Normalization of correlations

Correlation s often used as an approach to automated pattern recognition. Generally, correlation provides a measure of the similarity between a reference template and regions of an input image. This measure is also highly dependent on intensity variations in the input image, thereby hindering the performance of simple peak detection decision algorithms. Normalization can be used to achieve intensity invariance of correlation results. This paper addresses some aspects of normalization for a few filter types. For matched filters, the Cauchy-Schwarz inequality provides an effective method by taking into account the energy of the input image within the spatial region of support of the template. For many other types of filters being considered for pattern recognition applications, the regions of support are not always limited to the area occupied by the template pattern. This excessive support can produce undesirable effects in the correlation results whether normalized or not. Benefits of normalized correlation such as intensity invariance and resistance to high energy clutter are discussed along with some problems associated with regions of support. Matched filters, phase-only filters, and binary phase-only filters are investigated. Computer simulations of several cases is used to demonstrate results.