MESHED-MEMORY TRANSFORMER FOR IMAGE CAPTIONING IN VIETNAMESE

Abstract

Image captioning has been a topic of interest due to its various implementations such as impaired persons support, recommendations system, virtual assistants, image indexing, social media, and many other applications. Several studies using different architectures, such as LSTM, CNN, or RNN, have yielded promising results and have steadily improved over time. This thesis uses the mesh-memory Transformer approach to infer description sentences for images. The experimental results show that the meshmemory Transformer model effectively generates captions for images in Vietnamese. In particular, the BLEU-1, 2, 3, and 4 scores of the model on the Flickr8k dataset Vietnamese are 0.703, 0.589, 0.489, and 0.397, respectively.