CIGR VI 2019

講演情報

Oral Session

Others (including the category of JSAM and SASJ)

[4-1600-D] Other Categories (1)

2019年9月4日(水) 16:00 〜 18:15 Room D (4th room)

Chair:Satoshi Yamamoto(Akita Prefectural University), Kikuhito Kawasue(University of Miyazaki)

17:30 〜 17:45

[4-1600-D-07] A Deep Learning and MSM Machine Learning System for Recognition of Weed Infestation in Cabbage Field Using Unmanned Aerial Vehicle

*Tofael Ahamed1, Yan Zhang1, Linhuan Zhang1, Ryozo Noguchi1 (1. Faculty of Life and Environmental Sciences, University of Tsukuba(Japan))

キーワード:Deep Learning , Convolution Neural Network (CNN), Mutual Subspace Method (MSM) , Precision Agriculture, Spot Spraying

Cabbage is susceptible to grow due to weeds incidence and requires large amounts of herbicides in the small-scale Japanese farms. In precision application of herbicides, it is required to recognize the classifiers to minimize herbicides application. Therefore, the purpose of this research is to deal with recognition of weed infestation in the cabbage field using two classifiers: cabbage and weeds. A DJI Phantom UAV was flown with an onboard 4K RGB camera from 2m heights to identify the weed incidence in a cabbage field located at the Ibaraki Prefecture of Japan. Two videos were used and converted to figures: one for training and other for testing. In the pre-process stage, each original image with size of 1920x1080 was divided into 250x250 small size sub-graphs using a sliding window, with step of 250. Each sub-image could be defined as: cabbage and weeds. In the training, 676 datasets for cabbage, 667 datasets for weeds were taken from sub-images. Alexnet CNN deep learning and Mutual Subspace Method (MSM) machine learning were used to find the recognition of the two classes.The accuracy for recognizing the classifiers using MSM was 61%. To improve the accuracy of MSM, Histogram Oriented Gradient (HOG) method was used with MSM. The recognition accuracy was increased to 88% using MSM-HOG algorithm. The overall accuracy was achieved 94%for recognizing the classifiers using AlexNet CNN.In the deep learning process, the enlarging dataset and learning technology can be inherit with AlexNet CNN and MSM-HOG. Further research will be carried out using weights in each of the layer to improve the accuracy of the classifiers. This deep-learning approach has the potential to add in the spot sprayer for real-time application to minimize herbicides for UAV.