关键词:  蝴蝶, 卷积神经网络, GoogLeNet, 图像识别, 迁移学习

Abstract: 【Aim】Based on feature extraction by convolutional neural network, this study aims to explore the feasibility of accurately identifying wild butterflies by integrating the GoogLeNet network model with digital image feature information.【Methods】Nine species of Lepidoptera butterflies belonging to 4 families with high richness in Liancheng National Nature Reserve, Gansu Province, northwestern China were selected, among them, 3 704 adult specimen images were used to construct a butterfly image recognition model, and 376 butterfly images with interfering factors were collected to screen out the optimal model with the highest recognition precision. By modifying the internal structure and adjusting the parameters of the GoogLeNet network model, and using PhotoShop for post-processing of images, the activation function was optimized, the number of parameters was reduced, and the calculation speed and feature extraction ability were improved to accurately identify butterfly species.【Results】 Ten models were trained using MATLAB software platform, among them four models demonstrating the optimal performance in terms of model size, accuracy and recognition efficiency were screened. The best-performing model was selected to construct a diagnostic system for the nine representative butterfly species including Melanargia halimede, Pieris rapae, Colias fieldii, Papilio xuthus, Gonepteryx rhamni, Seokia pratti, Hipparchia autonoe, Aglais urticae and Neptis alwina. After 60-min training on the platform, the model achieved 100.0% diagnostic accuracy on the validation dataset. Even when artificial interference items were introduced, the modified GoogLeNet model maintained an accuracy of 94.7%. Upon completion of training with multiple data models, we confirmed that increasing dataset size and image quality directly improved the model’s recognition precision.【Conclusion】 To a certain extent, the model constructed in this study overcomes the shortcomings of identifying butterfly species using a single feature, improves the accuracy of wild butterfly identification, and has strong model stability. It can provide technical support for the intelligent identification of butterflies and lay the foundation for the subsequent development of the recognition platform.

Key words: Butterfly, convolutional neural network, GoogLeNet, image recognition, transfer learning