The main idea of our algorithm [8] is based on the known feature of affine transformations: transformation of a given rectangle is completely conditioned by transformation of three of its corner points.
As previously, we work inside a unit square Su of pixels for which we select three corner points (s0, st, sr), s0 being left lower corner, st left top corner, sr right lower corner. For every ( ) we only compute images of these three corner points, . Then we define offsets = , = , = , and = and use them in computing and , where - the angle with x axis and - the angle with y axis. This computation gives us the set of points which are transformed by fk onto and, therefore, the synaptic weights . Horizontal size and vertical size of the rectangle are determined by the contraction factors of fk on x and y axes and calculated only once for each fk. The inclination of each side of the rectangle is taken into account by and , respectively.
The dynamics of the neural network is the same as in the Stark's algorithm.