py-image-dedup
py-image-dedup is a tool to sort out or remove duplicates within a photo library. Unlike most other solutions, py-image-dedup intentionally uses an approximate image comparison to also detect duplicates of images that slightly differ in resolution, color or other minor details.
It is build upon Image-Match a very popular library to compute a pHash for an image and store the result in an ElasticSearch backend for very high scalability.
How it works
Phase 1 - Database cleanup
In the first phase the elasticsearch backend is checked against the
current filesystem state, cleaning up database entries of files that
no longer exist. This will speed up queries made later on.
Phase 2 - Counting files
Although not necessary for the deduplication process it is very convenient
to have some kind of progress indication while the deduplication process
is at work. To be able to provide that, available files must be counted beforehand.
Phase 3 - Analysing files
In this phase every image file is analysed. This means generating a signature (pHash)
to quickly compare it to other images and adding other metadata of the image
to the elasticsearch backend that is used in the next phase.
This phase is quite CPU intensive and the first run take take quite
some time. Using as much threads as feasible (using the -t parameter)
is advised to get the best performance.
Since we might already have a previous version of this file in the database
before analysing a given file the file modification time is compared to the
given one. If the database content seems to be still correct the signature
for this file will not be recalculated. Because of this, subsequent
runs will be much faster. There still has to happen some file access though,
so it is probably limited by that.
Phase 4 - Finding duplicates
Every file is now processed again - but only by means of querying the
database backend for similar images (within the given max_dist).
If there are images found that match the similarity criteria they are considered
duplicate candidates. All candidates are then ordered by the following
criteria (in this exact order):
- pixel count (more is better)
- EXIF data (more exif data is better)
- file size (bigger is better)
- file modification time (newer is better)
- distance (lower is better)
- filename contains "copy" (False is better)
- filename length (longer is better) - (for "edited" versions)
- parent folder path length (shorter is better)
- score (higher is better)
The first candidate in the resulting list is considered to be the best
available version of all candidates.
Phase 5 - Moving/Deleting duplicates
All but the best version of duplicate candidates identified in the previous
phase are now deleted from the file system (if you didn't specify --dry-run of course).
If duplicates_target_directory is set, the specified folder will be used as
a root directory to move duplicates to, instead of deleting them, replicating their original
folder structure.
Phase 6 - Removing empty folders (Optional)
In the last phase, folders that are empty due to the deduplication
process are deleted, cleaning up the directory structure (if turned on in configuration).
How to use
Install
Install py-image-dedup using pip:
pip3 install py-image-dedup
Configuration
py-image-dedup uses container-app-conf
to provide configuration via a YAML file as well as ENV variables which
generates a reference config on startup. Have a look at the
documentation about it
See py_image_dedup_reference.yaml
for an example in this repo.
Setup elasticsearch backend
Since this library is based on Image-Match
you need a running elasticsearch instance for efficient storing and
querying of image signatures.
Elasticsearch version
This library requires elasticsearch version 5 or later. Sadly the
Image-Match library still
specifies version 2, so a fork of the original project
is used instead. This fork is maintained by me, and any contributions
are very much appreciated.
Set up the index
py-image-dedup uses a single index (called images by default).
When configured, this index will be created automatically for you.
Command line usage
py-image-dedup can be used from the command line like this:
py-image-dedup deduplicate --help
Have a look at the help output to see how you can customize it.
Daemon
CAUTION! This feature is still very much a work in progress.
Always have a backup of your data!
py-image-dedup has a built in daemon that allows you to continuously
monitor your source directories and deduplicate them on the fly.
When running the daemon (and enabled in configuration) a prometheus reporter
is used to allow you to gather some statistical insights.
py-image-dedup daemon
Dry run
To analyze images and get an overview of what images would be deleted
be sure to make a dry run first.
py-image-dedup deduplicate --dry-run
FreeBSD
If you want to run this on a FreeBSD host make sure you have an up
to date release that is able to install ports.
Since Image-Match does a lot of
math it relies on numpy and scipy. To get those working on FreeBSD
you have to install them as a port:
pkg install pkgconf
pkg install py38-numpy
pkg install py27-scipy
For .png support you also need to install
pkg install png
I still ran into issues after installing all these and just threw those
two in the mix and it finally worked:
pkg install freetype
pkg install py27-matplotlib # this has a LOT of dependencies
Encoding issues
When using the python library click on FreeBSD you might run into
encoding issues. To mitigate this change your locale from ANSII to UTF-8
if possible.
This can be achieved f.ex. by creating a file ~/.login_conf with the following content:
me:\
:charset=ISO-8859-1:\
:lang=de_DE.UTF-8:
Docker
To run py-image-dedup using docker you can use the markusressel/py-image-dedup
image from DockerHub:
sudo docker run -t \
-p 8000:8000 \
-v /where/the/original/photolibrary/is/located:/data/in \
-v /where/duplicates/should/be/moved/to:/data/out \
-e PY_IMAGE_DEDUP_DRY_RUN=False \
-e PY_IMAGE_DEDUP_ANALYSIS_SOURCE_DIRECTORIES=/data/in/ \
-e PY_IMAGE_DEDUP_ANALYSIS_RECURSIVE=True \
-e PY_IMAGE_DEDUP_ANALYSIS_ACROSS_DIRS=True \
-e PY_IMAGE_DEDUP_ANALYSIS_FILE_EXTENSIONS=.png,.jpg,.jpeg \
-e PY_IMAGE_DEDUP_ANALYSIS_THREADS=8 \
-e PY_IMAGE_DEDUP_ANALYSIS_USE_EXIF_DATA=True \
-e PY_IMAGE_DEDUP_DEDUPLICATION_DUPLICATES_TARGET_DIRECTORY=/data/out/ \
-e PY_IMAGE_DEDUP_ELASTICSEARCH_AUTO_CREATE_INDEX=True \
-e PY_IMAGE_DEDUP_ELASTICSEARCH_HOST=elasticsearch \
-e PY_IMAGE_DEDUP_ELASTICSEARCH_PORT=9200 \
-e PY_IMAGE_DEDUP_ELASTICSEARCH_INDEX=images \
-e PY_IMAGE_DEDUP_ELASTICSEARCH_AUTO_CREATE_INDEX=True \
-e PY_IMAGE_DEDUP_ELASTICSEARCH_MAX_DISTANCE=0.1 \
-e PY_IMAGE_DEDUP_REMOVE_EMPTY_FOLDERS=False \
-e PY_IMAGE_DEDUP_STATS_ENABLED=True \
-e PY_IMAGE_DEDUP_STATS_PORT=8000 \
markusressel/py-image-dedup:latest
Since an elasticsearch instance is required too, you can
also use the docker-compose.yml file included in this repo which will
set up a single-node elasticsearch cluster too:
sudo docker-compose up
UID and GID
To run py-image-dedup inside the container using a specific user id
and group id you can use the env variables PUID=1000 and PGID=1000.
