This is the Undergraduate Thesis project (CS4490) that I completed in my final year of University.

This thesis is a feasibility study examining the automated segmentation of scanned 3D images for extraction of non-planar pages. Positive study findings indicate that this is indeed possible, however, more work is required to advance this area of study. Most previous researchers have used replicas of historical books/texts in their studies, while this project focuses on using a real historical manuscript to determine findings.

To study the feasibility of using a micro-CT scanner with a deep learning algorithm to splice images into individual legible pages from a closed book, scans were first obtained of the book Book of hours: Canon Grandel's Prayer Book. The scans were taken using a Nikon XT 225 ST industrial micro-CT scanner. To maximize image resolution the book was scanned in two sections. The larger of these two sections was the dataset used for the entire feasibility study.

The scans were then loaded into the software Dragonfly 4.1 by Object Research Systems (ORS).
A subset of the slices was separated from the book scan and then segmented manually. To do the segmentation each slice in the subset was examined and six pages were highlighted. Each page was highlighted a different colour. This took about 45 minutes per slice. The subset contained one slice for every 100 slices in the original scan of the book for a total of 17 slices in my subset.

Different deep learning algorithm were tested, U-Net, PsyNet, and FC-DenseNet. PsyNet and FC-DenseNet took over four times to train and PsyNet crashed Dragonfly when used to segment. Ultimately U-Net was selected as it trained the fastest and was able to produce accurate segmentation. Using the subset of slices, segmented manually, the U-Net deep learning algorithm was trained.

Once the deep learning algorithm had its neural network built, the algorithm was used to segment the entire scanned volume of slices, over 1,700 slices. The results from this partial book were evaluated to determine the feasibility of automated or semi-automated procedures to segment a scanned book to extract multiple pages at once.

The image used on this page is of one of pages I managed to extract. One can clearly see some letters on the page.

For additional information and images please see the entire thesis report which I have linked below.