Image Classifier

Docker based Python tool for classifying pictures based on the average color value.

Usage

• Launch the app with command

  • make for error level logging
  • make build-verbose for info level logging

• After the app is running, you can add images (or folders with images) into the system with command docker cp <PATH> module_parser:/images/unclassified/

• PATH can be path for an exact image or folder (with sub-folders) of images. Unsupported image formats or images without extension will be ignored.

• To check contents of classified files folder, you can use docker exec -it module_parser ls -LR /images/classified

• Run the tests with command make test

• Remove app containers, images and volume make clean

Requirements

• Docker 19.03.6 and up
• Docker-compose 1.27 and up

Overview

Image classifier uses Docker, NATs messaging system and OpenCV to process images, determine the average color using numpy and then classify it in a folder with given color (16 base Web colors).

Images are stored in a volume, which is attached to all 3 modules. This volume has given hierarchy:

-───unclassified
│   │   img1.png
|   |   ...
│   └───subfolder-with-images
│       │   img10.jpg
│       │   ...
│   
└───classified
    └───red
        │   red-image.png
        │   ...

After the image is classified, it is moved into correct directory without changing its name.

Supported image types

• jpg, jpeg
• png
• bmp
• tiff, tif

Communication

There are 3 main modules participating in the image classification process

1. Parser
   • scans the unclassified folder for new images
   • if a new image is found, message is sent with path to image using NATs
2. Image handler
   • Subscribes to parser's messages
   • After a new message is received, opens image with OpenCV and calculates average color using numpy
   • Sends color information and file path using NATs
3. Classifier
   • Subscribers to image handler's messages
   • After a new message is received, finds the exact or closest Web color out of 16 standard colors and moves the image into a folder

Color classification

If the average color does not fit any of the 16 standard Web colors, a closest Web color is determined by calculating Euclidean distance in the RGB space where each of 16 colors is compared to the average color and then smallest distance is chosen.

Example usage

1. make
2. docker cp ~/Downloads/images/ module_parser:/images/unclassified/
3. docker exec -it module_parser ls -LR /images/classified

And the result is:

/images/classified:
black  blue  lime  red	silver	white

/images/classified/black:
black_bmp.bmp

/images/classified/blue:
blue_jpeg.jpeg

/images/classified/lime:
green_jpg.jpg

/images/classified/red:
red_png.png

/images/classified/silver:
grey_tif.tif

/images/classified/white:
white_tiff.tiff