This is a fairly simple image classification Python script that I wrote as a project deliverable for my 2014 Machine Learning class at the Polytechnic University of Catalonia (UPC), back when I was doing the Master in Innovation and Research in Informatics program. It performs image classification using the SIFT (Scale-Invariant Feature Transform) module of the popular OpenCV Python library. SIFT is a feature detection and description algorithm first published by David Lowe in 1999.

The classifier works by first reading a set of training images, from which features are detected and computed using SIFT. It then builds models and determines the parameters necessary for classification. Later the classifier can be executed to perform classification on a set of input images (also known as query images). Only images in JPG and PNG formats are accepted.

DISCLAIMER: It's been years since I wrote this code and my understanding of image classification is now fuzzier than it was even then, but I didn't want to lose it so I decided to put it up. If you are an OpenCV/machine learning expert and have some feedback/corrections, please let me know!

ADD'L CREDITS: The k-means clustering implementation (Lloyd's algorithm) was obtained from the wonderful Data Science Lab.

This code was developed and tested on:

• Python 2.7 with numpy 1.15.1
• OpenCV 2.4.13.6 (NOTE: OpenCV 3.x no longer has the SIFT/SURF modules included by default)

Once all the dependencies have been installed, run:

$ python cv-classify.py

To get a list of options:

-t, --training <training-data-dir>       Enter training mode and use given directory for training data
-c, --classify <img-dir>                 Enter classify mode, using the images in the given directory as input
-v, --validate                           Perform validation (LOOCV) when clustering in training mode (default: false)
-o, --output <output-dir>                Location of training data file (in training mode) or classified images (in classify mode) 
                                         (default: <training-data-dir>/cv-classify.train training mode, <img-dir>_out in classify mode)
-a, --algorithm <classifier>             Use either "bf" (brute force) or "hist" (histogram) as the classifier algorithm (default: "bf")
-d, --data <training-data-file>          Use given file as source of training data
-r, --results <prefix-value>             [USE WITH -c] Check results after classification by inspecting filenames 
                                         (filename that starts with the given prefix means it should be classified as a positive)

Let's say you want to train the classifier with a set of images of circles located in the directory $HOME/cv-classify/training-circle.

You can do this by running

$ python cv-classify -t $HOME/cv-classify/training-circle

If all goes well the training data file cv-classify.train should have been created in the same directory:

If you want to change the location of the training data file, use the -o or --output option to specify the target directory.

Now let's try to classify the images located in $HOME/cv-classify/classify-circle:

To start classifying using the previously-generated training data:

$ python cv-classify -c $HOME/cv-classify/classify-circle -d $HOME/cv-classify/training-circle/cv-classify.train

By default the classifier will use the directory $HOME/cv-classify/classify-circle_out (creating and/or deleting it if necessary) and then copy the images found in $HOME/cv-classify/classify-circle to either $HOME/cv-classify/classify-circle_out/yes (positive classification):

or $HOME/cv-classify/classify-circle_out/no (negative classification):

depending on whether or not the classifier considered an input image as a "positive match" to its training data.

If the images to be classified have a prefix that identifies a successful match, the script can automatically verify the results of the classification and produce a confusion matrix for you.

Going back to our example of circle images, if the filenames of the input images of circles are prefixed with "circle" (e.g. "circle1.jpg", "circle-blue.png"), we can verify the results of the classification using the -r or --results option:

$ python cv-classify -c $HOME/cv-classify/classify-circle -d $HOME/cv-classify/training-circle/cv-classify.train -r circle

If all goes well you should output similar to the following:

-=-=-=-=-=- CLASSIFICATION RESULTS -=-=-=-=-=-
Results were computed by checking if the filename starts with "circle"
Column headers represent the truth, row headers represent the predictions done by the classifier
     YES   NO
YES    7    1
NO     2    6

Correctly classified: 13/16 (81.25% accuracy, 18.75% error rate)

According to this result, the classifier got it right 81.25% of the time. Not bad!

Some ideas on possible improvements:

• Validate understanding of image classification to improve success rate
• Provide ability to tweak internal parameters (e.g. match ratio, k-means cluster sizes, etc) from command line
• More testing to improve handling of errors/edge cases
• Make code more "pythonic" (I come from a Java background :))
• Implement other image classification algorithms (e.g. naive Bayes) and techniques (e.g. sliding window)