Fine-grained Image Classification of Birds

Computer vision and ecological research of the important topic of bird image classification. Bird species are so varied, and looks could so easily be similar, that fine classification is faced with many challenges. In this study, I extensively compare four models in order to select the most appropriate model for high resolution bird image classification. This study analyzes the performance of each model in the training stability, classification accuracy, and computational efficiency to help select the best model, which strikes a tradeoff between resource consumption and accuracy. We further investigate how to achieve better classification accuracy, including the exploration of model structure optimization, refined feature extraction (e.g. attention mechanisms) and the use of multi task learning to further enhance fine grained performance. Our goal in this work is to suggest a more effective fine grained image classification framework for capturing subtle features to distinguish similar birds at higher accuracy. To keep the research interesting for research and applicable in real situations, we coupled a large scale bird image dataset (CUB-200-2011) with many images of Malaysian birds. Data augmentation and transfer learning techniques have significantly improved the model’s generalization ability. Finally, our method achieves leading classification accuracy on standard datasets and effectively discriminate between similar categories. Furthermore, we also analyzed the model’s performance against different bird species and environmental conditions, and provide valuable references for future bird monitoring and conservation work.