부산대학교 도서관

GIF (Graphics Interchange Format) images have become ubiquitous in modern Internet lingo, but they are often inaccessible to people with vision impairments. This paper presents a project that aims to improve the accessibility of GIF images on web pages by auto-injecting commentaries into their alt text tag using machine learning and web browser extension techniques. This system uses a CNN-LSTM architecture to generate descriptive texts for GIF images, which outperforms the state-of-the-art Tumblr-GIF research with 7% improvement over BLEU accuracy metric, 6% improvement over ROUGE, and 10% improvement over METEOR. Usability studies performed with a focus group of individuals with and without vision disabilities further show that the system performs well for its targeted audience.