Open Source Testing Framework for image correlation, distance and analysis. Strongly related to : Douglas-Quaid
Building a generic library and services which can establish correlation between pictures. In order to achieve this goals, experiments needs to be conducted. This is the goal of this repository.
- Review of existing algorithms, techniques and libraries for calculating distances between images, State Of The Art : MarkDown | PDF version
See requirements.txt
(...)
(...)
(...)
in /lib_testing you just have to launch "python3 ./launcher.py" Parameters are hardcoded in the launcher.py, as :
- Path to pictures folder
- Output folder to store results
- Requested outputs (result graphe, statistics, LaTeX export, threshold evaluation, similarity matrix ...)
This is currently working on most configuration and will explore following algorithms for matching :
- ImageHash Algorithms (A-hash, P-hash, D-hash, W-hash ... )
- TLSH (be sure to have BMP pictures or uncompressed format at least. A function is available to convert pictures in /utility/manual.py)
- ORB (and its parameters space)
- ORB Bag-Of-Words / Bag-Of-Features (and its parameters space, including size of the "Bag"/Dictionnary)
- ORB RANSAC (with/without homography matrix filtering)
You can also manually generate modified datasets from your original dataset :
- Text detector and hider (DeepLearning, Tesseract, ...)
- Edge detector (DeepLearning, Canny, ...)
- PNG/BMP versions of pictures (compressed/uncompressed)
(...)
(...)
For the algorithms test library : See installation instruction
- Original project structure source
- Clean library implementation of algorithms
- Followed practice for logging
- Text detector model source
- Built-for-the-occasion manual image classificator
- Bibliography
PR are welcomed. New issues are welcomed if you have ideas or usecase that could improve the project.