Gabor features for detection and identification

Gabor features are considered for the problems of detection and identification of multiple distorted objects in high clutter scenes. Detection requires the location of all possible regions of interest in an input scene, while identification requires the classification of the detected regions of interest into the various object classes. Gabor filters are used in a correlator to extract features that are useful for solving both stages of the problem. Detection and identification are different problems that require different information, therefore different Gabor filters are used in each stage. One of the main issues addressed is the selection of the proper Gabor filter parameters for each task. The Gabor filter function is complex-valued in the image domain. We introduce guidelines for the selection of an initial set of Gabor parameters that enable the real part of the Gabor filter (RGF) to perform blob detection and the imaginary part (IGF) to perform edge detection. A neural network training algorithm is developed which iteratively refines the initial Gabor filter parameter values and selects the proper linear combination weights for a composite macro Gabor filter (MGF). We combine the outputs from the MGF and an IGF for detection. We show that fusion of these two Gabor filter outputs reduces the detection of false alarms.; For the identification problem, we calculate Gabor feature vectors which describe the local spatial characteristics of each object. This feature space is produced by correlating the input with a new MGF and sampling the output correlation at several internal locations. A neural network algorithm is developed which optimizes the Fisher ratio for these feature vectors. It iteratively refines the initial Gabor parameter values and combination weights used in the MGF. This produces an MGF which extracts improved features for identification.