The brown rice planthopper ( Nilaparvata lugens Stal) is one of the main pests of rice. The rapid and accurate detection of brown rice planthoppers (BRPH) can help treat rice in time. Due to the small size, large number and complex background of BRPHs, image detection of them is challenging. In this paper, a two-layer detection algorithm based on deep learning technology is proposed to detect them. The algorithm for both layers is the Faster RCNN (regions with CNN features). To effectively utilize the computing resources, different feature extraction networks have been selected for each layer. In addition, the second layer detection network was optimized to improve the final detection performance. The detection results of the two-layer detection algorithm were compared with the detection results of the single-layer detection algorithm. The detection results of the two-layer detection algorithm for detecting different populations and numbers of BRPHs were tested, and the test results were compared with YOLO v3, a deep learning target detection network. The test results show that the detection results of the two-layer detection algorithm were significantly better than those of the single-layer detection algorithm. In the tests for different numbers of BRPHs, the average recall rate of this algorithm was 81.92%, and the average accuracy was 94.64%; meanwhile, the average recall rate of YOLO v3 was 57.12%, and the average accuracy rate was 97.36%. In the experiment with different ages of BRPHs, the average recall rate of the algorithm was 87.67%, and the average accuracy rate was 92.92%. In comparison, for the YOLO v3, the average recall rate was 49.60%, and the average accuracy rate was 96.48%.

中文翻译：

---

基于深度学习的糙米飞虱（Nilaparvata lugens Stal）检测

糙米飞虱（Nilaparvata lugens Stal）是水稻的主要害虫之一。快速准确检测糙米飞虱（BRPH）有助于及时处理水稻。由于BRPHs体积小、数量多、背景复杂，对它们的图像检测具有挑战性。在本文中，提出了一种基于深度学习技术的两层检测算法来检测它们。两层的算法都是 Faster RCNN（具有 CNN 特征的区域）。为了有效利用计算资源，每一层都选择了不同的特征提取网络。此外，优化了第二层检测网络以提高最终检测性能。将两层检测算法的检测结果与单层检测算法的检测结果进行比较。对检测不同种群和数量的BRPHs的两层检测算法的检测结果进行了测试，并将测试结果与深度学习目标检测网络YOLO v3进行了对比。测试结果表明，两层检测算法的检测结果明显优于单层检测算法。在对不同数量的BRPHs的测试中，该算法的平均召回率为81.92%，平均准确率为94.64%；同时，YOLO v3的平均召回率为57.12%，平均准确率为97.36%。在不同年龄BRPHs的实验中，该算法的平均召回率为87.67%，平均准确率为92.92%。相比之下，对于 YOLO v3，平均召回率为 49.60%，