A Smart Face Recognition and Verification using Optimal Spatial and Spectral Feature Selection with Adaptive Multiscale Mobilenet

Abstract

Face recognition is complicated work, which is highly demanding while processing the inter-class similarities and intra-class differentiations in the acquired images in a wider range. However, the identification accuracy can be enhanced at some level through managing the system with non-matched templates. In the reality, face recognition is complicated owing to differentiations in the pose, background, illumination, and so on. On the other hand, the type of face recognition technique is recently dependent on machine learning-based facial features while avoiding the practical experiences in hand-craft features. Recent world, the face recognition technique using deep learning strategy is capable of learning efficient face features for getting a highly extraordinary efficiency. The face identification techniques use various conventional appearance methods, which are further applied for testing the efficiency via applying the benchmark facial images. The facial images gathered from the standard digital media are often noticed as problems regarding occlusion, lightning, and position conditions along with camera angle. The occluded images can be observed with the alignment, facial expressions, and human pose along with the camera axis. Thus, more attention must be taken care during the acquiring the facial images along with the coverage of the back-ground. Hence, it is necessary for considering the precise pre-processing steps along with the necessary steps for recognizing the face. In this research, a new face recognition model is investigated utilizing deep learning techniques. Initially, the images are gathered from the standard resources. Next, they are pre-processed to increase the quality of images for further processes. Then, the spatial and spectral feature extraction is performed, where the spatial features are extracted by DeepLabV3, and the spectral features are extracted using Discrete Wavelet Transform (DWT) and the Discrete Cosine Transform (DCT). Next, the optimal spectral features and optimal spatial features are attained by using a new hybrid heuristic algorithm known as Controlling Parameter-based Rock Hyraxes African Vultures Swarm Optimization (CP-RHAVSO) with the combination of Rock Hyraxes Swarm Optimization (RHSO) and African Vultures’ Optimization Algorithm (AVOA) techniques. Then, these optimally selected features are combined and fed to the weighted fused feature extraction process. Finally, weighted fused features are given to the Adaptive Multiscale Mobilenet (AM-Mob) for face recognition, where the hyper parameters of M-Mob are tuned by the same CP-RHAVSO. Investigational results show that the proposed face recognition method on the standard database, according to experimental findings, yields a highly valuable accuracy rate.