Karthikeyan S, Charan R, Narayanan Sathiya, Jani Anbarasi L
School of Electronics Engineering, Vellore Institute of Technology, Chennai, India.
School of Computer Science Engineering, Vellore Institute of Technology, Chennai, India.
Front Artif Intell. 2025 Aug 12;8:1640549. doi: 10.3389/frai.2025.1640549. eCollection 2025.
Technology is becoming essential in agriculture, especially with the growth of smart devices and edge computing. These tools help boost productivity by automating tasks and allowing real-time analysis on devices with limited memory and resources. However, many current models struggle with accuracy, size, and speed particularly when handling multi-label classification problems.
This paper proposes a Convolutional Neural Network with Squeeze and Excitation Enabled Identity Blocks (CNN-SEEIB), a hybrid CNN-based deep learning architecture for multi-label classification of plant diseases. CNN-SEEIB incorporates an attention mechanism in its identity blocks to leverage the visual attention that enhances the classification performance and computational efficiency. PlantVillage dataset containing 38 classes of diseased crop leaves alongside healthy leaves, totaling 54,305 images, is utilized for experimentation.
CNN-SEEIB achieved a classification accuracy of 99.79%, precision of 0.9970, recall of 0.9972, and an F1 score of 0.9971. In addition, the model attained an inference time of 64 milliseconds per image, making it suitable for real-time deployment. The performance of CNNSEEIB is benchmarked against the state-of-the-art deep learning architectures, and resource utilization metrics such as CPU/GPU usage and power consumption are also reported, highlighting the model's efficiency.
The proposed architecture is also validated on a potato leaf disease dataset of 4,062 images from Central Punjab, Pakistan, achieving a 97.77% accuracy in classifying Healthy, Early Blight, and Late Blight classes.
技术在农业中变得至关重要，尤其是随着智能设备和边缘计算的发展。这些工具通过自动化任务并允许在内存和资源有限的设备上进行实时分析来提高生产力。然而，许多当前模型在准确性、大小和速度方面存在困难，特别是在处理多标签分类问题时。
本文提出了一种具有挤压与激励恒等块的卷积神经网络（CNN-SEEIB），这是一种基于卷积神经网络的混合深度学习架构，用于植物病害的多标签分类。CNN-SEEIB在其恒等块中融入了注意力机制，以利用视觉注意力来提高分类性能和计算效率。使用包含38类患病作物叶片以及健康叶片的PlantVillage数据集，共计54305张图像进行实验。
CNN-SEEIB的分类准确率达到99.79%，精确率为0.9970，召回率为0.9972，F1分数为0.9971。此外，该模型每张图像的推理时间为64毫秒，适用于实时部署。将CNNSEEIB的性能与最先进的深度学习架构进行基准测试，并报告了诸如CPU/ GPU使用情况和功耗等资源利用指标，突出了该模型的效率。
所提出的架构还在来自巴基斯坦旁遮普省中部的4062张马铃薯叶病害数据集上进行了验证，在对健康、早疫病和晚疫病类别进行分类时准确率达到97.77%。
