Helping Visually Impaired People Take Better Quality Pictures.

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Perception Science

Background:

• Visually impaired individuals face challenges in capturing high-quality images due to technical distortions like blur, noise, and poor exposure.
• Existing image analysis tools often fail to address the unique quality issues present in visually impaired user-generated content (VI-UGC).

Purpose of the Study:

• To develop automated tools for assessing and providing feedback on the technical quality of images captured by visually impaired users.
• To create a large-scale dataset of VI-UGC for advancing research in image quality assessment for this population.

Main Methods:

• Construction of the LIVE-Meta VI-UGC Database, a unique dataset comprising 40K distorted images and 40K patches with 2.7M perceptual quality judgments and distortion labels.
• Development of an automatic picture quality and distortion predictor using a multi-task learning framework, trained on the created dataset.
• Creation of a prototype feedback system to guide users in mitigating image quality issues.

Main Results:

• The developed predictor achieves state-of-the-art performance on VI-UGC, significantly outperforming existing models.
• The LIVE-Meta VI-UGC Database provides a valuable resource for future research in perceptual image quality assessment.
• The prototype feedback system demonstrates the potential for guiding visually impaired users to capture higher-quality images.

Conclusions:

• Perception-based image analysis can empower visually impaired individuals to improve their photography and social media engagement.
• The developed dataset and models represent a significant advancement in addressing the technical quality challenges of VI-UGC.
• Future work will focus on addressing semantic quality issues in addition to technical quality.