Image-blur-based Robust Weed Recognition

Abstract

Image motion blur and defocus blur often occur when there is a relative motion between the imaging camera and the detected object. These two blurs will degrade the image quality and will also decrease the subsequent pattern recognition accuracy. In this paper, we propose a robust weed recognition scheme using the low quality color weed images with the above-mentioned image blurs. The proposed scheme consists of three steps. First, image matte is used to segment the soil and the plant. Second, a generative learning method is introduced in the training step to simulate blurred images by controlling blur parameters. Finally, weed recognition is performed by using the blurred color information based on the subspace method. We have experimentally proved that the effective use of image blurs improves the recognition accuracy of camera-captured weeds.