Lv Jiawei, Li Wenyong, Fan Mingyuan, Zheng Tengfei, Yang Zhankui, Chen Yaocong, He Guohuang, Yang Xinting, Liu Shuangyin, Sun Chuanheng
National Engineering Research Center for Information Technology in Agriculture, Beijing, China.
National Engineering Laboratory for Agri-product Quality Traceability, Beijing, China.
Front Plant Sci. 2022 Jul 7;13:939498. doi: 10.3389/fpls.2022.939498. eCollection 2022.
Light traps have been widely used as effective tools to monitor multiple agricultural and forest insect pests simultaneously. However, the current detection methods of pests from light trapping images have several limitations, such as exhibiting extremely imbalanced class distribution, occlusion among multiple pest targets, and inter-species similarity. To address the problems, this study proposes an improved YOLOv3 model in combination with image enhancement to better detect crop pests in real agricultural environments. First, a dataset containing nine common maize pests is constructed after an image augmentation based on image cropping. Then, a linear transformation method is proposed to optimize the anchors generated by the k-means clustering algorithm, which can improve the matching accuracy between anchors and ground truths. In addition, two residual units are added to the second residual block of the original YOLOv3 network to obtain more information about the location of the underlying small targets, and one ResNet unit is used in the feature pyramid network structure to replace two DBL(Conv+BN+LeakyReLU) structures to enhance the reuse of pest features. Experiment results show that the mAP and mRecall of our proposed method are improved by 6.3% and 4.61%, respectively, compared with the original YOLOv3. The proposed method outperforms other state-of-the-art methods (SSD, Faster-rcnn, and YOLOv4), indicating that the proposed method achieves the best detection performance, which can provide an effective model for the realization of intelligent monitoring of maize pests.
诱虫灯已被广泛用作同时监测多种农业和森林害虫的有效工具。然而，目前从诱虫图像中检测害虫的方法存在一些局限性，例如类分布极度不平衡、多个害虫目标之间存在遮挡以及物种间相似性。为了解决这些问题，本研究提出了一种改进的YOLOv3模型，并结合图像增强技术，以便在真实农业环境中更好地检测作物害虫。首先，在基于图像裁剪的图像增强后，构建了一个包含九种常见玉米害虫的数据集。然后，提出了一种线性变换方法来优化由k均值聚类算法生成的锚框，这可以提高锚框与真实值之间的匹配精度。此外，在原始YOLOv3网络的第二个残差块中添加了两个残差单元，以获取更多关于潜在小目标位置的信息，并在特征金字塔网络结构中使用一个ResNet单元来替换两个DBL（卷积+批归一化+泄漏整流线性单元）结构，以增强害虫特征的重用。实验结果表明，与原始YOLOv3相比，我们提出的方法的平均精度均值（mAP）和平均召回率（mRecall）分别提高了 6.3%和 4.61%。所提出的方法优于其他现有最先进的方法（SSD、Faster-rcnn和YOLOv4），表明所提出的方法实现了最佳检测性能，可为实现玉米害虫智能监测提供有效模型。
