Optické rozpoznávání rostlin pomocí neuronové sítě

Abstract

The increasing cost of manual work together with the increasing pressure from governments on minimizing of the chemical usage in agriculture intensifies the need for automation in the weed removal process. This work aims to present an algorithm capable of detection of an agricultural crop (sugar beet plant) in a so-called dangerous zone. The process of the automatic mechanical removal in the dangerous zone can only continue if the crop is not detected in this area. The detection algorithm is based on a neural network of the MobileNetV2 architecture. After the training on an acquired dataset of 73,600 images, the accuracy of detection using this algorithm was 95 %. To increase the accuracy of the detection even more significantly, it would be necessary to have more training images. The algorithm was deployed on a single-board computer NVIDIA Jetson Nano with the image processing pipeline speed of 40 frames per second.