itsjavi / mediatidy-python

Image Classifier and Photo Library Organizer using Python and Tensorflow

tidymypics

Command-line app written in Python to classify your photo library by year, separating different image classes (of your choice) using a trained Tensorflow model.

This CLI application consists of 2 parts:

• A command to generate and train a new model
• A command to classify your photo library and organize it by class and year

Requirements

• UNIX-compatible OS (only tested on macOS 12)
• Python 3.8+ with PIP 22.0+
• Tensorflow 2.9+
  • For macOS, follow this guide

Dataset required to train the model:

• Images correctly classified manually under the data/train directory of this project. One folder for every different class, for example: data/train/photo and data/train/screenshot.
• At least 1000 images (samples) for every class class.
• The more and the more diverse, the better.
• All classes should contain the same number of images (to avoid bias), or aproximately.
• Avoid duplicated images. Try not to add images that are too similar (e.g. photo bursts).

Usage

Setup

• Clone or download this repository
• Open your terminal and change the directory to this one
• Run pip install -r requirements.txt to install the dependencies

Creating a new model

To create a new model with your own data (pictures), use the python -m tidymypics --train TRAIN_PATH TEST_PATH.

A new model will be generated using TRAIN_PATH as the path where your train samples are stored (by class like mentioned before).

It also requires a TEST_PATH argument, which should be a directory similarly structured as TRAIN_PATH, to test your model with new and "unseen" images, and see how good are the predictions.

Using the Command-Line tool

Once you have a model, you can use again the command-line tool to organize your images. From the source code folder, run something like this:

# Usage: python -m tidymypics SOURCE_DIR OUTPUT_DIR
#   Example:
python -m tidymypics --year ~/Pictures/MyMessyAlbums ~/Pictures/OrganizedAlbums

All the images will be organized by class and year of when the picture was taken. Files will be also renamed using the creation timestamp, and the first 7 characters of the image's MD5 hash to avoid duplicate file names.

If you don't add the --year option, pictures won't be grouped by year, but they will all live in the same directory, separated only by class.

If you add the --move option, pictures will be moved instead of copied.

To know more about all the options and the usage, just run python -m tidymypics without any parameter.

Implementation Considerations

You can use anything as your classification class names (e.g. people, pet, landscape, artwork, text, nudity, etc.), but take in consideration the model architecture and layers:

• Adds random noise
• Flips images randomly (either vertically or horizontally)
• Reads the images as grayscale. If color information is important for your classification, you should modify the code and use rgb with 3 channels.
• Reads the images resizing them to 800x480px
• Uses a Drop Out layer

These transformations are only done in-memory, to compute the model, and not to the original images.

Check tidymypics/core.py, build_model() function to know more details about the implementation.