Abstract:
Bird species classification is significant for the ecological studies and the biodiversity
conservation. This paper introduces a novel method using deep learning to
automatically classify the bird species from the images. We study images of birds and
apply the convolutional neural network to examine the basic features like feather and
beak shape. This paper deals with the bird classification using deep learning methods
by using a novel dataset coming from Kaggle containing the images of the eighteen
native bird species in Bangladesh. The dataset, which is made up of 2704 carefully preprocessed images, goes through the rigorous pre-processing steps and is then divided
into different subsets for the purpose of model evaluation, which includes training,
testing and validation sets. Different deep learning architectures, such as MobileNetV2,
DenseNet201, VGG19, and Xception, are carefully studied individually to check their
effectiveness in the task of bird classification. Besides, hybrid models, which combine
the different architectural paradigms, are built in a way to find the synergies and thus
the classification accuracy is improved. A large number of evaluation metrics such as
accuracy, precision, recall and F1-score are used to precisely compare the performance
of the models. Findings reveal nuanced intricacies in model performance: The
MobileNetV2 gets an accuracy of 97. 15%, VGG19 87.22%, DenseNet201 98.27%,
Xception 98.33%, and the Hybrid MobileNetV2 and DenseNet201 model appears as
the winner, with the highest accuracy of 99.50%. This superb achievement shows the
possibility of hybrid architectures to beat the single-model approaches, giving us
information how the interrelation between architectural complexity, computational
resources, and task-specific requirements
