He Zhangzhang, Chen Xinyue, Gao Ying, Zhang Yu, Guo Yuheng, Zhai Tong, Wei Xiaochen, Li Huan, Zhu Haipeng, Fu Yongkun, Zhang Zhiliang
College of Food and Biology, Jingchu University of Technology, Jingmen, Hubei, China.
School of Computer Science, Yangtze University, Jingzhou, China.
Front Plant Sci. 2025 Aug 21;16:1663813. doi: 10.3389/fpls.2025.1663813. eCollection 2025.
Thrips can damage over 200 species across 62 plant families, causing significant economic losses worldwide. Their tiny size, rapid reproduction, and wide host range make them prone to outbreaks, necessitating precise and efficient population monitoring methods. Existing intelligent counting methods lack effective solutions for tiny pests like thrips. In this work, we propose the Thrip Counting and Detection Network (TCD-Net). TCD-Net is an fully convolutional network consisting of a backbone network, a feature pyramid, and an output head. First, we propose a lightweight backbone network, PartialNeXt, which optimizes convolution layers through Partial Convolution (PConv), ensuring both network performance and reduced complexity. Next, we design a lightweight channel-spatial hybrid attention mechanism to further refine multi-scale features, enhancing the model's ability to extract global and local features with minimal computational cost. Finally, we introduce the Adaptive Feature Mixer Feature Pyramid Network (AFM-FPN), where the Adaptive Feature Mixer (AFM) replaces the traditional element-wise addition at the P level, enhancing the model's ability to select and retain thrips features, improving detection performance for extremely small objects. The model is trained with the Object Counting Loss (OC Loss) specifically designed for the detection of tiny pests, allowing the network to predict a small spot region for each thrips, enabling real-time and precise counting and detection. We collected a dataset containing over 47K thrips annotations to evaluate the model's performance. The results show that TCD-Net achieves an F1 score of 85.67%, with a counting result correlation of 75.50%. The model size is only 21.13M, with a computational cost of 114.36 GFLOPs. Compared to existing methods, TCD-Net achieves higher thrips counting and detection accuracy with lower computational complexity. The dataset is publicly available at github.com/ZZL0897/thrip_leaf_dataset.
蓟马可危害62个植物科的200多种植物，在全球造成重大经济损失。它们体型微小、繁殖迅速且寄主范围广泛，容易爆发灾害，因此需要精确高效的种群监测方法。现有的智能计数方法缺乏针对蓟马等微小害虫的有效解决方案。在这项工作中，我们提出了蓟马计数与检测网络（TCD-Net）。TCD-Net是一个全卷积网络，由骨干网络、特征金字塔和输出头组成。首先，我们提出了一个轻量级骨干网络PartialNeXt，它通过部分卷积（PConv）优化卷积层，在保证网络性能的同时降低了复杂度。接下来，我们设计了一种轻量级的通道-空间混合注意力机制，以进一步细化多尺度特征，以最小的计算成本增强模型提取全局和局部特征的能力。最后，我们引入了自适应特征混合器特征金字塔网络（AFM-FPN），其中自适应特征混合器（AFM）取代了P层传统的逐元素相加，增强了模型选择和保留蓟马特征的能力，提高了对极小物体的检测性能。该模型使用专门为检测微小害虫设计的目标计数损失（OC Loss）进行训练，使网络能够为每个蓟马预测一个小斑点区域，实现实时精确的计数和检测。我们收集了一个包含超过4.7万个蓟马标注的数据集来评估模型的性能。结果表明，TCD-Net的F1分数达到85.67%，计数结果相关性为75.50%。模型大小仅为21.13M，计算成本为114.36 GFLOPs。与现有方法相比，TCD-Net以更低的计算复杂度实现了更高的蓟马计数和检测准确率。该数据集可在github.com/ZZL0897/thrip_leaf_dataset上公开获取。
