A system for weeds and crops identification based on convolutional neural network

Abstract

This paper presents an early step towards an autonomous weeding system. The system is based on the Deep Convolutional Neural Network (Deep ConvNet, CNN). CNNs reached state-of-the-art results in many computer vision tasks. However, their effectiveness is strongly related to the network architecture, as well as quality and quantity of the training data, and the data collection is a time-consuming process. In this paper, we will present how to find the first approximation of the network architecture and the data quantity, based on two sequences of 100 crop images. The obtained accuracy level equals to 96–98%. The presented approach will be used to train and test the CNN on larger datasets in the future work.