Improvisation of Dataset Efficiency in Visual Question Answering Domain

The technology revolution moves the world towards automation and most of the activities are performed with minimum human intervention. The medical domain is not an exception, few developments in the medical domain helps both the patient and physician to some extent. As a part of this advancement, Visual Question Answering (VQA) in the medical domain is evolved and which helps the physician and partially visually sighted people in clinical decision making and patient education. One of the main disadvantages in achieving this advancement is data limitation problem. In this paper, two methods for handling the data limitation problem are explained and validated using appropriate pretrained models like VGGNet and ResNet. The methods namely label smoothing and mixup are used to reduce the hard samples and augmentation of the medical data. From the performance analysis, it has been inferred that the highest accuracy and BLEU score are obtained for improved dataset as 0.297 and 0.313 for ResNet with a significant improvement of 7.9% and 5.9% respectively.