Shape context based clutter defiance in logarithmic correlation filters

Correlation based classification decisions are challenged by the occurrence of clutter in the real life input scenes. Clutter tends to cause side-lobes in the output correlation plane thus making correlation based classification decision nontrivial. The pronounced side-lobes may dominate correlation peak for true class objects leading to a false alarm. We propose a new hybrid scheme involving shape context and correlation filters to avoid impact of clutter in correlation output. The proposed combination provides a dual validation strategy where presence of true class object in test image is confirmed at two different stages before final class decision. Clutter participation in correlation process is denied by matching shape contexts of each object in test image with the true class object. The recommended addition in optimal logarithmic maximum average correlation height filter (LMACH) results in a performance gain ranging from 30% to 70% in peak to side-lobe ratio leading to a correct classification decision. The improvement remains consistent throughout the range of in-plane angular distortions against synthetically cluttered challenging images.