Blur Detection

Blur Detection estimates sharpness by seeking contours in an image. It starts out with a very strict contrast threshold for contour growth while seeking a minimum number of contours. If it can't find that minimum, it relaxes the contrast threshold and tries again. Sharper images will tend to have contours that are easier to find, while blurry images will require very low contrast thresholds or have no contours at all.

Setup

This package has a number of dependencies. These can be installed by running:

python -m venv .
Scripts/activate
pip install -U -r requirements.txt

This creates a virtual environment for the script to run in so that it won't conflict with other scripts.

Usage

The repository has a script, process.py which lets us run on single images or directories of images. The blur detection method is highly dependent on the size of the image being processed. To get consistent scores we fix the image size to HD, to disable this use --variable-size. The script has options to,

# run on a single image
python process.py -i input_image.png

# run on a directory of images
python process.py -i input_directory/ 

# or both! 
python process.py -i input_directory/ other_directory/ input_image.png

In addition to logging whether an image is blurry or not, we can also,

# save this information to json
python process.py -i input_directory/ -s results.json

# display blur-map image
python process.py -i input_directory/ -d

# move images into subfolders based on their evaluation
python process.py -i input_directory -m

The saved json file has information on how sharp an image is, the higher the value, the sharper the image.

{
    "images": ["/Users/demo_user/Pictures/Flat/"],
    "fix_size": true,
    "results": [
        {
            "blurry": false,
            "input_path": "/Users/demo_user/Pictures/Flat/IMG_1666.JPG",
            "score": 6984.8082115095549
        },
    ],
    "threshold": 100.0
}

History / Changes from parent branch

This was forked from Will Brennan's BlurDetection2. The major changes from that version are:

• Switched to a completely different (mildly novel?) algorithm for estimating blur.
• Support for raw files via RawPy
  • Note: this has only been tested with NEF and CR2 files!
• The original script processed every file twice in Windows environments due to a difference in how globbing works between Linux and Windows. I'm a Windows-focused developer, and my "bugfix" for this effectively moves the bug to Linux.
  • If running this script on Linux, note that it will only seek lowercase filenames by default.
• An additional flag that causes images to be moved into subfolders based on their score (blurry/semi_blurry)
  • Note: This hasn't been tested extensively and may interact poorly with having multiple input paths.
• A summary section after everything has been processed that shows statistics for the run.
• Bugfixes
  • A fix for the scaling system to increase the range of possible input image sizes
  • A fix to file detection that prevents files from being processed twice

Citations

This is based upon the blog post Blur Detection With Opencv by Adrian Rosebrock.