Paper Bag-of-visual-words-augmented Histogram of Oriented Gradients for efficient weed detection

As season-long weeds competition produces important yield losses, early detection of these plants is essential to sustain productivity. Machine vision as a non-destructive surveying technique requires features that can describe weeds in a real field case. Colours and shapes provide good results in controlled conditions. However, when different crops or weeds appear in clusters, such solutions fail to meet satisfactory performance. Therefore, considering features that are less specific to field conditions is crucial for integrated weed management. In this study, we provide effective use of the Histogram of Oriented Gradients (HOG) to improve its performance for weed detection. The concept is based on the Bag-of-Visual-Words (BOVW) approach. We use the HOG blocks as keypoints to generate the visual-words, and the features vectors are the histograms of these visual words. Next, we use the Backpropagation Neural Network to detect weeds and classify plants for three different crop fields. Namely, we consider sugar-beet, soybean, and carrot as target crops. Results demonstrate that the proposed weed detection system can locate weeds for site-specific treatment and selective spraying of herbicides. The proposed BOVW-based HOG can discriminate between weeds and crops with an accuracy of 97.7%, 93%, and 96.6% in sugar-beet, carrot and soybean fields respectively. For plant classification, our method can classify plants with an accuracy of 90.4%, 92.4%, and 94.1% in sugar beet, carrot and soybean fields respectively. Our results turn out 37.6% better than the classical HOG that produces an accuracy ranging from 71.2% to 83.3% in weed detection and 49.1%-82.1% in plant classification. (c) 2020 IAgrE. Published by Elsevier Ltd. All rights reserved.

Automated summary

Early detection of weeds is crucial for sustaining productivity, and the use of Histogram of Oriented Gradients (HOG) within a Bag-of-Visual-Words (BOVW) framework can significantly improve weed detection accuracy. The proposed method demonstrates satisfactory results in practical tests across three different crop fields.