Weed and crop discrimination through an offline computer vision algorithm

With the recent global interest in organic farming and cultivation, many people are turning away from chemical-based herbicides and moving towards alternate methods to extirpate weeds living amongst their crops. Of the methods proposed, robotic weed detection and removal is the most promising because of its possibility to be completely autonomous. Several robust, fully-autonomous robots have been developed, although none have been approved for commercial use. This paper proposes a weed and crop discrimination algorithm that utilizes an excessive green filter paired with principal component analysis to detect specific spatial frequencies within an image corresponding to different types of weeds and crops. This method also works to reduce dimensions in data by representing an image as a small set of values obtained from a projection. This technique optimizes performance while allowing for simpler calculations. These calculations were used to develop thresholds for weeds, crops, and soil for discrimination purposes. The algorithm resulted in an overall classification rate of 77%. 46% of all crops were identified correctly; 78% of all weeds were identified correctly; and 91% of all soil was identified correctly. The low rate of correct crop classification was due to poor edge detection by the algorithm, but could be improved in future research by applying one or more edge-detection algorithms. This technique can be adapted in the future with other image-analysis techniques to be used on low-cost systems