Structure

• src

  • Source code of the backpropagation algorithm

• data

  • Training and validation sets

• data/train

  • Data set generated from two images

• data/train2

  • Data set generated from ten images

• src/FourierMellin

  • MATLAB source code of the Fourier-Mellin transform

How to generate the training set

Go to data/train or data/train2 and execute ./import2.pl (edit the script to change the size of the data set).

Use src/FourierMellin/applyCavanagh.m to generate the raw and invariant features for training (raw.txt and features.txt).

Use data/tableto{lisp,svm}.pl to convert the data set to a format which can be read by lisp or the support vector library:

../tabletolisp.pl < features.txt > features.lisp

The script data/train2/scale.R can be used to scale the data appropriately.

How to train a neural network

Load 'learning-algorithm.lisp' into lisp and execute

(defparameter features
  (make-instance 'learning-algorithm
                 :network (make-instance 'ff-ann :topology (list :layers '(1023 15 10) :transf *tanh*))
                 :data    (make-instance 'data :file "features.lisp" :percentage 10)
                 :algorithm 'backprop-incremental))

to create a neural network with 1023x15x10 neurons, tanh activation function, where an incremental backprop algorithm is used. 10% of the data set will be used as validation set.

Then execute

(run features #'(lambda (a) (>= (epoch a) 200)) :learning-rate 0.001)

to train the network (the algorithm stops after 200 epochs).

To save/restore the weights of a network use:

(save-weights    features "features.weights")
(restore-weights features "features.weights")