基于深度学习的害虫图像识别与分类方法研究

害虫侵扰一直是农业生产中回避不了的问题，每年都会造成巨大的经济损失。为了能够有效地预防和控制病虫害问题，需要实现对农田害虫的快速、准确识别，对此提出了一种基于深度学习的农田害虫识别方法，可按照害虫特征区分二化螟、白背飞虱、褐飞虱属、八点灰灯蛾、蟋蟀等多种害虫类别。一阶段对害虫数据集进行分析校正，加入图像整理、剪切等操作，合理划分数据集，添加一系列数据增强处理，进行农田害虫的训练检测。二阶段为增加数据集规模，使用EfficientNet网络对未标注图片进行识别分类，得到伪标签后继续半监督学习。最后，将分类的验证集和训练集合并，做进一步训练加强。实验结果表明，该模型对相关害虫识别效率高，识别效果好，可移植性强，可为农作物害虫的高效快速检测提供参考。

Research on Recognition and Classification Method of Pest Images Based on Deep Learning

Pest infestation has always been an unavoidable problem in agricultural production, which causes huge economic losses every year. In order to effectively prevent and control diseases and insect pests, it is necessary to realize the rapid and accurate identification of farmland pests. For this, a deep learning-based farmland pest identification method is proposed. According to the characteristics of the pests, a variety of pest categories can be screened out, such as Dillite borer, white-backed planthopper, brown planthopper, eight-pointed gray light moth, cricket and so on. In the first stage, the pest data set is analyzed and corrected, and operations such as image sorting and cutting are added, the data set is divided reasonably, and a series of data enhancement processing is added to conduct training and detection of farmland pests. In the second stage, to increase the scale of the dataset, use the EfficientNet network to identify and classify unlabeled pictures, and continue semi-supervised learning after obtaining pseudo labels. Finally, the classified validation set and training set are merged for further training enhancement. The experimental results show that the model has high identification efficiency for related pests, good identification effect and strong portability, which can provide a reference for efficient and rapid detection of crop pests.