Scale and rotation invariant pattern recognition using complex-log mapping and translation invariant neural network

In this paper, we propose a scale and rotation invariant pattern recognition system using complex-log mapping (CLM) and translation invariant neural network (TINN). CLM is very useful for extracting scale and rotation invariant features. However, the results are given in a wrap-around translated form, which requires subsequent wrap-translation invariant recognition steps. This problem can be solved by using an augmented second order neural network (SONN). It requires, however, a connection complexity O(n²) for input feature extraction which is too high to be implemented. The proposed method reduces the connection complexity to O(n*log(n)) by using TINN. Experimental results show that the recognition performance of the proposed method is almost the same as that of SONN while its network size is significantly reduced