Optimal Gabor Kernel's Scale and orientation selection for face classification

2D Gabor-based face representation has attracted much attention. However, owing to the fact that Gabor features are redundant and too high-dimensional, appropriate feature dimension reduction appears to be much more paramount. Allowing for each individual Gabor feature constructed by a combination of scale and orientation pair, we equate feature dimension reduction problem to optimal Gabor kernels’ scales and orientation selection problem. Genetic algorithms (GAs) have represented a useful tool for optimal subset selection. However, population premature and optimization stagnancy problems exist in traditional GAs. Here we present an improved algorithm: Hybrid Genetic algorithms-based (HGAsb), which introduces the concept of the simulated annealing into traditional GAs to effectively solve the problems mentioned above and to improve optimization efficiency. Experimental results on IMM face database demonstrate that in contrast to GAs, our proposed algorithm can provide 4.25 improvements. The distributions of orientations and scales of the selected features by HGAsb are also analyzed. Results indicate that the features in the larger scales have equal importance as those in the smaller scales in discriminating nuance of faces. The features in horizontal, vertical and 225° orientations have more discriminative power.