-
Notifications
You must be signed in to change notification settings - Fork 0
Home
A senior capstone course project for UK Computer Science by **Alan and His Merry Men **
- Nolan Chancellor
- Siyuan Chen
- Michael Dingess
- Tim Watson
AKA CS499-001 Team 11
Archival of printed documents often requires retention of the digitized copy of the document. When there is a large quantity of documents, storage size becomes a critical concern.
Lossy compression offers a solution, but the quality of the reconstructed image may suffer. For example, JPEG compression artifacts such as ringing or blocking can distort text within the image, making it difficult to read.
The goal of this project is to restore and enhance the lossy document image to resemble the original, scanned document as closely as possible. The project will explore machine-learning approaches to transform the compressed image into a higher-quality image.
In addition to developing a single, multi-purpose model (i.e., for all types of images), the project may consider additional models customized for specific types of compression or image content.
In addition to subjective visual evaluation (i.e., looking at the image), the project should consider one or more objective quality criteria. Examples include fidelity metrics (such as SNR and SSIM) and functional metrics (such as OCR accuracy).
The customer will provide full-resolution (ground-truth) scanned document images. Using these images, the project team will develop training data.
Some training samples should use JPEG compression across a range of quality factors. Other training samples should apply simple spatial and/or tonal resolution reduction (i.e., lower resolution and/or quantization of tone levels). The project team may wish to include other lossy compression algorithms in addition.
In all cases, test images need to include an appropriate range of compression.
The customer will evaluate and approve the proposed collection of training samples.
The deliverable product consists of functional machine-learning models that the customer can run and evaluate. The project team should clearly document and explain their approach so that the customer can readily duplicate their results.
~ Brian Cooper : Lexmark