With image classification we try to describe images by looking at the image itself. This is a very interesting approach because today all image browsers or search engines (like google or yahoo) just search images by analysing the image name. This method however returns very few results as most of the images have generically created image names like for example img00035.jpg.

Colour-Science combines different object recognition and image segmentation techniques to automatically classify images in important categories.

 

 

 

Using these categories the user can sort his images for example to find good images with people or he can for example show all the bad images so that he can delete them. Also combinations are possible. So you can for example find people in the nature or people in the city.

 

This project is still in a beta phase, but prototypes for the different classes are available. So we have for example a prototype which sorts images in a scale going from nature to urban area or we can sort images by the surface or number of detected faces.

 

Here is a sample of a image folder content which was sorted in images beginning with nature images and ending with urban area images. We have a Matlab prototype to sort a folder with images by this type of classification.

 

 

The first screenshot shows the first 25 images which are most likely nature images with no human constructions in it:

 

 

Here are much more images inbetween …………………

 

 

These are the last 25 urban area images with a lot of man made constructions in it.