An improved face recognition with T2FSN based noise reduction in unconstrained environment

Face recognition is a significant biometric system employed to determine a person’s identity from digital images. The application of a face recognition system is mainly for monitoring and security purposes. Several methods are introduced to enhance face recognition performance for security applications in an unconstraint environment, among which few have achieved impressive outcomes. However, its performance is minimized due to various aspects like illumination, pose variations and expression. Recently, there have been more studies for face recognition based on deep learning and have accomplished positive outcomes. This paper proposes a novel optimization-driven-improved deep learning model (OD-IDLM) for identifying the human face in various disturbing environments. Initially, the proposed OD-IDLM performs pre-processing using a type 2 fuzzy system and Nelder mead algorithm (T2FSN) for noise removal. Then, feature extraction uses a morphable convolutional neural network (MCNN) to detect complex features and fit a 3D face from an uncontrolled environment. Further, based on the extracted features, face recognition is accomplished using Shuffled Shepherd Bird Search (SSBS)-ResNet (SSBS-ResNet) to enhance face verification performance. SSBS-ResNet’s weights are updated using SSBS, the hybridization of shuffled shepherd optimization (SSO) and bird swarm optimization (BSO) to expand the recognition rate. The proposed OD-IDLM is implemented in the Python platform using LFW, CASIA WebFace and YTF datasets, and performance is assessed in different evaluation metrics. In addition, the performance of a proposed OD-IDLM is compared with recent existing classifiers to determine the efficacy of face recognition. The simulation outcomes show that the maximum recognition accuracy obtained by the proposed OD-IDLM is 99.8%, 99.28% and 97.45% for LFW, CASIA WebFace and YTF datasets. It proves that the proposed OD-IDLM is accurate and superior to existing classifiers.