Potato-ShuffleNetV2: A Field Potato Disease Identification Model Based on the Improved ShuffleNetV2

Abstract

Accurate and rapid identification of potato diseases is essential for preserving yield and quality. However, existing deep learning models suffer from low accuracy and redundancy issues. Moreover, models trained on datasets with idealized backgrounds often fail to meet the requirements for potato disease identification under real-field conditions. To address the aforementioned challenges, this study proposes a potato disease identification model named Potato-ShuffleNetV2, based on an improved ShuffleNetV2 architecture. First, a grouped efficient channel attention (GECA) mechanism is introduced to suppress interference from natural field backgrounds. Subsequently, a feature fusion module is constructed to reduce the loss of critical information. The ReLU activation function is then replaced with Gaussian error linear unit (GELU) to enhance model robustness. Finally, model pruning is applied to remove redundant feature layers and reduce model complexity. The final model achieves an accuracy of 97.78% on the potato disease test dataset, with only 0.13 M parameters and 0.03G FLOPs. Compared with state-of-the-art models such as ResNet50, DenseNet121, and EfficientNetV2_m, the proposed model improves classification accuracy by 1.53%, 2.29%, and 5.41%, respectively, while reducing the number of parameters by 99.45%, 98.13%, and 99.75%, respectively. Experimental results demonstrate that the proposed model achieves significant advantages in both recognition accuracy and model compactness, making it highly suitable for deployment in real-world field environments.

Plant phenotyping relevance

圃場画像からジャガイモ病害を識別する深層学習モデルを開発し、精度・モデル規模・実環境での性能を評価しており、植物病害状態の表現型取得が中心である。

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.