Wavelet transform to discriminate between crop and weed in perspective agronomic images

Abstract : We proposed testing and validating the accuracy of four image processing algorithms (wavelet transforms and Gabor filtering) for crop/weed discrimination in synthetic and real images. A large panel of wavelet bases (33) was tested and the two best wavelets and the worst one were selected for detailed study. Based on a confusion matrix the crop/weed classification results of wavelet transforms were compared to the results of Gabor filtering that was initially chosen to develop a machine vision system for a real-time precision sprayer. The accuracy of these algorithms was compared and showed that wavelets were well adapted for perspective images: the best results were with Daubechies 25 and discrete approximation Meyer wavelets. They provided better results than Gabor filtering not only for crop/weed classification but also in processing time.