Image restoration
DeblurGAN helps prepare photographs affected by blur for subsequent recognition.
03 / Computer Vision · Accessible AI
From blurred
to understood.
A multimodal shopping-assistance prototype for people with low vision. Restoring an image is not an end in itself: the goal is to make everyday product information easier to understand and access.
BEng thesis / Ritsumeikan University / 2025
DeblurGAN · Vision-language models · Speech interaction
Real images from the project repository
01 / Image restoration
See the difference.
Drag directly on the image (or use the slider and buttons) to compare image clarity. This is a visual comparison, not a numerical quality or recognition benchmark.
Visual comparison / retail photographsSET 01 / 02
BLURREDRESTOREDThe original research image could not load. You can still explore the architecture below or open the source repository.
Set 01: thesis repository DeblurGAN images. Set 02: original flower-shop photograph supplied by the author.
Set 01 shows paired images from the thesis repository. Set 02 uses your supplied flower-shop photograph and a deliberately blurred version for visual comparison; it is not a DeblurGAN output.
02 / System pipeline
See. Describe. Speak.
One of the thesis's important design questions was that a user with low vision might not know whether an image is sufficiently sharp to support visual understanding. The system therefore integrates image restoration before visual description and speech interaction.
Vision-language model
The enhanced image and question are supplied to the visual understanding system.
Speech interface
Audio output and spoken follow-up enable an alternative to visual reading.
This is an explanatory flow based on the project architecture; there is no live inference or external user-data processing on this webpage.
Code & thesis ↗03 / Contribution & evidence
From an assumption
to an experiment.
Individual bachelor’s thesis, Ritsumeikan University, January 2025. I framed the shopping-information problem, prepared paired image data, integrated the restoration, vision-language and speech components, and evaluated image quality and description overlap.
The initial concept assumed that users could provide a usable photograph. I recognised that I could judge blur and retake a photo because I could see it, while the intended users might not be able to do so. This motivated investigating restoration before visual interpretation.
| Measure | Blurred input | Restored input |
|---|---|---|
| SSIM ↑ | 0.45 | 0.87 |
| PSNR, dB ↑ | 18.2 | 26.5 |
| BLEU ↑ | 0.75 | 0.90 |
SSIM and PSNR measure image similarity. BLEU measures overlap with reference descriptions, not factual correctness or usability. The study evaluates restoration and generated descriptions under controlled conditions; it does not establish benefits in real shopping with low-vision users.
Evaluation details and limitations ↗