Tourist Attraction Classification for Supporting Thoughtful Indonesia Program Using Siamese Neural Networks

Abstract

Tourist attractions both on a local scale in Indonesia and on an international scale are very numerous. Nowadays, more and more information on tourist attractions is represented as images rather than text. Tourists are interested in the specific tourist attraction shown in the picture, do not know the attraction's name, and cannot do a text search to get more information about the attraction in question. Convolutional neural networks (CNNs) perform well on large data sets of images. However, due to the diversity of tourist attractions in Indonesia, not all tourist attractions in Indonesia have a large sample image. So, this paper will discuss adopting one-shot learning with the Siamese network to solve the problem of the availability of a small sample of tourist data. Siamese networks are a type of twin network with two or more identical subnets. The settings and weights are the same for all subnets. The parameters of the Siamese network are modified by operating together in all its subnets. In addition, the Siamese network can learn well even with limited input. This study resulted in an image classification of 102 tourist attractions in Indonesia. With each class, five samples resulted in a validation accuracy of 93%.