基于迁移学习和Mask R-CNN的稻飞虱图像分类方法

针对当前稻飞虱图像识别研究中自动化程度较低、识别精度不高的问题,提出了一种基于迁移学习和Mask R-CNN的稻飞虱图像分类方法。首先,根据稻飞虱的生物特性,采用本团队自主研发的野外昆虫图像采集装置,自动获取稻田稻飞虱及其他昆虫图像;采用VIA为数据集制作标签,将数据集分为稻飞虱和非稻飞虱两类,并通过迁移学习在Res Net50框架上训练数据;最后,基于Mask R-CNN分别对稻飞虱、非稻飞虱、存在干扰以及存在黏连和重合的昆虫图像进行分类实验,并与传统图像分类算法(SVM、BP神经网络)和Faster R-CNN算法进行对比。实验结果表明,在相同样本条件下,基于迁移学习和Mask R-CNN的稻飞虱图像分类算法能够快速、有效识别稻飞虱与非稻飞虱,平均识别精度达到0. 923,本研究可为稻飞虱的防治预警提供信息支持。

Rice Planthopper Image Classification Method Based on Transfer Learning and Mask R-CNN

In order to deal with the problem of low automation and low recognition accuracy in the current rice planthopper image recognition research, an image classification algorithm based on transfer learning and Mask R-CNN was proposed. Firstly, according to biological characteristics of rice planthopper, the self-developed wild insect image collection device was utilized to obtain insect images automatically. Then, the dataset was divided into two categories: rice planthopper and non-rice planthopper by the image label tool VIA, and was trained in the ResNet50 framework with transfer learning. Finally, the Mask R-CNN image classification experiments were carried out based on rice planthopper images, non-rice planthopper images, insect images with disturbances and those images which were adhesive and overlapping, respectively. Moreover, experiments were compared with SVM, BP neural network, which were traditional image classification algorithms, and Faster R-CNN algorithm. Experiment results showed that the method based on transfer learning and Mask R-CNN could distinguish the rice planthopper and non-rice planthopper images effectively and the average classification accuracy reached 0.923 under the same sample conditions, which could provide information support for the prevention and early warning of rice planthoppers.