Abstract: Lepidopteran pests are the most important and common pests in vegetable crops. Due to the complex vegetable garden background, sunlight and pest gestures, conventional methods of pests image recognition, detection and counting may not be accurate. In order to recognize, detect and count lepidopteran pests quickly and accurately in vegetable garden,  recognition model, detect and count model were introduced based on deep convolutional neural network, respectively. According to five kinds of common lepidopteran pests, such as Pieris rapae, Helicoverpa armigera, Spodoptera exigua, Plutella xylostella and Spodoptera litura, this paper  built recognition image dataset, detection and counting image dataset and then carried out experiments, the average recognition accuracy could reach 94.5% and the mean average precision of detection was 76.6%, the comparison to conventional methods proved that our method was preeminent. The experiment results showed that the proposed method was a feasible way to recognize, detect and count vegetable lepidopteran pests, and it  reached the level of practical application.

Key words: lepidoptera pest, auto recognition, detection and counting, deep learning, convolutional neural network

摘要： 鳞翅目害虫是蔬菜作物中最重要且常见的一类害虫。由于受到复杂田间背景，光照及害虫姿态等的影响，传统的害虫自动识别与检测计数方法准确率比较低。为实现在田间快速准确地对目标害虫进行自动识别和检测计数，分别提出了基于深度卷积神经网络的识别模型和检测计数模型。针对菜粉蝶、棉铃虫、甜菜夜蛾、小菜蛾、斜纹夜蛾这5种常见且容易混淆的蔬菜鳞翅目害虫，构建了分类识别数据集和检测计数数据集，并分别进行了实验，平均识别率达到94.5%，检测均值平均精度(mAP)达到76.6%，与传统方法相比，证明了此方法的优越性。实验结果表明，该方法对于蔬菜鳞翅目害虫的识别和检测计数是可行的，且达到了实际应用水平。

关键词: 鳞翅目害虫, 自动识别, 检测计数, 深度学习, 卷积神经网络