A symbolic/neural hybrid approach to multiple shape recognition

Abstract

Recognizing multiple shape is one kind of human visual behavior. People usually recognize several distinct emergent subshapes from multiple shapes and give them different interpretations. This paper presents a symbolic/neural hybrid system to provide computers with this kind of ability. Through this approach, the recognition system is divided into three modules. Source images are sent to the first module, that is a neural network, of the hybrid system. The network is responsible for transforming the source image into abstract visual data, named pre-attention distribution and local feature information. Then, the abstract visual data are processed in the second module that is a symbolic subsystem. The subsystem is responsible for making decision in the visual search attention processes and for managing the features of the whole shape. Finally, another neural network takes the previous results from the symbolic subsystem and performs the final recognition.