外文摘要：Plant health conditions are becoming more prevalent in smart farming, and diagnosing plant diseases usually involves a visual inspection and a physical examination by experts. Detecting the signs of a disease at an early stage may prevent the spreading of pathogens in the farmland. However, these examinations are computationally costly and time-consuming. Complex machine learning algorithms such as convolutional neural networks cannot run on the resource constraint IoT devices since they require high-configuration hardware for training and testing. Therefore, this paper proposes an energy-efficient and computationally lightweight solution deployed on an IoT device, i.e., esp32-cam, that can recognize nine (9) different plant diseases using a quantized CNN (Q-CNN) approach. Q-CNN speeds up computation and reduces the model size while maintaining optimal performance. The memory overhead of the convolutional layers and fully connected layers are compressed via model quantization. The size of the developed model is only 28 KB with exclusive int8 quantization and is utterly attainable on low computational edge devices. Experimental outcomes demonstrate that the proposed approach can perform better than existing methods, gaining an overall F-1 accuracy of 98% on resource-constrained IoT devices.
外文关键词：Quantized CNN;Plant disease recognition;IoT edge device;TinyML
作者：Rakib, Abdul Fattah;Rahman, Rashik;Al Razi, Alim;Hasan, A S M Touhidul
作者单位：Univ Missouri;Univ Asia Pacific
期刊名称：ARABIAN JOURNAL FOR SCIENCE AND ENGINEERING
期刊影响因子：0.0
出版年份：2024
出版刊次：49（3）
原文传递申请：江苏省科技资源（工程技术文献）统筹服务平台