부산대학교 도서관

Video decaptioning is the process of automatically removing subtitles from video frames and inpainting the captioned regions. However, directly transferring the deep-learning-based image inpainting methods to video decaptioning scenarios always requires the masks of subtitled regions, which is unavailable for subtitled video frames. To address these issues, we propose a two-stage lightweight framework. The first caption mask prediction stage uses an encoder-decoder full convolutional network with residual blocks to predict the caption mask. The second background inpainting stage uses an encoder-decoder structure with attention modules and the skip connection to repair the background areas. Extensive experiments demonstrate that our proposed model can produce better visual results and outperform state-of-the-art methods.