Transfer Learning Framework Using CNN Variants for Animal Species Recognition

Automatic recognition of species is the task of identifying and counting animal or bird species from pictures taken from camera traps. Such recognition systems help ecologists automatically analyse and monitor animal behaviour without human intervention. In this work, we exploit transfer learning using convolutional neural networks (CNN) to identify animal species. The overall framework uses a pre-trained network as a backbone to learn general features before the classification layer. Using this framework, several models are developed by using the EfficientNet, ResNet, Inception, and VGG as the backbone networks. Each model is trained over the animal species dataset. The models are evaluated over test data, and it is observed that the EfficientNet-based model exhibits the best performance.