MATLAB Eigenfaces Toolbox

This MATLAB Toolbox implements the 'Eigenface' algorithm for feature extraction and classification of faces. For details see http://en.wikipedia.org/wiki/Eigenface.

This MATLAB scripts are the results of a project for the lecture course 'Data Fusion in Sensor Systems' of the study program 'Human-Centered Computing' at the University of Applied Sciences Upper Austria (http://www-en.fh-ooe.at).

Project members: Christoph Absenger, Christian Salomon, Mario Winterer.

See eigenfaces.m for examples of how to load images, train a model and use it for classification. It also contains several sections for validation.

For testing we suggest The Database of Faces (formerly 'ORL').

The function eigenfaces_load(path, ext) loads all images with extension ext from all subfolders of a folder specified by path. Every subfolder is treated as separate class (= face). So the image folder structured should be as follows:

path
+-- face1           (class 1)
|   +-- img1.ext    (image 1 / class 1)
|   +-- img2.ext    (image 2 / class 1)
|   \-- img3.ext    (image 3 / class 1)
|
+-- face2           (class 2)
|   +-- img1.ext    (image 4 / class 2)
|   +-- img2.ext    (image 5 / class 2)
|   \-- img3.ext    (image 6 / class 2)
...

(Note that the names of subfolders and files do not matter, but all images must have exactly the same size).

The result of calling eigenfaces_load(...) is a h x w x n matrix (h = height of an image, w = width of an image, n = number of images) as well as a 1 x n vector containing the classes (target values) of all faces as numbers. These classes (target values) are numbered in ascending order according to the order in which the faces-subfolders are listed:

% load images and their classification
[I, C] = eigenfaces_load('<path>', '*.jpg');

% get 5th image
img5 = I(:,:,5);

% get class (=face) of 5th image
class5 = C(5)

This results can be used directly for training the eigenface model using eigenfaces_model(I, C).

The function eigenfaces_flatten(I) rearranges the image matrix into a matrix of size n x (h*w) as follows:

• Every row contains one observation (= image)
• The image-data is flattened into a one dimension array (row)

The results can easily be combined into one big matrix to be used in MATLAB functions like datasample or crossval:

% flatten images
I_flat = eigenfaces_flatten(I);

% combine into big matrix
data = [I_flat C'];

In order to perform classifications, it is required to train an Eigenface model using eigenfaces_model(I, C). The result is a 1x1 structure array containing all required information about the trainied model (but not the original images themselves):

efm = eigenfaces_model(I, C);
disp(efm);
% result:
       meanface: [1x10304 double]
     eigenfaces: [10304x400 double]
    eigenvalues: [400x1 double]
        weights: [400x400 double]
          class: [1x400 double]

To classify an image, either use eigenfaces_classify(efm, image) directly, or compute the weights of the image to classify via eigenfaces_weights(efm, image) first and use any classifier (e.g. kNN) then to find the best match of the resulting weights within efm.weights - the weights of the trained eigenface model.

Internally, eigenfaces_classify uses kNN(k=1, euclidean distance) to find the best match. The results are the index and the class (target value) of the best match:

[face, dist, idx] = eigenfaces_classify(efm, image);
fprintf('best match: image #%d; face #%d (distance: %f)\n', idx, faceId, dist);
% result
  best match: image #68 face #33 (distance: 834.564822)

The function eigenfaces_test(xtrain, xtest) is thought for being used by different validation algorithms like crossval. It uses the given training data set xtrain to train an Eigenface model (using eigenface_model) and then tries to classify the given test set xtest, returning the classification rate (number of correctly classified samples / total number of samples).

The two data sets xtrain and xtest must be n x (h x w + 1) matrices, where every row contains the image data of an observation and the classification (= face id) of that image in the last column. This structure can easily be established by using the alternative date format (eigenfaces_flatten(I)):

I_flat = eigenfaces_flatten(I); % flatten images
data = [I_flat C'];              % append class column

Using this test data, it is now easy to perform a 10-fold cross-validation by calling crossval:

result = crossval(@eigenfaces_test, data);
disp(mean(result));

For visualization, some functions accept optional parameters:

%% create model and display Eigenfaces and weights
efm = eigenfaces_model(I, C, 'Show', { 'Eigenfaces', true, 'Weights', true});

%% classify image and show distances to all training faces and weights of input image
[face, dist, idx] = eigenfaces_classify(efm, image, 'ShowDistances', true, 'ShowWeights', true);

See the two examples for more info.

• All images must be grayscale.
• All images must have equal size.
• All images are loaded into memory at once, so this toolbox may not be suitable for large data sets.