A multiresolution based method for recognizing weeds in corn fields

In order to reduce the quantities of herbicides applied to cereals, we have to detect and localize by image processing the places where weeds are located. The herbicide spray is only performed if the weed density is sufficient, otherwise spraying is not performed because the weeds will lose the competition with the cereal. A color camera provides 4 images red (R), green (G), blue (B) and infrared (IR) bands of agronomic scenes. The resolution is 2036/spl times/3060 pixels. The color images (R,G,B) are filtered by L-filters. We have studied a lot of color representations and find that the most efficient one was (L, a, b). For separating vegetation and soil, the components a and IR are fused after thresholding from histograms. The recognition of weeds is based on the skeletons of leaves. A method which provides 4-connected skeletons gives good results. Finally, a small expert system based on color and geometrical features enables us to recognize the classes of the different parts of the skeletons of connected components. A conditional dilation using these labeled parts of skeletons and the binary image vegetation/soil is carried out. On a few connected components, we obtain a mixture of different labels. In order to solve these ambiguities, we compute geometrical features and topological relationships of leaves and the decision is taken with the technique of Bayesian nets. On images with small quantities of overlapping, vegetal species are well recognized. At the and, we compute the covers and locations of maize and weeds in order to get the densities of vegetal species. When the corn and weeds are not at an earlier stage, the vegetal species recognition process is much more complex, a multiresolution technique based on wavelets is performed in order to classify weeds and corn with a better accuracy.