Kohonen (SOM) Implenetation

We want to viualize 3D input (rgb) into 2D using SOM network.

The program uses two different methods to extract colors:

• K-means clustering to spit the image in RGB color space to the n_colors clusters. The cluster mean points are selected as colors in the palette.
• Kohonen Self-Organizing Map (SOM). Random selection of original image pixels are given to the SOM algorithm, which then maps them to a n_colors x n_colors two-dimensional palette.

I have three files containg main.py processing.py and som.py you only have to start with main.py the som.py contains the main class of SOM which the logic of SOM networks is implemented there the prcessing.py contains the other details of converting which you can see in the code(the have comments :D)

Test Results

The results of my program are attached to testcases folder

How To run:

Start with main.py and put your image link in the img_link: select n_colors which can be any number you want but in this homework it is 40 (40*40 = 1600) If you wanna give the rgb numbers directly into a list goto peocessing.py in the sownload_image function and set this value no_download = False In this way you can give the data manually

Efficiency

Because it was really boring and took alot of memory I set the no_download flag to false and directly convert image to numpy array( not python list which is terrible) and that makes my program efficent

Dependencies

you need to have the follwoing libraries:

• numpy
• pandas
• matplotlib (for visualization)
• sklearn.cluster( for kMean)
• sklearn.utils (for shuffle)
• mpl_toolkits.mplot3d(3D display)
• PIL (pillow for images)
• pickle
• scipy
• multiprocessing (used in som.py)

Notice:

Run with Python3

Any ideas making it better ?

Send me message at zahra_9877@yahoo.com