Integrating complete gabor filter to the random forest classification algorithm for face recognition

Abstract

Local feature approach to face recognition such as Local Binary Pattern (LBP) and Gabor has demonstrated to be an excellent facial descriptor. However, due to the large dimension features generated from the Gabor filter, the computation time for feature extraction is lengthy. Likewise, LBP generates lengthy histograms, which ultimately slow down the recognition time. This paper proposes a hybrid face recognition technique called Complete Gabor Filter with Random Forest (CG-RF) in biometrics technologies. CG-RF uses Gabor Magnitude Responses (GMR) and Oriented Gabor Phase Congruency Image (OGPCI) with Random Forest as features classification. For best features, Monte Carlo Uninformative Variable Elimination Partial Least Squares (MC-UVE-PLS) regression is used to select the important features generated from the Gabor filters. The upside of the proposed technique reduces computation time significantly without compromising on the accuracy of face recognition. Further experiments were conducted on Georgia Tech (GT) and Facial Evaluation Recognition Test (FERET) face databases with regards to varied face images such as head positions and orientations, occlusion and light illumination. The outcome of the experiments on both databases demonstrated a short turnaround in computation time and high recognition rates-GT (1735.2 seconds; 94.7%), FERET (8875.6 seconds; 96.7%).