Detection of diseases and pests on images captured in uncontrolled conditions from tea plantations

摘要

Timely and accurate recognition of health conditions in crops helps to perform necessary treatment for theplants. Automatically localizing these conditions in an image helps in estimating their spread and severity, thussaving on precious resources. Automated disease detection involving recognition as well as localization helpsin identifying multiple diseases from one image and can be a small step forward for robotic farm surveyingand spraying. Recent developments in Deep Neural Networks have drastically improved the localization andidentification accuracy of objects. We leverage the neural network based method to perform accurate and fastdetection of the diseases and pests in tea leaves. With a goal to identify an accurate yet efficient detector interms of speed and memory, we evaluate various feature extraction networks and detection architectures. Theimages used to train and evaluate the models are with different resolutions, quality, brightness and focus asthey are captured with mobile phones having different cameras through a participatory sensing approach. Theexperimental results show that the detection system effectively identifies and locates the health condition on thetea leaves in a complex background and with occlusion. We have evaluated YOLO based detection methods withdifferent feature extraction architectures. Detection using YOLOv_3 achieves mAP of about 86% with 50% IOUwhile making the system usable in real time.