluster, for the basic feature domain. We demonstrate competitive results compared to other state-of-the-art prototypical part methods on the CUB-2011-200 dataset and Stanford Cars dataset using our proposed self-explanatory part-aware proto-aligned network (PaProtoPNet).