Mobile Application for Bird Species Identification Using Transfer Learning

Abstract

Bird populations are declining worldwide, and several species have gone extinct in historical times. Hence for ornithologists and birdwatchers, exploration of rarely found bird species has become a challenging task. We have developed a deep learning based android application to help users recognize 260 Species of birds, making bird classification a lot more user-friendly. In this paper, we use Convolutional Neural Networks (CNN) pre-trained on ImageNet Dataset as freeze layers of the network, and train the last output layer, which consists of 260 different classes. CNN models such as EfficientNet-lite0, Xception, MobilenetV2, ResNet-50, InceptionV3, and InceptionResNetV2 have been compared based on the accuracy, and working of the mobile app is explained. Maximum accuracy of 99.82% on train data and 98.61% on test data is achieved.