Wavelet transform to discriminate between crop and weed in agronomic images

Abstract

In precision agriculture, the reduction of herbicide applications requires an accurate detection of weed patches. From image detection, to quantify weed infestations, it would be necessary to identify crop rows from line detection algorithm and to discriminate weed from crop. Our laboratory developed several methods for line detection based on Hough Transform, double Hough Transform or Gabor filtering. The Hough Transform is well adapted to image affected by perspective deformations but the computation burden is heavy and on-line applications are hardly handled. To lighten this problem, we have used a Gabor filter to enhance the crop rows present into the image but, if this method is robust with parallel crop rows (without perspective distortions), it implies to deform image with an inverse projection matrix to be applied in the case of an embedded camera. We propose, in order to implement a filter in the scale / space domain, to use a discrete dyadic wavelet transform. Thus, we can extract the vertical details contained in various parts of the image from different levels of resolution. Each vertical detail level kept allows to enhance the crop rows in a specific part of the initial image. The combination of these details enable us to discriminate crop from weeds with a simple logical operation. This spatial method, thanks to the fast wavelet transform algorithm, can be easily implemented for a real time application and it leads to better results than those obtained from Gabor filtering. For this method, the weed infestation rate is estimated and the performance are compared to those given by other methods. A discussion concludes about the ability of this method to detect the crop rows in agronomic images. Finally we consider the ability of this spatial-only approach to classify weeds from crop.