An efficient scale and rotation invariant 2-D object recognition method

This paper proposes an efficient scale and rotation invariant 2-D object recognition method using Complex-Log Mapping (CLM) and Translation Invariant Neural Network (TINN). CLM is known as very useful transform 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. Recently, a new method using an augmented second order neural network (SONN) was introduced as a solution. It requires, however, a connection complexity O(n²) for input feature extraction which is too high to be implemented. In this paper, we propose a method reducing 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