Skip to content

Latest commit

 

History

History
64 lines (50 loc) · 2.04 KB

README.md

File metadata and controls

64 lines (50 loc) · 2.04 KB
Raw

About

This is a set of utilities to find perceptual similar images in a small to medium data sets. It's not usefull on data sets which contains more than approx. 1000000 images since method used becomes not very accurate.

Building

  1. Install vcpkg.
  2. cmake -B build -S . -DCMAKE_TOOLCHAIN_FILE=/path/to/vcpkg/scripts/buildsystems/vcpkg.cmake -DVCPKG_TARGET_TRIPLET=x64-linux-release -DCMAKE_BUILD_TYPE=Release
  3. cmake --build build
  4. Binaries will be located in the build directory.

Compilation was tested on Ubuntu 18.04+ with packages from the list below installed by stantard means (apt-get install etc.).

Requirements

All other dependencies will be managed by vcpkg & cmake.

Usage

Example dataset can be downloaded from here

  • You need to build perceptual hashes for the images.
$ ./imghash --data /tmp/imgdupl-dataset-example --result /tmp/hashes.txt
  • You need to export results into SQLite database.
$ ./export2db hashes /tmp/hashes.txt /tmp/imgdupl.db
  • You need to clusterize images.
$ ./clusterizer /tmp/imgdupl.db 32 2 >/tmp/clusters_32.txt
  • You need to export results of clusterization stage into SQLite database.
$ ./export2db clusters /tmp/clusters_32.txt /tmp/imgdupl.db clusters_32
  • Print clusters.
$ ./print-clusters --database /tmp/imgdupl.db --table clusters_32 --min-size 3

You may vary perceptual hashes matching threshold constant (second argument to the clusterizer utility) to improve quality.

If you want to view clusterization results more visually you can run viewer:

  1. python3 -m venv .venv
  2. source .venv/bin/activate
  3. pip install --upgrade pip
  4. pip install -r viewer/requirements.txt
  5. ./viewer/webstand.py 127.0.0.1:9090 /tmp/imgdupl.db clusters_32