Finding Duplicate Images

Finds equal or similar images in a directory containing (many) image files.

Official home page: https://github.com/lene/DuplicateImages

Development page: https://gitlab.com/lilacashes/DuplicateImages

PyPI page: https://pypi.org/project/duplicate-images

Usage

Installing:

$ pip install duplicate_images

Printing the help screen:

$ find-dups -h

Quick test run:

$ find-dups $IMAGE_ROOT 

Typical usage:

$ find-dups $IMAGE_ROOT --parallel --progress --hash-db hashes.json

Supported image formats

• JPEG and PNG (tested quite thoroughly)
• HEIC (experimental support, tested cursorily only)
• All other formats supported by the pillow Python Imaging Library should work, but are not specifically tested.

Explicitly allow huge images

The PIL image library, which is used as backend, limits the size of images to 178956970 pixels by default, to guard against memory exhaustion. For larger images, specify the maximum image size in pixels with the --max-image-pixels option.

Image comparison algorithms

Use the --algorithm option to select how equal images are found. The default algorithm is phash.

ahash, colorhash, dhash, dhash_vertical, phash, phash_simple, whash: seven different image hashing algorithms. See https://pypi.org/project/ImageHash for an introduction on image hashing and https://tech.okcupid.com/evaluating-perceptual-image-hashes-okcupid for some gory details which image hashing algorithm performs best in which situation. For a start I recommend using phash, and only evaluating the other algorithms if phash does not perform satisfactorily in your use case.

Image similarity threshold configuration

Use the --hash-size parameter to tune the precision of the hashing algorithms. For the colorhash algorithm the hash size is interpreted as the number of bin bits and defaults to 3. For all other algorithms the hash size defaults to 8. For whash it must be a power of 2.

Use the --max-distance parameter to tune how close images should be to be considered duplicates. The argument is a positive integer. Its value is highly dependent on the algorithm used and the nature of the images compared, so the best value for your use case can oly be found through experimentation.

NOTE: using the --max-distance parameter slows down the comparison considerably with large image collections, making the runtime complexity go from O(N) to O(N²). If you want to scan collections with at least thousands of images, it is highly recommended to tune the desired similarity threshold with the --hash-size parameter alone, if that is at all possible.

Pre-storing and using image hashes to speed up computation

Use the --hash-db ${FILE}.json or --hash-db ${FILE}.pickle option to store image hashes in the file $FILE in JSON or Pickle format and read image hashes from that file if they are already present there. This avoids having to compute the image hashes anew at every run and can significantly speed up run times.

Handling matching images either as pairs or as groups

By default, matching images are presented as pairs. With the --group CLI option, they are handled as a group containing all matching images.

Example: 1.jpg, 2.jpg and 3.jpg in the current folder . are equal.

$ find-dups .
1.jpg 2.jpg
1.jpg 3.jpg
2.jpg 3.jpg
$ find-dups . --group
1.jpg 2.jpg 3.jpg

Actions for matching image groups

Use the --on-equal option to select what to do to pairs of equal images. The default action is print.

• delete-first or d1: deletes the first of the files in the group
• delete-last or dl: deletes the last of the files in the group
• delete-biggest or d>: deletes the file with the biggest size
• delete-smallest or d<: deletes the file with the smallest size
• symlink-smaller: delete the smaller files and replace them to a symlink to the biggest file
• eog: launches the eog image viewer to compare the files in the group (deprecated by exec)
• xv: launches the xv image viewer to compare the files in the group (deprecated by exec)
• print: prints the files in the group
• print_inline: like print but without newline
• quote: prints the files in the group quoted for POSIX shells
• quote_inline: like quote but without newline
• exec: executes a command (see --exec argument below)
• none: does nothing; may be useful for benchmarking and testing

The --exec argument allows calling another program when the --on-equal exec option is given. You can pass a command line string like --exec "program {1} {2}" where {1} and {2} are replaced by the matching pair files (or first two files in a group), quoted so the shell recognizes the files properly. The wildcard {*} expands to all files in a matching group, which when called with the --group argument may be more than two images considered equal.

Examples:

• --exec "open -a Preview -W {1} {2}": Opens the files in MacOS Preview app and waits for it.
• --exec "ls -s {*}": Prints the size (in blocks) next to all files.
• --exec 'for i in {*}; do dirname $i; basename $i; done': Shows the directory and the filename separately for all files.

Parallel execution

Use the --parallel option to utilize all free cores on your system for calculating image hashes. Optionally, you can specify the number of processes to use with --parallel $N.

To execute the --on-equal actions in parallel, use the --parallel-actions option, which also can take an optional number of processes to use as argument.

