Directional Shape Feature Extraction Using Modified Line Filter Technique for Weed Classification

Abstract

Precision agriculture is gaining attention as it employs modern technologies and intelligence for automation in agricultural practices. In the area of weed management, automation is advantageous to select the appropriate herbicide and manage the amount used, which consequently reduced the cost and minimizes the environmental impact. Selective spraying using a sprayer boom can be implemented using automatic detection of weed type. This paper presents a weed classification method based on a modified line filter image analysis technique that can effectively detect the morphological differences, mainly directional shape features, between two types of weeds. After the result for binary classification has been verified, a third dataset is introduced which is mixed leaves which consists of an approximately balanced amount of broadleaves and narrow leaves. The weed images were pre-processed using the adaptive histogram method and difference of Gaussian to improve the image contrast and delineate the edges of the weed. The images were then processed using the proposed modified line filter feature extraction technique. The filter is based on the evaluation of pixel response that corresponds to the pre-defined lines at different orientations from 0o to 360o. The pixel strength of each line was compared to determine the overall response of the filter. The proposed method achieved around 97% classification, superior compared to previously reported methods such as Gabor wavelet as well as a combination of Gabor and Gradient Field Distribution (GFD).