# Fine-grained few-shot class-incremental identification of medicinal plants via frequency-aware contrastive learning

## TL;DR

This paper introduces a new deep learning framework for identifying medicinal plants with limited labeled data, improving accuracy and reducing forgetting.

## Contribution

The FGDE framework uses frequency-aware contrastive learning to enable few-shot class-incremental learning for plant identification.

## Key findings

- FGDE outperforms existing methods on plant identification tasks with limited labeled examples.
- The framework effectively reduces catastrophic forgetting and overfitting during incremental learning.
- Multi-frequency fusion and contrastive learning enhance feature representation and discriminative power.

## Abstract

Developing robust algorithmic tools for accurately identifying diverse medicinal plant species is critical for advancing precision medicine. Although deep learning methods have shown considerable promise, they generally require large-scale annotated datasets, which are often difficult to acquire given the vast taxonomic diversity and limited labeled samples available for many plant species. To address this, we propose a novel Frequency-Aware Guided Domain Enhancement Contrastive Learning (FGDE) framework, designed to incrementally learn new categories from few annotated examples while alleviating catastrophic forgetting and overfitting. Our approach integrates high- and low-frequency components to refine feature representations, using multi-frequency fusion to preserve detail-enhanced information. Contrastive learning is further employed to strengthen multi-semantic aggregation and extract discriminative features across both visual and label domains. Additionally, we introduce a multi-objective loss function to enhance semantic compactness within base classes and improve separation among incremental classes. Extensive experiments demonstrate that FGDE significantly outperforms state-of-the-art methods on our collected dataset and two public benchmarks. These results underscore the potential of our model to support practical applications in intelligent plant identification and precision agriculture.