Hybrid meta-heuristic algorithm based deep neural network for face recognition

Abstract

Face recognition has been active research in the security domain. Human face recognition gains importance for developing a secured environment for the organization and also enhances the usage of artificial intelligence for security. Face recognition has been studied over the years for accurate recognition of complete face images. However, in the real case, the presence of occlusion and noise in the image significantly affects the performance of the recognition. Even though a lot of research has been carried out in handling the occluded and noisy image, more refinement is required to achieve high accuracy. This paper proposes a simple and efficient face recognition system with occlusion and noisy faces using the deep learning concept, as it has the advantage of handling all of it. The developed model undergoes four main steps like (a) preprocessing, (b) cascaded feature extraction, (c) optimal feature selection, and (d) recognition. Initially, the preprocessing of the face image is focused in terms of face detection by Viola-Jones algorithm. Further, a set of features termed as Local Diagonal Extrema Number Pattern (LDENP), Gradient-based directional features, and Gradient-based wavelet features are extracted for the cascaded feature extraction. As the collection of features is in a cascaded manner, it leads to providing irrelevant information of features. Hence, there is a need for optimal feature selection. The hybrid meta-heuristic concept, namely Multi-Verse with Colliding Bodies Optimization (MV-CBO), is developed with the integration of Colliding Bodies Optimization (CBO) and Multi-Verse Optimizer (MVO), and it is used for performing the optimal feature selection. Further, the optimally selected features are subjected to the optimized Deep Neural Network (DNN) for recognizing the faces, in which the proposed MV-CBO is used for optimizing the activation functions (sigmoid, tanh, Relu, ArcTan, and RRelu). The experimental findings on diverse datasets with occlusion and noises prove that the extensive experiments on several benchmark databases prove the ability of the proposed model over the existing face recognition approaches.