关键词: 林业图像, 微调, 森林管护, 卷积神经网络, 迁移学习

Abstract:

Objective: A fine-tuning convolution neural network method based on transfer learning technology was proposed for automatic classification of forestry images, which was expected to be beneficial for forestry management department to make reasonable disposal plan or scheduling decision in a series of forest incidents, so as to improve the level of forest management and to protect forest resources and ecological safety. Method: Pre-training with the large-scale image set(ImageNet), CaffeNet was trained with fine-tuning by using the forestry image data. The first 5 layers' parameters of the model were obtained by migration, including convolution layer, activation function and pool layer; and the parameters of full connection layer and Softmax were determined by training. Result: CaffeNet model of the pre-training fine-tuning convolution neural network presented a good classification accuracy for forestry images. With four classes forestry image data sets established according to forestry business requirements, the average recognition accuracy was stable to 97.5% after a certain number of iterations. Further feature visualization showed that the trained feature maps obtained from different layers of the deep convolution neural network got the forestry image classification ability from different aspects. Compared with the traditional feature extraction method, our method's recognition rate was also increased by 10.8% even adding one class into the number of forestry categories. Conclusion: It might be feasible to classify the forestry images by using CaffeNet model. Compared with the traditional feature extraction and recognition methods, the forest image classification model based on convolution neural network could have a stronger feature extraction ability and classification ability. So it would play an important role in the application of forest management in the future.

Key words: forestry images, fine-tuning, forest manage and protect, convolutional neural networks, transfer learning