FaceNet recognition algorithm subject to multiple constraints: Assessment of the performance

Abstract

Literature has it that the performance of most face recognition algorithms still decline in multiple constrained environments (Occlusions and Expressions), despite the achieved successes of deep learning face recognition algorithms. Using expression variant test face images synthetically occluded at 30% and 40% rates, the study evaluated the performance of FaceNet deep learning model for face recognition under the aforementioned constraints and when three (3) statistical multiple imputation methods (Multivariable Imputation using Chain Equations (MICE), MissForest and Regularized Expectation Maximization (RegEM)) are adopted for occlusion recovery. Results of the study showed improved recognition rates of the study algorithm when the imputation-based recovered faces were used for recognition compared with using their multiple constrained counterparts. However, test faces reconstructed with the MissForest imputation method were more accurately recognized using the FaceNet deep learning algorithm. Furthermore, the study demonstrated that some simple augmentation schemes sufficed to further enhance the performance of the FaceNet model. Specifically, the FaceNet algorithms gave the highest average recognition rates (85.19% and 79.5% for 30% and 40% occlusion levels respectively) under augmentation scheme IV (slight rotations, horizontal flipping, shearing, brightness adjustments, and stretching) using MissForest as the de-occlusion mechanism. The study also found that, no disparity existed in its performance with the choice of either Support Vector Machines (SVM) or City Block (CB) for classification under augmentation scheme IV. The study recommends using the MissForest imputation method in dealing with moderately high occluded test faces with varying expressions to enhance the performance of the FaceNet face recognition model.