Invariant pattern recognition using higher-order neural networks

Abstract

This paper explores the use of a higher-order neural networks to implement a pattern recognition system that is insensitive to transformations, i.e., translation, rotation, and scaling. The proposed implementation of the invariant system consists of a feature extractor (a third-order neural network) and a trainable classifier (a single-layer linear associative memory). A single parameter, sphericity, which represents the similarity of two triangles, is introduced into the third-order neural network structures, from which the invariant feature vector is extracted. In this way, the invariant pattern recognition problem can be formulated and the invariance property can be proven under the assumption that the input pattern is continuous. The vast storage requirement usually encountered in higher-order networks is overcome, since only the activated pixels have to be processed in our scheme. Translation invariance is guaranteed by our invariant structure for the grid transformation of the binary image. Simulation results for typed numerals with different feature vector lengths show that the invariant system achieves 100% recognition accuracy for rotated and scaled patterns, respectively. Accuracy up to 95.11% is achieved for the random combination of rotated and scaled patterns. A 99.60% success rate for combined transformation is achieved for the recognition of various aircraft figures.