View Proposal

---

Proposer

Hind Zantout

Title

Evaluating the impact of real-world image quality on Lightweight Vision-Language Models for Visually Impaired users 

Goal

Description

The project investigates how defects in images taken by Visually Impaired (VI) individuals affect the reliability of lightweight VLMs in Visual Question Answering (VQA). VizWiz, a public dataset of images taken by VI users will be used. The dataset is distributed under the Creative Commons Attribution 4.0 International (CC by 4.0) licence. Images will be categorised by their natural defects such as blur, over/under-exposure and improper framing. Three lightweight VLMs suitable for assistive applications will be evaluated. Image quality will be analysed using quality labels and Laplacian variance, while the VLMs' reliability will be evaluated using VQA performance. The project may additionally investigate the impact of visual enhancement techniques on the VQA score and Laplacian variance. 

Resources

Background

Url

Difficulty Level

Moderate

Ethical Approval

None

Number Of Students

0

Supervisor

Hind Zantout

Keywords

Degrees