This is a tool for performing image classification. This project primarily serves as a starting point for developing and experimenting with more sophisticated image classification algorithms using c++. This project is intentionally small for the purpose of being flexible and easy to understand.

ImClass is only known to build and run on Mac OS 10.8, however other systems will probably work as well.

ImClass has the following dependencies:

• OpenCV (Open Source Computer Vision)
• Boost C++
• UnitTest++ (For testing)

If you are on a mac and have homebrew, you can install each of them with the following:

brew tap homebrew/science
brew install opencv
brew install boost
brew install unittest-cpp

Once you have all of the dependencies installed, you can compile the program and run the tests using CMake and Make as follows:

mkdir build
cd build
cmake ..
make
make test

The example programs are relatively simple. The program visual_vocabulary generates the visual vocabulary which is basically a set of image feature descriptors that will be counted in each image.

$> visual_vocabulary directory/with/images [output.vv]

The program classifier uses a set of classified images in a directory expected to have the layout:

images/apples/foo.png ... images/apples/bar.png
images/bananas/alpha.png ... images/bananas/omega.png
...

Along with the visual vocabulary to train a bag-of-words classifier.

$> classifier directory/with/images vocab.vv [classifier.cls]

The last program classify uses a visual vocabulary and a classifier to determine the class of an unknown image.

$> classify mysteryimage.png vocab.vv classifier.cls
5
$>

It prints out a number corresponding to the internal representation of the determined class.

In the test directory, there are a number of test image sets.

• images - a set of images that the classifier performs well for
• photos - a set of images that the classifier performs poorly on
• bad_images - a set of images that are designed to break the classifier