Saleem Muhammad Hammad, Velayudhan Kesini Krishnan, Potgieter Johan, Arif Khalid Mahmood
Department of Mechanical and Electrical Engineering, School of Food and Advanced Technology, Massey University, Auckland, New Zealand.
Massey AgriFood Digital Lab, Massey University, Palmerston North, New Zealand.
Front Plant Sci. 2022 Apr 25;13:850666. doi: 10.3389/fpls.2022.850666. eCollection 2022.
The accurate identification of weeds is an essential step for a site-specific weed management system. In recent years, deep learning (DL) has got rapid advancements to perform complex agricultural tasks. The previous studies emphasized the evaluation of advanced training techniques or modifying the well-known DL models to improve the overall accuracy. In contrast, this research attempted to improve the mean average precision (mAP) for the detection and classification of eight classes of weeds by proposing a novel DL-based methodology. First, a comprehensive analysis of single-stage and two-stage neural networks including Single-shot MultiBox Detector (SSD), You look only Once (YOLO-v4), EfficientDet, CenterNet, RetinaNet, Faster Region-based Convolutional Neural Network (RCNN), and Region-based Fully Convolutional Network (RFCN), has been performed. Next, the effects of image resizing techniques along with four image interpolation methods have been studied. It led to the final stage of the research through optimization of the weights of the best-acquired model by initialization techniques, batch normalization, and DL optimization algorithms. The effectiveness of the proposed work is proven due to a high mAP of 93.44% and validated by the stratified k-fold cross-validation technique. It was 5.8% improved as compared to the results obtained by the default settings of the best-suited DL architecture (Faster RCNN ResNet-101). The presented pipeline would be a baseline study for the research community to explore several tasks such as real-time detection and reducing the computation/training time. All the relevant data including the annotated dataset, configuration files, and inference graph of the final model are provided with this article. Furthermore, the selection of the DeepWeeds dataset shows the robustness/practicality of the study because it contains images collected in a real/complex agricultural environment. Therefore, this research would be a considerable step toward an efficient and automatic weed control system.
杂草的准确识别是特定地点杂草管理系统的关键步骤。近年来，深度学习（DL）在执行复杂农业任务方面取得了快速进展。以往的研究侧重于评估先进的训练技术或修改知名的DL模型以提高整体准确率。相比之下，本研究试图通过提出一种基于DL的新方法来提高八类杂草检测和分类的平均精度均值（mAP）。首先，对单阶段和两阶段神经网络进行了全面分析，包括单阶段多框检测器（SSD）、你只看一次（YOLO-v4）、高效检测器（EfficientDet）、中心网络（CenterNet）、视网膜网络（RetinaNet）、基于区域的更快卷积神经网络（Faster RCNN）和基于区域的全卷积网络（RFCN）。接下来，研究了图像缩放技术以及四种图像插值方法的效果。通过初始化技术、批量归一化和DL优化算法对最佳获得模型的权重进行优化，从而进入研究的最后阶段。所提工作的有效性通过93.44%的高mAP得到证明，并通过分层k折交叉验证技术进行了验证。与最适合的DL架构（Faster RCNN ResNet-101）默认设置所获得的结果相比，提高了5.8%。所呈现的流程将为研究界探索诸如实时检测和减少计算/训练时间等多项任务提供基线研究。本文提供了所有相关数据，包括带注释的数据集、配置文件和最终模型的推理图。此外，DeepWeeds数据集的选择显示了该研究的稳健性/实用性，因为它包含在真实/复杂农业环境中收集的图像。因此，本研究将朝着高效自动杂草控制系统迈出重要一步。
