Species recognition technology based on migration learning and data augmentation
Yisheng Song, Zhijie Lin · 2018
At present, the biggest obstacle to the protection of endangered species in China is the lack of basic information. In order to help the relevant departments to collect species information better, a deep convolutional neural network method based on migration learning and data enhancement technology is proposed to realize real-time identification of species. First extract the bottleneck descriptor of the model on the ImageNet dataset, and then use the TensorFlow official script to train the new dataset. In the experiment, using data enhancement technology and real-time image distortion, the Inception3 and MobileNet models were trained respectively, and their own optimal parameters selection scheme was proposed. During the experiment, the best training model MobileNet_1.0_224 was obtained by weighing the time complexity and space complexity of the model. The accuracy of the test set was 89%. Finally, the trained model is transplanted into the Android device to realize the real-time classification of rare animals. The experiment proves that the proposed method has high accuracy and stable running performance.