Rotation Invariant Neural Network with An Edge Detection Network
Minoru Fukumi, Sigeru Omatu, Fumiaki Takeda, Toshihisa Kosaka · IEEJ Transactions on Electronics Information and Systems · 1992
This paper presents a rotation invariant neural pattern recognition system with an edge detection architecture. The system consists of an edge detection network and a trainable multilayered network. The edge detection network preprocesses an input pattern to produce the edge features, which are the activity pattern of orientation specificity cells (OSC), The OSC pattern is an input pattern to the multilayered network. This network can be invariant to rotation by any number of degrees. The network weights are completed through the back-propagation training.Finally, computer simulation for coin recognition shows the effectiveness of the system, and comparison of the system with the conventional one is described.