Improvements Based on ShuffleNetV2 Model for Bird Identification

Abstract

Bird identification and classification are of great significance in bird conservation. Through proper bird classification, the changes in bird populations in a given area can be effectively predicted, thereby ensuring their effective protection. Nowadays, deep learning has achieved high accuracy in classifying bird images. However, most existing models suffer from poor generalization ability and high complexity. To enhance the model’s generalization ability, this paper constructed a large dataset consisting of 275 bird species and the dataset contains bird pictures in different situations, such as rainy days, foggy days. To reduce model complexity, a lightweight neural network based on ShuffleNetV2 was constructed. In ShuffleNetV2 network, there is no feature fusion module and efficient attention mechanism to assist the feature learning of the model. Therefore, this paper adds a feature fusion module and two attention mechanisms to make up for this shortcoming. A multi-channel feature fusion structure (MCF) was adopted to improve the network’s adaptability to extract information from multiple channel scales. By introducing Squeeze-and-Excitation (SE) Module and Coordinate attention (CA) in the Block module, the model’s ability to refine the global features was enhanced. The experimental results show that the accuracy of the model in identifying 275 bird species on the self-built dataset is 92.3%, which is 6.1% higher than the accuracy of ShuffleNetV1 (86.2%) and 1.8% higher than the accuracy of ShuffleNetV2 (90.5%). At the same time, with smaller parameters and floating point operations (FLOPs), its accuracy is 1.2% higher than ResNet50’s accuracy (ResNet50’s accuracy is 91.1%), which can save the cost better.

Keywords

• Deep learning
• image classification
• lightweight
• attention mechanism
• multi-channel feature fusion