Vision-Language Integration: Transformer-Based Generative Ai For Image Captioning To Improve Accessibility And Visual Content Understanding

Image captioning has become a problematic activity in computer vision and natural language processing, filling in the gap between visual cognition and language description. The current captioning systems, however, amusingly do not have the contextual flexibility, integration of accessibility, and multi-modality output required in the real world. The presented work suggests a generative framework based on transformers that combines vision language modeling and accessibility-driven improvements. Using the model of BLIP (Bootstrapping Language Image Pretraining), we apply both conditional and unconditional caption generation to various situations of images. The system supports the use of Google Text-to-Speech to provide multi-language audio output, adjustable voice settings, and capture history, to enhance accessibility, most especially to the visually impaired users. It is built with PyTorch and deployed with Streamlit, and hence offers an interactive and user-friendly interface. The system is shown to be able to provide correct and contextually detailed captions on heterogeneous sets of images through experimental assessment, and to provide features of accessibility that allow inclusivity in promoting user engagement. The study will help to promote the development of multimodal AI systems because it will allow both content creators, accessibility advocates, and ordinary users in comprehending images accurately, easily, and individually.

Keywords: Image captioning, Image-text model, Vision Language model, Transformers, text-to-speech (TTS), accessibility technology, PyTorch, streamlit, generative AI, multi-modal system

Bhuvaneswari S, Kavya S, Chethana B, Akshitha M, Egneswari K (2026). Vision-Language Integration: Transformer-Based Generative Ai For Image Captioning To Improve Accessibility And Visual Content Understanding. International Journal of Innovative Research in Multidisciplinary Education & Technology (IJIRMET), 11(3).