UnsharpDetector

AI-application to automatically identify and delete blurry photos.

Installation

Load the 64 bit version of Python from python.org and install it. You can skip this step on Linux because Python is usually included in your distribution.

Get the code using the download link or clone this repository with git. If you are using Windows or macOS get git from git-scm.com. Use your package manager on Linux.

The next step is to create a virtual environment. This makes sure that the following steps can not interfere with other python programs. Create the virtualenv on macOS and Linux with the following command (in the directory where you downloaded the code):

python3 -m venv env

Windows does not find python.exe by default. So you may have to specify the full path:

..\..\AppData\Local\Programs\Python\Python36\python.exe -m venv env

Start the virtualenv on Windows with env\Scripts\activate.bat. On Linux and macOS use source env/bin/activate.

Install all other dependencies with pip:

pip install -r requirements.txt

If you are using a Nvidia GPU and CUDA you may use requirements-gpu.txt. TensorFlow will use a version which uses your GPU to run the neural network.

Usage

To use the program activate the virtualenv first with env\Scripts\activate.bat (Windows) or source env/bin/activate (Linux and macOS).

Run the graphical Application with:

python inference_gui.py

The program starts after a couple of seconds (initialization of TensorFlow). Initially it displays an empty list. You fill the list by clicking the Button in the upper left corner and selecting a path. The software will load all the images in this folder which may take a couple of seconds depending on the number and the size of the images. The classification starts immediately in the background.

You may immediately mark images for keeping or removal using the mouse. The neural network will analyze all the images you do not classify manually. The dashed line around these images indicates the decision of the network. Green means definitely sharp. Red means blurry. Brown is something in between and a good candidate to override the decision.

If the thumbnail is too small to decide if an image is sharp enough to keep you may click on the thumbnail. This opens the image in full resolution in the preview area on the right.

If you are happy with all the decisions for the images click on the red button in the upper right corner. This deletes all the images which were marked for removal (red border) without further questions.

Training

This repository comes with weights and settings for a pretrained neural network. If you want to experiment with different network architectures you can simply change the config and run train.py.

The code uses sacred to keep track of the experiments. To use this magic create a file named secret_settings.py which defines two variables:

1. mongo_url: The url of your mongodb with credentials.
2. db_name: The database name you are using in the mongodb.

If you are training on a dedicated server you can create queued experiments with -q on your notebook and start queue_manager.py on the server. It will automatically fetch queued experiments from your database and run them.

Most network architectures will learn some specifics of the generated datasets after 2-5 epochs. Training for 50 epochs (my default setting) leads to something which looks like overfitting. So if you want the best accuracy on validation data you may want to train for only 2-5 Epochs. But this also depends on the size of your dataset.

Also make sure you have no blurry images in your training dataset. This greatly reduces the accuracy. My intention was to use images from vacations where I had already manually deleted all blurry images. I trained with ca. 2000 images.

Instabilities

I stumbled upon some instabilities of this program. Sometimes it crashes with a not very informative segmentation fault. This has something to do with the C-code from Qt or TensorFlow. It happened randomly. If you run into this problem try doing the same thing again. It may just work at the second attempt. If you have any idea what triggers these crashes please create an issue with your idea.