Ding Biao, Tian Yunxiang, Guo Xiaojun, Wang Longshen, Tian Xiaolin
School of Electronic and Information Engineering, Faculty of Innovation Engineering, Macau University of Science and Technology, Macau 999078, China.
School of Computer Science and Engineering, Faculty of Innovation Engineering, Macau University of Science and Technology, Macau 999078, China.
Life (Basel). 2025 Jun 4;15(6):910. doi: 10.3390/life15060910.
Rice yields are expected to drop significantly due to the increasing spread of rice pests. Detecting rice pests in a timely manner using deep learning models has become a prevalent approach for rapid pest control. However, current datasets related to rice pests often suffer from limited sample sizes or poorly annotated labels, which compromises the training accuracy of deep learning models. Building upon the large-scale IP102 dataset, this study refines the rice pest segment of IP102 by separating adult specimens and larva specimens, acquiring additional pest images via web crawler techniques, and re-annotating all adult samples. The pest category names, originally in English, are replaced with the Latin scientific names of the corresponding families to improve both clarity and scientific accuracy. The resulting dataset, designated RP11, includes 11 adult categories with 4559 images and 7 larval categories with 2467 images. All annotations follow a labeling format compatible with YOLO model training. The sample count in RP11 is approximately four times that of the rice-specific subset in IP102. In this work, YOLOv11 was employed to evaluate RP11's performance, with IP102 serving as a comparison dataset. The results demonstrate that RP11 outperforms IP102 in precision (83.0% vs. 58.9%), recall (79.7% vs. 63.1%), F1-score (81.3% vs. 60.9%), mAP50 (87.2% vs. 62.0%), and mAP50-95 (73.3% vs. 37.9%).
由于水稻害虫的传播日益广泛，预计水稻产量将大幅下降。利用深度学习模型及时检测水稻害虫已成为快速控制害虫的一种普遍方法。然而，当前与水稻害虫相关的数据集往往存在样本量有限或标注标签质量差的问题，这影响了深度学习模型的训练准确性。基于大规模的IP102数据集，本研究通过分离成虫标本和幼虫标本、利用网络爬虫技术获取额外的害虫图像以及重新标注所有成虫样本，对IP102的水稻害虫部分进行了优化。原来用英文表示的害虫类别名称被相应科的拉丁学名所取代，以提高清晰度和科学准确性。由此产生的数据集命名为RP11，包括11个成虫类别，共4559张图像，以及7个幼虫类别，共2467张图像。所有标注都遵循与YOLO模型训练兼容的标注格式。RP11中的样本数量约为IP102中水稻特定子集的四倍。在这项工作中，使用YOLOv11来评估RP11的性能，并将IP102作为比较数据集。结果表明，RP11在精度（83.0%对58.9%）、召回率（79.7%对63.1%）、F1分数（81.3%对60.9%）、mAP50（87.2%对62.0%）和mAP50 - 95（73.3%对37.9%）方面均优于IP102。
