张善文 邵彧 齐国红 许新华

摘要： 田间作物害虫检测是精确防治虫害和减少农药使用量的前提。由于田间害虫种类多，同种害虫个体间差异大，田间同一只害虫的大小、颜色、姿态、位置和背景变化多样、无规律，而且田间背景复杂、对比度低，使得传统的作物害虫检测方法的性能不高。现有的基于深度学习的作物害虫检测方法需要大量高质量的标注训练样本，而且训练时间长。在VGG16模型的基础上，本研究提出一种基于多尺度注意力卷积网络（Multi-scale convolutional network with attention， MSCNA）的作物害虫检测方法。在MSCNA中，多尺度结构和注意力模型用于提取多尺度害虫检测特征，增强对形态较小害虫的检测能力;在训练过程中引入二阶项残差模块，减少网络损失和加速网络训练。试验结果表明，该方法能较好地检测到农田中各种各样、大小不同的害虫，检测平均准确率为92.44%。说明该方法能够实现自然场景下作物害虫的精准检测，可应用于田间作物害虫自动检测。

关键词： 作物害虫检测;注意力机制;卷积神经网络;多尺度注意力卷积网络

中图分类号： TP391.41;S432 文献标识码： A 文章编号： 1000-4440（2021）03-0579-10

Crop pest detection based on multi-scale convolutional network with attention

ZHANG Shan-wen， SHAO Yu， QI Guo-hong， XU Xin-hua

（School of Electronics and Information Engineering， Zhengzhou SIAS University， Zhengzhou 451150， China）

Abstract： Detection of crop pests in field is the prerequisite for accurate pest control and reduction of pesticide dosage. The performance of the traditional detection methods for crop pests is not high， due to the reasons such as various varieties of pests in the field， the difference between different pest individuals of the same variety is great. Besides， the size， color， posture， position and background of the same pest in the field are various and irregular， and the field background is complex and has low contrast. The existing crop pest detection methods based on deep learning require a large number of labeled training samples with high quality， and the training time is long. A multi-scale convolutional network with attention （MSCNA） method based on VGG16 model was proposed for crop pest detection. In MSCNA， the multi-scale structure and attention model were used to extract the detection features of pests on multi-scale and to enhance the ability in detecting smaller pests. Second-order term residual module was introduced in the training process to reduce network loss and accelerate network training. The experimental results showed that， the proposed method could detect various pests with different sizes in the farmland preferably， and the average detection accuracy was 92.44%. The results indicated that this method can detect crop pests accurately in natural scenes and can be applied in the automatic detection of crop pests in the field.

Key words： crop pest detection;attention mechanism;convolutional neural network （CNN）;multi-scale convolutional neural network with attention （MSCNA）

作物害蟲检测和识别是害虫防治的一个重要步骤。目前已有很多作物害虫检测和识别方法。Martineau等[1]综述了44种昆虫分类方法，并介绍了害虫图像采集、特征提取和测试数据集构建等。Yaakob等[2]提取昆虫图像的形状特征的6种不变矩，并验证了该方法的有效性。Fedor等[3]介绍了基于数字图像分析和人工神经网络系统的半自动害虫识别和监测工具，并提取害虫的头部、锁骨、翅膀、产卵器长度和宽度等形态特征，构成特征向量，最后由神经网络进行害虫识别。……