Classification of patterns using a self-organizing neural network

Abstract The objective of this study is to evaluate the performance of Fukushima's neocognitron model when it is applied to complex imagery. In his original report, Fukushima demonstrated that this system could discriminate between simple alphabetical characters represented in fields of 16 × 16 pixels, and that shift invariance can be achieved through a proper choice of design parameters. The present work describes results for expanded neocognitron architectures operating on complex images of 128 × 128 pixels. These neural network systems were simulated on a VAX-8600 minicomputer. Wire frame models of three different vehicles were used to test the properties which Fukushima had demonstrated. The expanded neocognitron systems were able to classify these objects and to identify their critical features. After training, each object was placed at different positions in the plane, and the neocognitron's shift invariance property was tested. With complex ( 128 × 128 ) imagery, it was difficult to achieve proper classification and maintain shift invariance using only a few levels. In another experiment, the neocognitron trained on polar transforms of objects in the training set. Objects in the training set were rotated, and polar transforms of the rotated images were submitted as input. In this manner, the neocognitron's shift invariance was exploited to recognize rotated imagery. These investigations gave insight into the role of various model parameters and their proper values, as well as demonstrating the model's applicability to complex images.