Vehicle Feature VQA: Visual Question Answering for Vehicle Feature

Abstract

Visual Question Answering (VQA) can automatically produce the predict answers for questions and real-world images. In this paper, we propose the VQA dataset for Vehicle Feature to know the knowledge of Vehicle. We develop the VQA model using RestNet50 in Convolutional Neural Networks (CNN) for feature extraction of images and Long Short-Term Memory (LSTM) for question feature extraction and answer generation. The experimental result describes the training loss, evaluation loss, Blue Score, and VQA accuracy for epochs 20 and epochs 30. In epochs 20, after VQA model generated the training loss 1.1949 , evaluation loss 1.7953, Blue Score 0.6180, and VQA accuracy 0.0493, this model predicted the one correct answer for question and image. In epochs 30, the VQA model predicted the five correct answers in fifteen test data for vehicle feature questions and image according to generate the training loss 0.8780, evaluation loss 1.6634, Blue Score 0.6775 and VQA accuracy 0.0627.