To run the script just call the main.m MATLAB script file.

I am using the formula:

           W = Y^T * X * (R * c * I)^-1 , (*)

to compute the weights for the linear classifier.

I have the following set of features:

cross_feature: consideres the dot product of the current image, with the 'perfect' representation of the digits (i.e. the representation which results when you average the grey values for each pixel of the training set for each digit). This features measures in a way the similarity of the tested digit and the perfect written one.

row/col_stripe_feature: this features examines every strip the image (row and column wise) and counts the number of strips that it intercepts, given certain threshold value. That is so because as the first features measures similarity in a way, this should take care if the digit is skewed in some firection for example.

bfs_feature: I use this feature to examine the closed regions (islands) of the image, excluding the ones that are touching the borders. It should help with distinguishing between 0 and 8, for example, since both of them have a similar shape, however, 0 has only one 'island' and 8 has 2.

intensity_feature: Here I break the image into 4x5 quadrant blocks, indexed 1 through 12, where we have 4 rows of 3 such blocks on each, and I compute the average grey scale value in each one of those. This feature should account again in a way for the relative structure of the digits, where some skewness and unaccuracy is expected within the blocks.

With the compute_feature_vectors.m script I have tried to make the feature engineering (addition, removal, modification) as modular as possible, where we can have an external script computing different features, and in the compute_feature_vectors.m script we just put together the building blocks.

I use 10-fold cross-valudation (find_best_tychonov.m) for the normalization coefficient c for (*), the result of which over the test portion of the training data set you can find in cross-valudation.jpg.

Running the main.m script with the coefficient we got from the x-validation yields error rate of 3.3%, where we have no error when recognizing 2 and 7, but we have the most at 3, 5 and 8, which look similar to some extend.