Recognition of Local Birds Using Different CNN Architectures with Transfer Learning

Abstract

The global world is dependent and integrated with all the ecosystems. To survive in the race, we must need to know about the birds and their habitats and importance to the existence of the human race on earth. However, it is difficult to recognize several species of birds, animals, and so on. In this paper, we presented a methodology to recognize local birds of Bangladesh using transfer learning techniques. The whole research work has been done using transfer learning in six different CNN architecture namely DenseNet201, InceptionResNetV2, MobileNetV2, ResNet50, ResNet152V2, and Xception. As to defeat the lack of much image data, augmentation is performed on the collected image data too. All the models are trained by 2800 data images and tested by 700 data images. Among all the discussed models, MobileNetV2 model exhibits the best performance in terms of various indicators such as F1-score, precision, recall, and accuracy. The accuracy, precision, recall, and F1 Score of MobileNetV2 are 96.71%, 96.93%, 96.71%, 96.75%. Then a comparative analysis has been performed for this work among the approaches as well. The obtained result shows that the working method is optimal and efficient for recognizing local birds of Bangladesh.