Weeds and Corn Classification by Image Processing and Neural Network Techniques

Abstract

Expensive and laborious job of weed control can be facilitated if automatic weeding machines are employed. Site-specific
managements of the weeds in the field need accurate discrimination between the crop and the weeds. There are distinct species of
the weeds so called “common weeds” for cultivation in a specific region. Three species of the weeds commonly grow in corn fields
are considered in this study, which are Convolvulus arvensis, Chenopodium album, and Amaranthus retroflexus. There are distinct
differences between the shapes of the plants especially in early growing stages. Therefore, ten shape features of the leaves were
considered for discrimination between the weeds and corn plants. An image processing algorithm was developed and combined
with the artificial neural network (ANN) for classification of corn and weeds. Several images of the leaves of each plant were
taken. The ten shape features extracted from the images by image processing algorithm were fed as the input to the ANN classifier.
A number of the corn and weeds leaves’ images were used to train the network. Several topologies of ANN including single and
multi layer perceptrons (MLPs) with various transfer functions such as MLP-GDM, MLP-RP and MLP-SCG were used. Finally,
the ability of the ANN models for classifying weeds and corn plants were evaluated using new image data. Results revealed that
the ANN could discriminate corn from weeds with an accuracy of 98.5%. However, the algorithm had less accuracy for classifying
the weeds from each other which was limited to 78.5%.