Large Scale Bird Species Classification Using Convolutional Neural Network with Sparse Regularization

Bird is one of many creatures with so many species worldwide. Every bird species has many differences, from the shape of its limbs, behavior, and food. Sometimes some scientists have difficulty when making their observations. To this end, accurate artificial intelligence system using deep learning has to be developed to help scientists detect the existence of certain birds automatically. Convolutional Neural Networks (CNN) can help classify the species of birds based on their characteristics. Nevertheless, to train a CNN that can classify the species of birds correctly, it requires a large dataset. 285-birds is a suitable dataset consisting of 43780 digital images with 285 class labels. Bird classification training utilizes a Transfer-learning method using pre-trained weights on ImageNet such as AlexNet and Resnet34. In addition, the hyper-parameters of training of 100 epochs, an Adam optimizer with a learning rate of 0.00001, a batch size of 64, and a Cross-Entropy loss. Utilizing a Sparse regularization in loss function improves the performance model by reducing unnecessary features while also focussing on the important ones. The result of this research is that ResNet model with sparse regularization can recognizes large number of wild birds with the most robust performance compared to other models and thus we suggest our proposed methods to be applied in large scale birds recognition systems.