Automatic Indonesian Image Caption Generation using CNN-LSTM Model and FEEH-ID Dataset

Abstract

Image captioning is a challenge in computer vision research. This paper extends research on automatic image captioning generation in the Indonesian dimension. Description in Indonesian sentences is generated for unlabeled images. The dataset used is FEEH-ID, this is the first Indonesian image captioning dataset. This research is crucial due to unavailability of a corpus for image captioning in Indonesian. This paper will compare the experimental results in the FEEH-ID dataset with English, Chinese and Japanese datasets using the CNN and LSTM models. The performance of the model proposed in the test set provides promising results of 50.0 for the BLEU-1 score and 23.9 for BLEU-3, which is above average of the Bleu evaluation results in other language datasets. The merging model between CNN and LSTM displays pretty good results for the FEEH-ID dataset. The experimental results will be better with a larger dataset.