Skip to content

Latest commit

 

History

History
321 lines (222 loc) · 9.59 KB

NOTES.org

File metadata and controls

321 lines (222 loc) · 9.59 KB

Identifying Woodblocks from printed Bukan Books

This is kind of backwards…

Work Log [12/19]

CW45/2019 [3/3]

Manually going through Bukan Collection

Skim “Digital Image Processing” book

  • Continuous Image Characterization
    1. Continuous Image Mathematical Characterization
    2. Psychophysical Vision Properties
    3. Photometry and Colorimetry

Prepare short presentation about topic

CW46/2019 [3/3]

Continue with “Digital Image Processing

  • Digital Image Characterization
  • Discrete Two-Dimensional Processing

Some quantitative Image Analysis

I want to get some categories for the image processing pipeline.

Look for related papers

  • ICDAR (International Conference on Document Analysis and Recognition)
  • CBDAR (International Workshop on Camera-Based Document Analysis)
  • Both seem to be useless at the moment

CW47/2019 [4/4]

Set up work PC

Create small subset for testing

Use the 袖珍武鑑 (shuuchinbukan.csv) books. These are 56 editions over 89 years.

Evaluation of pHashes

I thing I need a low-level approach here since my computer is so slow. So: C++/C/Rust (don’t know how its FFI works here) Result: Did not work

Continue with “Digital Image Processing”

  • Image Improvement

CW48/2019 [2/2]

Manually prepare annotations for 袖珍武鑑 test set

Finish “Digital Image Processing”

CW49/2019 [4/4]

Prepare for meeting with Prof Kitamoto

Start with “Computer Vision” book

  • [X] Introduction
  • [X] Image formation
  • [X] Image processing

Skim OpenCV Documentation and make notes

At least a few bullet points for each chapter of the official docs

Implement a first feature matching algorithm

Starting out with the OpenCV tutorials

  • [X] I used ORB because not patented and from OpenCV itself. Matching looks good.
  • [ ] I now like to have some metrics for comparing matching algorithms.
  • [ ] Furthermore, I like to proceed with feature based alignment. Maybe building a first prototype.

CW50/2019 [5/5]

Meeting with Prof Kitamoto

Continue with “Computer Vision” book

  • [X] Feature Detection and Matching
  • [X] Segmentation
  • [X] Feature-based alignment

Implement and describe a simple baseline

  • I this I’ll best use ORB at first

Finish “Computer Vision” book

  • [X] Skim rest of the book,
  • [X] Especially Image-based rendering (what is this?)

Experiment with different Feature Detectors

CW51/2019 [3/3]

Meeting with Prof Kitamoto

  • Prepare some slides
  • Ask how to best proceed

Start with writing a first draft of research results

Use some more matchers

  • [X] AKAZE with rotational invariance
  • [X] SIFT
  • [X] SURF

CW52/2019 [2/2]

Pipeline Optimizations

Run on more test data data

Filter out one- and two-paged images

Calculate page averages

Just for fun. Doesn’t seem to be meaningful.

HOLIDAY CW01/2020

CW02/2020 [3/3]

Meeting with the Prof

Preparing two presentations:

  • [X] A general introduction of the topic
  • [X] My current results

Examining pipeline

Which step produces which effect?

Examining some particular images

It’s not only about the numbers. I need to see which images succeeded and failed.

CW03/2020 [3/3]

Cleaning up code

How to build a system out of this?

Run on full dataset

CW04/2020 [4/4]

Major Bukan Version recognition possible?

Visualizing changes

Meeting with Prof

Writing report

CW05/2020 [1/1]

Implemented webapp prototype

CW06/2020 [3/3]

Recalculated features with split up pages

Finalized webapp

Meeting at Institute for Japanese Literature

CW07/2020 [5/5]

Setup Django project

Create ORM model

Database migrations

Look into WebGL for image processing

Work in your report!

CW08/2020 [0/0]

CW09/2020 [0/0]

CW10/2020 [0/0]

CW11/2020 [0/0]

CW12/2020 [0/0]

Overview

We have 366 scanned books with around 90,000 pages. Now we want to find some links for better understanding the data.

We have no ground truth!

So first, let’s apply some techniques from classical image processing.

Since this worked out better than expected, let’s see how one might use this for building a Bukan comparison platform out of this.

Problems

Easy?

Seem to be solved with standard tools; just need to find the right parameters.

  • Page detection

Medium?

There are some current papers on this; harder than it seems but there are some working approaches.

  • Page binarization
  • Visualization of image changes. This shouldn’t be too sensitive to pixel changes. If thresholding and opening/closing doesn’t work I’m out of ideas. Maybe it is also possible to just paint a rectangle around some cluster of matches…

Hard?

There are no (useful) existing approaches and therefore no existing tools. But it seems this isn’t the problem here. The task is too easy. ;)

Various Open Questions

  • [X] Is there a difference between simple 武鑑 and 武鑑大全? Not sure, maybe just a different edition.

Historical and Cultural Background

Visit woodblock printing museums [0/3]

Ota Memorial Museum of Art

Printing Museum

Sumida Hokusai Museum

Reading some Books

  • [X] The Elements of Japanese Design
  • [ ] A Modern History of Japan

Working with the Data itself

Manually examine the collection [100%]

Usable in general? [352/366]

Automatic filtering the books by quantitative measures

  • Do we have enough books from the same location?
  • Does the number of pages match?

Technical Stuff

Preprocessing

Convert to Greyscale

Do this in memory

Convert to binary (Black/White)

You might want to use Histograms for finding good thresholds “Document Image Binarization” … Adaptive Thresholding

Finding Major Differences

With perceptual hashes using pHash Result: Did not work!

Finding Minor Differences

Aligning/Registering the images and doing pixelwise comparison