Automatic visual Weed Recognition - Detection and Classification of Weed in Row Cultures combining Machine Vision and Artificial Intelligence
Abstract
This thesis was motivated by the use of machine vision and artificial intelligence in agricultural robotics. More efficient agricultural production is needed as the worlds population grows. It is also important, both from an economical and environmental point of view, to reduce the amount of applied herbicidal products in agriculture. The use of machine vision and artificial intelligence are already a part of the modern agriculture and will play a important role in the technology to come.
The aim of this thesis is to implement a program for leaf detection in row crops and investigate the use of different classifiers to detect weed. This work will form the base of the weed detection part in a bigger project, Asterix, owned by Adigo AS.
The program uses images, in this thesis from a carrot field, to detect leaves by segmentation and connected components analysis. From each leaf, ten features are extracted to be used in the classification process. By a graphical user interface, a user can label leaves into given classes to create training sets.
By the use of a resulting training set, leaf-objects were classified by six different basic classifiers. By this classification, a success rate of over 89% total correct classified leaves was reached. Furthermore, by raising the lower boundary of the leaf size and using votes from all the classifiers to determine the class, total classification success rate surpassed 95%.