Image Sorter

This project aims to classify images based on their technical quality. We will be able to separate flawed images (blurred, grainy, incorrectly exposed, and so on) from the set when there are better ones to choose from. Our goal is to make this useful to photographers who want to spend less time sorting out poor images and more time improving good ones.

We began this project in November, 2010 and the rest of this web page will give you information on the original work. Our planned completion date for an updated, more accurate project is mid-March.

Our GIT Repository can be viewed here. Keep in mind that the code there may not be stable.

Image Sorter will rate a photo based on its overall quality. A photo will be rated on a 0 to 9 scale (0 for poor and 9 for excellent). The quality is determined by weighted ratings from the following five categories:

1. Duplicate Images

   In the case of duplicates, the one of highest quality will be selected. Duplicate or similar pictures in a submitted set are omitted. To detect a duplicate, a photo is divided into 9 quadrants, and the average color is calculated for each quadrant. If corresponding quadrants of each picture are similar, then they are duplicates.

2. Exposure

   The amount of light that enters the camera at the time of the shoot can affect the quality of the picture. Too much light will make everything too bright and too little may make everything indistinguishable. The over/under exposure algorithm takes into account RGB as well as YRGB values by looking at the luminosity clippings and locality of the different levels of exposure. MORE ABOUT EXPOSURE
   TURK EXPOSURE RESULTS VS PROGRAM RESULTS

3. Middle Gray

   Aesthetically pleasing photographs exhibit the amount of light exposure as close to a certain "gray" color as possible, known as middle gray. And is about the equivalent of 18% light exposure in a photograph or 12%-13% in digital cameras. This criteria tests how close the entire picture is, on average, to middle gray. The middle gray algorithm rates on how close to middle-gray the image is.

4. Color Pallette

   By looking at saturation levels and the amount of harmonious colors, the aesthetic quality of the colors are determined. MORE ABOUT COLOR

5. Blur

   Photographs that are too blurred are obviously not very pleasing pictures. Blur Detection calculates the contrast between a foreground subject's edges and its background to calculate the amount of foreground blur while allowing background depth-of-field.

RELATED WORK

878 photos were rated by both real people and Image Sorter. Our user study consisted of 300 paid users through Amazon Turk.

To the left you can see a graph comparing what our algorithm ranked to what users ranked.

Here are two example ratings: Rank: 2.1 Rank: 8.4

MORE RESULTS

The application is written in Qt by Nokia (a C++-based GUI builder). Our source code is provided free-of-charge under the GNU General Public License 3.0. The license agreement can be viewed here. You are free to use, change, and share the application under these terms.
Download source code [.zip] (1.71 MB)
Instructions for running the program are provided in the README.