Abstract:

Objective: In this study, based on the bird images, we construct a bird population identification model with the deep convolutional neural network theory by combining the global and local features fusion method, in order to provide a new approach for monitoring and management of forests and wetlands. Method: First of all, according to the physiological process of the object identification of human from entireness to part, the jump structure was used to implement the interaction between global and local information. In the proposed model, two model frameworks are mainly used to extract the global and local features of birds, and the jump structure is used to propose the fusion module structure for feature fusion, which transfers the global feature information to local feature extraction module. In the training stage of the model, we need to provide the labeling information on the local parts of birds, while in the test stage, we use Faster R-CNN model to automatically extract the labeling information on the local parts of birds. Secondly, we discussed the effects of different bird local image information on the model. Finally, the validity and applicability of the model are verified by comparing different network classification models and bird datasets. Result: The bird species classification model proposed in this paper has high classification accuracy, and the overall classification accuracy is over 90%. For the image information of different parts of a bird, the classification accuracy of the model shows a certain difference, among which the overall classification accuracy of the network classification model based on the bird's head image is the highest. The Faster R-CNN model has a high accuracy in bird part image locating. There is little difference in the overall accuracy between the manually labeled local image tag and the local image tag predicted by Faster R-CNN model in the test stage. Compared with the overall classification accuracy of the network classification models such as Inception-V1、ResNet-101、DenseNet-121 and Bilinear CNN, the overall classification accuracy of the model proposed in this paper is relatively high, that verifies the effectiveness of the proposed classification model of bird. Compared with the classification accuracy by using NABirds bird dataset, the overall classification performance of the proposed model is better, which verifies the applicability of the proposed model. Conclusion: The proposed bird species classification model has good identification results and effectiveness, which can provide a reasonable and effective basis for monitoring and management of forests and wetlands.

Key words: bird identification, deep convolutional neural network, global and local components