Summary
| Purpose | Matchbox: Duplicate detection tool for digital document collections. |
| Homepage |
https://github.com/openplanets/scape/tree/master/pc-qa-matchbox |
| Source Code Repository |
https://github.com/openplanets/scape/tree/master/pc-qa-matchbox |
| License |
Open source |
| Debian Package | http://packages.debian.org/stable/package |
Description
The Matchbox tool is responsible for finding duplicatre pairs in a collection of digital documents based on SIFT features and SSIM methods. Consequently the tool takes a collection path with associated parameters as input. Currently three scenarios are implemented. These are:
- Duplicate search in one turn (parameter ‘all’)
- Professional duplicate search (experienced user can execute particular step in ‘FindDuplicates’ workflow)
- Quick check if two documents are duplicates (based on previous BoW dictionary).
Further parameters that influence and adjust duplicate analysis are currently investigated.
Image processing method:
The image processing algorithm can be described in 4 steps:
1. Document feature extraction
- Interest point detection (applying Scale Invariant Feature Transform (SIFT) keypoint extraction)
- Derivation of local feature descriptors (invariant to geometrical or radiometrical distortions)
2. Learning visual dictionary
- Clustering method applied to all SIFT descriptors of all images using k-means algorithm
- Run over collection and collect local descriptors in a visual dictionary using Bag-Of-Words (BoW) algorithm
3. Create visual histogram for each image document
4. Detect similar images based on visual histogram and local descriptors. Evaluate similarity score – pair-wise comparison of corresponding keyword frequency histograms for all documents. Conduct structural similarity analysis applying Sturctural SIMilarity (SSIM) approach (1 means identical and 0 means very different)
- Rotate
- Scale
- Mask
- Overlaying
Usage:
FindDuplicates script can be invoked from command line. For standard usage two parameters are required: path to the collection documents and ‘all’.
scape/pc-qa-matchbox/Python# python2.7 FindDuplicates.py -h
usage: FindDuplicates.py [-h] [\--threads THREADS] [\--sdk SDK] [\--precluster PRECLUSTER] [\--clahe CLAHE] [\--config CONFIG] [\--featdir FEATDIR] [\--bowsize BOWSIZE] [\--csv] [-v] dir all,extract,compare,train,bowhist,clean
User Experiences
currently installed at Austrian National Library