关键词:

深度学习, 数据增强, 迁移学习, 微调, 自动识别

Abstract:

Species intelligent identification is a hot spot of current research. Based on transfer learning， this paper has developed a spider species intelligent identification platform combining with data augmentation technology. First， to enhance the data based on prominent prospects and traditional methods； second， to use the method of pre-training and fine-tuning in transfer learning to take the VGG-16 model parameters pre-trained by ImageNet as the initial parameters and freeze the convolutional layer， and only the fully connected layer is trained； finally， to combine mobile the research and construction of mobile technology， including the intelligent recognition platforms of Android and WeChat recognition systems， and a background management system. After testing， the system had an average accuracy of more than 95% for the recognition of five types of spiders（4 families and 1 species）. The platform could provide basic identification services， thus effectively reduce the difficulty of species identification， improve the identification rate and the stability of identification. At the same time， it could further collect a large number of species image data uploaded by users， and integrate online expert resources， so as to continuously iterated and improved the identification system. In the related research of arachnids resources， it has important application and theoretical values.

Key words:

 , deep learning, data augmentation, transfer learning, fine-tuning, automatic identification