Enhancing THAI Image Captioning Performance using CNN and Bidirectional LSTM

This research has designed a deep learning model to create Thai captions using a convolutional neural network (CNN) in VGG16 format to extract image features, and it is used to procreate captions using bidirectional LSTM. The data warehouse used for training and testing is Flickr8k, which combines customized traffic-related image and caption information. For the first set of data, that is Flickr8k, all subtitles had to be translated from English to Thai using Google Translate, and ways to deal with the data before training were to remove special characters to prevent the Thai language description from being distorted. Then, to evaluate the result of the captions the model produced compared to default captions, the BLEU metric was used to measure the score. The resulting average score was effective because it was higher than the compared models. The score values were paralleled up to 4 grams.