Automated Alt-Text Generation for EPUB Accessibility using Vision-Language Models
An end-to-end pipeline that generates descriptive alt text for images in EPUB e-books: it parses the EPUB and locates images, extracts visual features with CLIP and ViT, fuses them with the surrounding document text, and feeds the result to a fine-tuned transformer language model that writes the alt text, then rewrites the EPUB metadata for accessibility compliance. Students learn how vision-language models are chained with an LLM and how the output is validated with BLEU, cosine similarity, and user studies.
2501.00113
Alt text is the textual description attached to an image that lets a screen reader convey the image to a visually impaired reader; many EPUB e-books ship without it. AltGen is an end-to-end pipeline …