AUTHOR=Zhang Enxu , Zhang Ning , Li Fei , Lv Cheng 

TITLE=A lightweight dual-attention network for tomato leaf disease identification

JOURNAL=Frontiers in Plant Science

VOLUME=Volume 15 - 2024

YEAR=2024

URL=https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2024.1420584

DOI=10.3389/fpls.2024.1420584

ISSN=1664-462X

ABSTRACT=Tomato disease image recognition plays a crucial role in agricultural production. However, utilizing machine vision for disease identification faces challenges such as unclear image features, uneven spatial distribution of disease characteristics, small inter-class differences, and significant intra-class variations. To address these issues, this paper proposes a method for tomato leaf disease classification and recognition. Firstly, image data is enhanced using Segmentation Linear Transformation combined with adjustable thresholds and the AutoAugmen method to highlight detailed image features. Additionally, oversampling is employed to handle imbalanced datasets. Subsequently, a lightweight network model, LDAMNet, is constructed using the Dual Attention Convolutional Block (DAC). Within the DAC module, Hybrid Channel Attention (HCA) and Coordinate Spatial Attention (CSA) are applied to enhance the model's capability to extract and retain image features by processing channel and spatial information of the images. Finally, a Robust Cross-Entropy (RCE) loss function robust to noisy labels is proposed to optimize the training of the LDAMNet model. Experimental results demonstrate that LDAMNet trained with the RCE loss function achieves an average recognition accuracy of 98.71% on the tomato disease dataset, outperforming large-scale models such as ConvNeXt V2, DenseNet, as well as lightweight models like MobileNet. Moreover, LDAMNet maintains significant advantages when tested on other crop disease datasets.