Method of target detection in images by moment analysis of correlation peaks

Automatic target detection and recognition in images often is attempted by use of a linear correlation filter (matched filter), whose output is interpreted by a single pointwise detector (detection based on only one point). I examine a technique for significantly improving the performance of this target detection approach by supplementing the pointwise detector with several neighborhood correlation peak detectors (detection based on a domain of many points extending over much of the peak). The neighborhood detectors extract peak shape information through a moment analysis of correlation plane peaks. I describe the design of statistically quasi-optimal correlation peak discriminators based on second-order geometric moments.