Logarithmically Pre-processed EMACH Filter for Enhanced Performance in Target Recognition

Abstract

The image transformation with respect to mammalian retina-based sensor geometry enables us to resolve some major difficulties in object recognition. Logarithmic mapping is a mathematical transform that converts scale and angular variations to image shifts which helps in classifying target objects correctly. The primary issues dealt with in the design of correlation filters include the ability to suppress clutter and noise, easy detection of the correlation peaks and distortion tolerance. The correlation pattern recognition concept is applied for developing filters that provide invariance to distortions by adopting a suitable training mechanism. But, a separate training requirement for each distortion type is a drawback in correlation filters. By joining a recent variant of correlation filters, i.e. the Extended Maximum Average Correlation Height Filter with the strengths of logarithmic translation, we get a supplementary advantage which is useful in most of the recognition applications. The proposed combination yields a design which can manage multiple distortions in a single training instance with pronounced detection results against complex input scenes.