Very Fast Rotation and Scale Invariant Multiple Template Matching Using Histogram-based Associative Network

This paper introduces a very fast method for histogram multi-template matching using a fast Multi-layer Perceptron (MLP) based auto associative memory. The method achieves a high-speed improvement by rapidly evaluating the hidden unit outputs. We demonstrate that the hidden unit outputs can be rapidly calculated for any region and scale in a constant time using the integral image technique. Compared to traditional methods like the conventional, distributive, and integral histogram methods, the proposed approach outperforms them with speedups of up to thousands of times. Furthermore, it preserves the desirable properties of rotation and scale invariance.