VI-OCR: "Visually Impaired" optical character recognition pipeline for text accessibility assessment

Area of Science:

• Computer Vision
• Ophthalmology
• Human-Computer Interaction

Background:

• Low vision significantly impairs daily activities, especially reading.
• Current product design often neglects the needs of low vision readers due to challenges in quantifying text accessibility.

Purpose of the Study:

• To introduce VI-OCR, a novel text accessibility assessment pipeline.
• To bridge the gap between computer vision and low vision research.
• To estimate text recognizability for low vision readers based on visual acuity and contrast sensitivity.

Main Methods:

• Developed VI-OCR using state-of-the-art Optical Character Recognition (OCR) models.
• Benchmarked OCR and vision-language models on letter acuity (ETDRS charts), word acuity (MNREAD charts), and scene text recognition.
• Compared model performance to normal vision participants under simulated visual deficits.

Main Results:

• Identified limitations in some models, including poor generalizability and difficulty with contrast reduction.
• Observed that some models overperformed human capabilities rather than mimicking them.
• Demonstrated that advanced models like Qwen2.5-VL and GPT show human-like performance.

Conclusions:

• VI-OCR is a feasible approach for assessing text accessibility for the visually impaired.
• Human-like performance in AI models supports the utility of VI-OCR for inclusive design.
• Further research can refine models to better mimic human visual perception for low vision readers.