Abstract: In order to solve the problem of rice insect pests of artificial identification difficult and time-consuming, an automatic identification method of rice insect pests was studied. Rice pests images of Diostrombus politus Uhler, Nezara viridula, Naranga aenescens Moore, Riptortus pedestris, Sesamia inferens were selected from 2 500 images to construct weed rice pests set, and the features of the rice pest dataset were automatically extracted using deep convolutional neural network, and then the optimal weights of documents were got by training the stochastic gradient descent method. During the test, the trained model was used to select test sets, and YOLOv3 was also compared with Faster-RCNN and SSD algorithm. The results showed that average precision (AP) of YOLOv3 algorithm was the highest among all the deep learning models. the AP of Diostrombu politus Uhler, Nezara viridula,Naranga aenescens Moore, Riptortus pedestris, Sesamia inferens were 97.40%, 88.76%, 85.74%, 92.96% and 94.78%, respectively, and mAP in the five kinds of rice insect pests was 91.93%. The mAP of YOLOv3 increased by 1.43 percent and the detection time per image decreased by 833.68 ms compared with Faster-RCNN algorithm. The mAP of proposed method was 5.56 percent higher than that of SSD algorithm, and the detection time per image decreased by 2.90 ms. Five types of relatively representative images of rice insect pests were selected for testing. For the cases of leaf occlusion targets and similar backgrounds, YOLOv3 algorithm could correctly identify without missing or wrong detection, and the recognition precision was greater than 98%. It was feasible to introduce YOLOv3 algorithm into the identification of rice insect pests under complex field conditions, and it had higher precision and faster detection speed, which can accurately identify rice insect pests. It was of great significance for promoting fine cultivation of rice pest control and variable field spraying.

Key words: convolution neural network, rice pest, faster-RCNN, SSD, YOLOv3

摘要： 为解决人工识别水稻虫害难度大、费时费力的问题，研究了一种自动识别水稻虫害的方法。选取2 500张红袖蜡蝉、稻绿蝽、稻螟蛉、点蜂缘蝽、大螟图片建立数据库，利用深度卷积神经网络提取水稻虫害数据集特征，采用随机梯度下降法训练，得到最优权重文件。测试训练好的模型，并对比Faster-RCNN、SSD和YOLOv3算法的效果。结果表明，YOLOv3算法的平均精度最高，其在红袖蜡蝉、稻绿蝽、稻螟蛉、点蜂缘蝽、大螟5种水稻虫害中分别为97.40%、88.76%、85.74%、92.96%、94.78%，五类水稻虫害mAP为91.93%。与Faster-RCNN算法相比，平均准确率高1.43个百分点，单张图像检测耗时减少853.68 ms；与SSD算法相比，平均准确率高5.56个百分点，单张图像检测耗时减少2.9 ms。选择5类比较具有代表性的水稻虫害图片进行测试，对于叶片遮挡目标和相似背景等情况，YOLOv3算法能够正确识别不会出现漏检错检，且识别准确率大于98%。将YOLOv3算法引入田间复杂情况下的水稻虫害识别是可行的，具有较高的平均准确率以及较快的检测速度，能够准确识别水稻虫害，这对于水稻虫害防治和田间喷药等方面具有重要意义。

关键词: 卷积神经网络, 水稻虫害, Faster-RCNN, SSD, YOLOv3