Thai-IC: Thai Image Captioning based on CNN-RNN Architecture
Lawankorn Mookdarsanit, Pakpoom Mookdarsanit · 2020
The trend of news are represented in an image with a short description (or caption) and quickly shared on social media. Most short captions in many languages (e.g., English, Indonesian, Myanmar, Chinese, Arabic, etc.) are manually written by human. Instead of human labor, the visual objects within an image have enough information to autonomously generate the caption, called image captioning. Thai image captioning (Thai-IC) is such a new problem in Thai natural language processing (Thai-NLP) to make the model understand the image. This paper proposes an end-to-end deep learning model to generated Thai image caption. The model consists of encoding stage by convolutional neural network (CNN) and decoding stage by recurrent neural network (RNN). Visual geometry group in 16-layers (VGGNet-16) is used to extract visuals from an image as CNN-encoder. The visuals are used to generate Thai captions by Long-short-term memory (LSTM) as RNN-decoder. Thai captioning corpus is constructed by secondary and primary data that has 10,732 images. This Thai-IC is evaluated by Bilingual Evaluation Understudy (BLEU) on the 10-fold cross validation.