A Comparative Analysis of Transformer-Based Models for Document Visual Question Answering

Abstract

Visual question answering (VQA) is one of the most exciting problems of computer vision and natural language processing tasks. It requires understanding and reasoning of the image to answer a human query. Text Visual Question Answering (Text-VQA) and Document Visual Question Answering (DocVQA) are the two sub problems of the VQA, which require extracting the text from the usual scene and document images. Since answering questions about documents requires an understanding of the layout and writing patterns, the models that perform well on the Text-VQA task perform poorly on the DocVQA task. As the transformer-based models achieve state-of-the-art results in deep learning fields, we train and fine-tune various transformer-based models (such as BERT, ALBERT, RoBERTa, ELECTRA, and Distil-BERT) to examine their validation accuracy. This paper provides a detailed analysis of various transformer models and compares their accuracies on the DocVQA task.