应用迁移学习的提高小数据集蝴蝶种类识别精确度的方法
A Method of Improving the Accuracy of Butterfly Species Identification in Small Data Sets Using Transfer Learning

野生蝴蝶品种繁多、分布广泛，且其对于生存环境的变化较为敏感，监测范围内的蝴蝶种群生存情况，可以为监测衡量该地区生态环境变化和衡量生态环境的质量提供帮助。现有的蝴蝶种类数据集大多数据量较小，识别的准确率较低。针对该问题，本文提出了基于VGG16模型的迁移学习及微调方法，对蝴蝶图像进行种类识别，以提高识别准确率。首先对蝴蝶种类数据集进行数据增强，再利用大型图像数据集对VGG16模型进行预训练，对预训练模型进行参数迁移，对卷积层和池化层进行“冻结”，修改全连接层和分类层，再“解冻”部分卷积层对参数进行微调，得到识别结果。实验证明：通过迁移学习加微调的方法，该网络对于蝴蝶种类识别的准确率得到有效提高，识别准确率从最初的76.67%，在迁移及微调后达到83.95%。
There are many species and wide distribution of wild butterflies, and they are sensitive to changes in the living environment. The survival of butterfly populations within the monitoring range can provide help for monitoring and measuring changes in the ecological environment in the region and measuring the quality of the ecological environment. Most of the existing butterfly species data sets have small data volume and low recognition accuracy. In order to solve this problem, this paper proposes a migration learning and fine-tuning method based on the VGG16 model to recognize the butterfly image in order to improve the recognition accuracy. First, the butterfly species data set is enhanced, and then the large image data set is used to pre-train the VGG16 model, migrate the parameters of the pre-training model, “freeze” the convolution layer and pool layer, modify the full connection layer and classification layer, and “thaw” part of the convolution layer to fine-tune the parameters to get the recognition results. The experiment shows that the accuracy of the network for butterfly species recognition has been effectively improved by the method of migration learning and fine tuning, and the recognition accuracy has reached 83.95% after migration and fine tuning from 76.67% at the beginning.