Excluding subfolders

Use the --exclude-dir option to exclude subfolders of $IMAGE_ROOT from the search. The argument is a regular expression matching the subfolder names to be excluded. Multiple arguments can be passed to --exclude-dir to exclude multiple subfolders.

The argument(s) given to --exclude-dir may be regular expressions. These regular expressions are matched only against the directory name, not the file name.

Examples

Exclude subfolder $IMAGE_ROOT/foo:

$ find-dups $IMAGE_ROOT --exclude-dir $IMAGE_ROOT/foo

Exclude all subfolders named foo or bar:

$ find-dups $IMAGE_ROOT --exclude-dir foo bar

Slow execution

find-dups can also use an alternative algorithm which exhaustively compares all images to each other, being O(N²) in the number of images. This algorithm is selected automatically if --max-distance is not 0.

You can use the --slow option to use this alternative algorithm specifically. The --slow switch is mutually exclusive with the --group switch.

Progress bar and verbosity control

• --progress prints a progress bar each for the process of reading the images, and the process of finding duplicates among the scanned image
• --debug prints debugging output
• --quiet decreases the log level by 1 for each time it is called; --debug and --quiet cancel each other out

Development notes

Needs Python3, Pillow imaging library and pillow-heif HEIF plugin to run, additionally Wand for the test suite.

Uses Poetry for dependency management.

Installation

From source:

$ git clone https://gitlab.com/lilacashes/DuplicateImages.git
$ cd DuplicateImages
$ pip3 install poetry
$ poetry install

Running

$ poetry run find-dups $PICTURE_DIR

or

$ poetry run find-dups -h

for a list of all possible options.

Test suite

Running it all:

$ poetry run pytest
$ poetry run mypy duplicate_images tests
$ poetry run flake8
$ poetry run pylint duplicate_images tests
$ poetry run bandit -r duplicate_images

or simply

$ .git_hooks/pre-push

Setting the test suite to be run before every push:

$ cd .git/hooks
$ ln -s ../../.git_hooks/pre-push .

Publishing

A tag is created and the new version is published automatically by GitLab CI on every successful merge to master.

Prerequisites

For every Merge Request to master it is checked that:

• the version number in pyproject.toml is not an already existing git tag
• the CHANGELOG.md contains an entry for the current version number

PyPI

There is a job in GitLab CI for publishing to pypi.org that runs as soon as a new tag is added, which happens automatically whenever a MR is merged. The tag is the same as the version in the pyproject.toml file. For every MR it needs to be ensured that the version is not the same as an already existing tag.

To publish the package on PyPI manually:

$ poetry config repositories.testpypi https://test.pypi.org/legacy/
$ poetry build
$ poetry publish --username $PYPI_USER --password $PYPI_PASSWORD --repository testpypi && \
  poetry publish --username $PYPI_USER --password $PYPI_PASSWORD

(obviously assuming here that username and password are the same on PyPI and TestPyPI)

Updating GitHub mirror

The GitHub repo git@github.com:lene/DuplicateImages.git is set up as a push mirror in GitLab CI, but mirroring is flaky at the time and may or may not succeed. The CI job PushToGithub should take care of mirroring to GitHub after every merge to master.

To push to the GitHub repository manually (assuming the GitHub repo is set up as remote github):

$ git checkout master
$ git fetch
$ git pull --rebase
$ git tag  # to check that the latest tag is present
$ git push --tags github master 

Creating Releases on GitHub

The CI job CreateGithubRelease creates a Release on GitHub, which can then be found under https://github.com/lene/DuplicateImages/releases.

Profiling

CPU time

To show the top functions by time spent, including called functions:

$ poetry run python -m cProfile -s tottime ./duplicate_images/duplicate.py \ 
    --algorithm $ALGORITHM --action-equal none $IMAGE_DIR 2>&1 | head -n 15

or, to show the top functions by time spent in the function alone:

$ poetry run python -m cProfile -s cumtime ./duplicate_images/duplicate.py \ 
    --algorithm $ALGORITHM --action-equal none $IMAGE_DIR 2>&1 | head -n 15

Memory usage

$ poetry run fil-profile run ./duplicate_images/duplicate.py \
    --algorithm $ALGORITHM --action-equal none $IMAGE_DIR 2>&1

This will open a browser window showing the functions using the most memory (see https://pypi.org/project/filprofiler for more details).

Contributors

• Lene Preuss (https://github.com/lene): primary developer
• Mike Reiche (https://github.com/mreiche): support for arbitrary actions, speedups
• https://github.com/beijingjazzpanda: bug fix