부산대학교 도서관

Shapes of rice leaves look very similar with those of wild millets. This makes the discrimination of weeds from rices on images to be extremely difficult. As a result, estimation of their locations becomes a very hard task. By virtue of the recent development of computer technology, dramatic development of Deep Learning Neural Network (DNN) technology has been made. Thus, the DNN is considered as a promising solution for the agricultural automation which was impossible with the conventional technologies. In this thesis, a neural network-based method discriminating wild millets from rices on rice fields is proposed. Particularly, Encoder-Decoder Neural Network is investigated in deep. Based on the Encoder-Decoder network, the location, orientation, as well as the class of plants are identified and its results are represented on images using graphic forms. Three major Encoder-Decoder networks such as Basic Encoder–Decoder, Pooling–Unpooling Encoder–Decoder, Skip Encoder–Decode are investigated. The training of the networks has been performed with 4,828 cases of rice, 2,396 cases of wildmillet and respectively. Among these 3 networks, Skip Encoder–Decoder Network shows the best identification performance with 78.7%, 83.02% and 73.19% for, and , respectively. In the aspect of identification speed, the skip-network shows 277.77 fps which is much faster than ordinary live video speed of 30fps. Considering classification accuracy and speed, the Skip-Encoder and Decoder network would be the best option to build an intelligent vision system for automatic weeding machine on rice fields.