Deep Learning Approach for high Energy efficient Real-Time Detection of Weeds in Organic Farming

Organic farming, vision based detection and classification systems can be used to reduce chemical and synthetic pesticides. Automated weeding control requires in addition to classification also localization through bounding boxes. By using tensor processing units, the typical implementation of convolutional neural networks by means of graphics cards should be improved. The main objective is to improve energy consumption without deteriorating accuracy and frame rate. In this approach a Google Coral USB Accelerator with an Edge-TPU and a Raspberry PI 4 Model B is used. The MobileNetV2-SSD was chosen for this application because of its ability to run on embedded systems. The use of tensor processors can enable time-saving calculations. An accuracy of up to 62.3% is achieved. A maximum power of 5.7 W is obtained. Compared to a tiny-YOLO network accelerated on a Jetson TX2, the speed of the CNN is outperformed and the energy efficiency is increased. In addition, the acquisition costs are lower. Nevertheless, the accuracy of a tiny-YOLO network accelerated on a Jetson TX2 was not achieved.