<![CDATA[Vision-language models could be very useful for assistive technology, but they struggle significantly with processing low-quality, noisy photos that visually impaired individuals often take. The primary contribution of this research is a data-centric pipeline that combines Fast Fourier Transform (FFT) blur filtering with Parameter-Efficient Fine-Tuning (PEFT) utilizing Low-Rank Adaptation (LoRA) to directly confront and eradicate the “Quality Paradox” bias in accessibility datasets like VizWiz. We use the AdamW optimizer for memory-efficient training and perform comprehensive statistical ablation studies with FFT-based tertile splits to see how image quality affects the results. The optimized model achieves a CIDEr score of 0.4481, a strong BERTScore F1 of 0.8815, and a BLEU-4 of 0.0511. Statistical analysis (p = 0.622, Cohen’s d = -0.038) shows that this bias has been successfully removed, which means that performance is balanced on both severely blurry and crisp images. Further qualitative testing shows that the model can recognize text and objects even in very low light without needing an extra OCR module, demonstrating its viability as a stand-alone navigational aid.]]>
Copyrights © 2026
