Shift and rotation invariant texture recognition with neural nets

Abstract

Shift and rotational invariant approaches based on neural nets have been proposed in the literature for the purpose of a training for classification of patterns independently of their position and orientation. The training usually requires the presentation of preprocessed patterns in the different positions and orientations that are involved in the testing. This means high dimensional input vectors and large number of training patterns. This situation can be simplified by using rotation insensitive information in the construction of the input vectors. The use of representative input vectors from a pool containing all possible input vectors (ideal) accounts for the shift invariance characteristic of the training. This approach leads to a considerable time saving in the learning phase of the net because of the reduction of the dimensionality of the input vectors and the reduction of the number of training patterns required. This in turn reduces the number of hidden nodes thus reducing the size of the processor itself upon implementation. The example used here is the classification of textures that show directionality. >