Visual Question Answering Apparatus Using Selective Residual Learning and Method Thereof

摘要

The present embodiments selectively substitute zero vectors for some vectors that do not correspond to an important region in an attention-based correct answer inference model including a self-attention unit, a selective attention unit, and a guide attention unit, and correspond to an important region in the attention map Provided are a visual question answering apparatus and method in which the accuracy of correct answers to questions is improved by selectively masking some columns that are not.