Adaptive recognition network for few-shot plant diseases and pests based on homeostatic neuromodulation and meta-plasticity

Abstract

Introduction Static networks often exhibit limited generalization on few-shot data, particularly given the scarce samples and unstructured background noise inherent to precision agriculture. To address these limitations, an adaptive recognition network for few-shot plant diseases and pests based on homeostatic neuromodulation and meta-plasticity (HNeuroNet) is proposed. Methods This framework incorporates dynamic plasticity inspired by biological systems to mitigate the data dependency paradox. First, a Neuro Modulatory Generator (NMG) is constructed utilizing a hypernetwork architecture. Simulating neurotransmitter gating mechanisms, affine transformation parameters are dynamically generated for feature channels based on support set samples. Consequently, instantaneous weight reconstruction is enabled without expensive gradient fine-tuning, thereby overcoming structural rigidity and catastrophic forgetting during rapid adaptation. Second, a Homeostatic Suppression Mechanism (HSM) integrating visual perception is introduced. Leveraging Bienenstock-Cooper-Munro (BCM) theory, an adaptive activation function is employed to regulate neuron thresholds based on historical feature map statistics. High-frequency noise from complex environments is suppressed, significantly enhancing feature extraction and target saliency in low signal-to-noise ratios. Finally, an end-to-end Dynamic Meta-Plasticity (DMP) strategy is implemented. By coupling parameter generation and threshold regulation within a bi-level optimization framework, biological homeostatic adaptation is simulated to adjust perception strategies. Context-dependent feature interaction patterns are established to secure robust discriminative boundaries under extreme few-shot conditions. Results Experimental results demonstrate that HNeuroNet significantly outperforms state-of-the-art methods on IP102, PlantDoc, and Mini-ImageNet. Notably, 5-way 1-shot accuracy on the PlantDoc dataset surpasses the second-best baseline by 4.33%. Furthermore, a 1-shot accuracy of 71.36% is achieved on the cross-domain Mini-ImageNet task. Discussion These results confirm the potential of bio-inspired computing in addressing data scarcity.

Plant phenotyping relevance

植物画像から病害・害虫状態を認識する適応型深層学習手法を開発し、PlantDoc等で性能検証しており、病害状態の推定と手法自体が研究の中心である。

Code and data availability

The supplied blocks contain no data availability statement, code repository, or author-deposited assets. The paper evaluates on public benchmarks (IP102, PlantDoc, Mini-ImageNet), but these are cited prior-work datasets, not paper-specific assets, and no authors' public URL for code or models is present.

No evidence-backed public reproduction asset is currently recorded.