Vision-Language Assistive Navigation for Visually Impaired Using BLIP Fine-Tuning

Abstract

This paper presents a novel approach to assistive navigation for visually impaired individuals using vision-language models. We fine-tune the BLIP (Bootstrapping Language-Image Pre-training) model using a three-stage LoRA (Low-Rank Adaptation) strategy to generate contextual navigation descriptions from visual input. Our method achieves significant improvements in BLEU scores and inference speed compared to baseline approaches, demonstrating the effectiveness of parameter-efficient fine-tuning for assistive AI applications.

Keywords

Citation

@inproceedings{ponnada2025vision,
  title={Vision-Language Assistive Navigation for Visually Impaired Using BLIP Fine-Tuning},
  author={Ponnada, Charan Sai and Kumar, S. and Patel, R.},
  booktitle={Proceedings of ISAECT 2025},
  year={2025}
}