A MOBILE APPLICATION FOR BIRD SPECIES RECOGNITION USING DEEP LEARNING TECHNIQUES

ALVAREZ BERAI, UCHONG (2023) A MOBILE APPLICATION FOR BIRD SPECIES RECOGNITION USING DEEP LEARNING TECHNIQUES. [Final Year Project Report] (Unpublished)

	PDF Alvarez Berai 24pgs.pdf Download (528kB)
	PDF Alvarez Berai ft.pdf Download (2MB)

Abstract

The development of a mobile application for bird species identification has become a significant area of research due to its potential to engage bird enthusiasts, facilitate citizen science initiatives, and contribute to conservation efforts. In this project, we present a comprehensive approach to building a mobile application for bird species identification, encompassing the development of a machine learning model in Jupyter Lab, model optimization using Google Cloud Platform with Vertex AI, then building the API to process the request then send prediction and the creation of an application through Android Studio. The system utilizeses a convolutional neural network (CNN) architecture through transfer learning trained on a dataset of local Sarawak birds. The images were preprocessed to enhance relevant features and remove noise. The CNN transfer learning model chosen was EfficientNet-B4. To validate the robustness of the model, 5 famous transfer learning model (EfficientNet-B4, ResNet-50, InceptionV3, MobileNetV2 and VGG16) were compared through parameters such as training time, training loss, validation loss, training accuracy, validation accuracy, test loss, and test accuracy. The chosen EfficientNet-B4 model achieved an accuracy of 93.06%, test loss of 0.3156, training accuracy of 96.81%, training loss of 0.1081, validation accuracy of 93.87%, validation loss of 0.2853 and average training time of 285 seconds for 1 epoch with a total of 20 epochs, demonstrating its effectiveness in accurately classifying the bird species. The successful completion of this project makes a valuable contribution to the field of bird species identification and conservation. The developed mobile application provides an accessible and user-friendly tool for bird enthusiasts, researchers, and citizen scientists, promoting active engagement in bird conservation efforts and facilitating data collection for monitoring bird populations and habitats. Future directions for this project include expanding the model's dataset to cover a broader range of bird species and environmental conditions, as well as integrating real-time updates and environmental sensor data to enhance the application's functionality and provide users with dynamic and contextually rich bird identification information.

Actions (For repository members only: login required)

View Item