Kang Huangyi, Ai Luxin, Zhen Zengyi, Lu Baojia, Man Zhangli, Yi Pengyu, Li Manzhou, Lin Li
College of Information and Electrical Engineering, China Agricultural University, Beijing 100083, China.
College of Plant Protection, China Agricultural University, Beijing 100083, China.
Insects. 2023 Jul 24;14(7):660. doi: 10.3390/insects14070660.
In this work, an attention-mechanism-enhanced method based on a single-stage object detection model was proposed and implemented for the problem of rice pest detection. A multi-scale feature fusion network was first constructed to improve the model's predictive accuracy when dealing with pests of different scales. Attention mechanisms were then introduced to enable the model to focus more on the pest areas in the images, significantly enhancing the model's performance. Additionally, a small knowledge distillation network was designed for edge computing scenarios, achieving a high inference speed while maintaining a high accuracy. Experimental verification on the IDADP dataset shows that the model outperforms current state-of-the-art object detection models in terms of precision, recall, accuracy, mAP, and FPS. Specifically, a mAP of 87.5% and an FPS value of 56 were achieved, significantly outperforming other comparative models. These results sufficiently demonstrate the effectiveness and superiority of the proposed method.
在这项工作中，针对水稻害虫检测问题，提出并实现了一种基于单阶段目标检测模型的注意力机制增强方法。首先构建了一个多尺度特征融合网络，以提高模型在处理不同尺度害虫时的预测精度。然后引入注意力机制，使模型能够更专注于图像中的害虫区域，显著提升了模型的性能。此外，还为边缘计算场景设计了一个小型知识蒸馏网络，在保持高精度的同时实现了较高的推理速度。在IDADP数据集上的实验验证表明，该模型在精度、召回率、准确率、平均精度均值（mAP）和每秒帧数（FPS）方面优于当前最先进的目标检测模型。具体而言，实现了87.5%的mAP和56的FPS值，显著优于其他对比模型。这些结果充分证明了所提方法的有效性和优越性。
