PR-CNN: A Multiscale Attention Relation Network for Accurate Bean Leaf Disease Image Recognition

Abstract

Abstract Accurate recognition of plant leaf diseases from images is essential for intelligent agriculture and precision crop protection. However, reliable disease identification remains challenging because lesion regions often exhibit subtle visual differences, complex backgrounds, and large intraclass variations, especially when available disease samples are limited. To address these challenges, this study proposes PR-CNN, a deep learning framework that integrates convolutional neural networks, pyramid split attention, and a relation network for bean leaf disease image recognition. The convolutional backbone is first used to extract visual features from support and query images. Then, the pyramid split attention module enhances multiscale spatial and channel feature representation, enabling the model to focus on discriminative lesion regions while suppressing redundant background information. Finally, the relation network learns a nonlinear similarity metric between paired samples and generates relation scores for disease category prediction. Experimental results show that PR-CNN achieves an overall classification accuracy of 99.24% on the primary bean leaf disease dataset, outperforming representative models, including ResNet50, DenseNet, Inception v4, and EfficientNet B7, in terms of recognition accuracy and adaptability. In addition, PR-CNN was evaluated on four publicly available plant disease datasets, including CGIAR, Plant Diseases, LWDCD 2020, and Plant Pathology, achieving an average accuracy of 99.84%. These results demonstrate that PR-CNN can effectively improve image based plant disease recognition and provides a robust visual classification framework for intelligent crop disease diagnosis.

Plant phenotyping relevance

豆葉画像から病徴・病害を認識する深層学習手法を開発し、複数データセットで性能検証しており、植物の病害状態の画像ベース計測が研究の中心である。

Code and data availability

The paper's self-collected bean leaf disease image dataset (11,903 images) is not publicly deposited; the data availability statement only offers access upon request. No author code, models, or public repository URLs are provided. Publicly used datasets (CGIAR, Plant Pathology, etc.) are generic external resources, not

No evidence-backed public reproduction asset is currently recorded.