Lightweight Deep Learning Model for Weed Detection for IoT Devices

Increased crop productivity is largely dependent on weed control, thus making weed detection an important aspect of smart agriculture systems. To spray pesticides, it is necessary to distinguish between the weed and the crop. Owing to the recent advancements in Computer Vision algorithms and techniques, researchers have proposed numerous deep learning based pipelines for weed detection. However, deep learning is computationally expensive due to its dependency on power-ful GPUs (Graphics Processing Unit). This study attempts to tackle the problem of maintaining the trade-off between cost-effectiveness and performance of algorithms by utilizing You Only Look Once version 4 - tiny(YOLOv4-tiny). YOLOv4-tiny is a high-performance fast deep learning model that can run on machines with less computing capacity, providing farmers with a cost-effective IoT weed identification solution. Beginning with a literature survey on data collection, preparation, and model architectures, we propose the usage of YOLOv4-tiny for weed detection which is an object detection model and finally explain the deployment of our trained lightweight model on Raspberry Pi 4 Model B.