Phuket Landmark Recognition using Fine-Tuned Convolutional Neural Network

In this paper, we proposed the application of machine learning for Phuket landmark recognition. The dataset includes 15 landmarks in Phuket by web scraping images. The system is trained with pre-trained convolutional neural network models in the families of DenseNet, EfficientNet, Inception, MobileNet, ResNet, and Xception. We conducted three fine-tuning weights experiments (freezing weights, half unfreezing weights, and all unfreezing weights) to identify the optimal performance model. The best recognition result in each model will be chosen as a candidate model. The forward model selection process selected a few representative models from a set of candidate models to obtain the more robust classifier. Ensemble learning with a soft voting classifier algorithm was used to boost the recognition system's performance. The web application was developed to receive image data from the user. The image was transferred to the process at the web server. Information such as histories, maps, or nearby recommended spots are retrieved from the database for users who want more details about the landmark location in Phuket.