ACPNet: adaptive concept-prototype network for interpretable fine-grained image classification

Shiv Singh

doi:10.1007/s13735-026-00419-6

ACPNet: adaptive concept-prototype network for interpretable fine-grained image classification

2026-09-16

0

PRE

AI

DOI：10.1007/s13735-026-00419-6

Original

Save

Abstract

En 中文

Fine-grained image classification, like bird species classification in the Caltech-UCSD Birds-200-2011 (CUB-200-2011) dataset, requires high accuracy and explainability for trustworthiness in high-stakes applications. However, these are hard to achieve with the complexity of deep models and fine-grained visual contrasts in fine-grained tasks. We introduce the $${\textbf {Adaptive Concept-Prototype Network (ACPNet)}}$$ , a new model that combines spatial attention, concept bottlenecks, and prototype learning to make accurate and human-explainable predictions. ACPNet utilizes a ResNet-50 backbone, a 1x1 convolutional attention mechanism to attend to discriminative regions, a 10-concept bottleneck for semantic representation, and 10 learnable prototypes for class-specific matching, and is trained using multi-task loss. Tested on CUB-200-2011, ACPNet shows competitive Top-1 accuracy of $${\textbf {95\%}}$$ and Top-5 accuracy of $${\textbf {97\%}}$$ , comparable with state-of-the-art interpretable models. Interpretability metrics show high clarity (0.60), robustness (0.99), and concept sparsity (0.35), supported by visualizations of the top 2 attention regions (e.g., beak, wings, etc.) and top concept activation (e.g., “beak”, etc). Contributions include a hybrid architecture for spatially adaptive, semantically grounded explanations, memory-efficient implementation with mixed precision and gradient accumulation, and extensive evaluation on accuracy, interpretability metrics, and visualizations. ACPNet is an improvement of the state of the art for interpretable AI on fine-grained tasks. It offers an extensible solution for tasks such as biodiversity monitoring and medical imaging, where transparency is paramount.

Keywords:

Explainable AI

Concept bottleneck

Prototype learning

Interpretability

Fine-grained image classification

ACPNet: adaptive concept-prototype network for interpretable fine-grained image classification

Abstract

Journal

Organization

Cited Papers