Article Preview
Top1. Introduction
In today’s world of technology, the researchers in the platform of computer vision and pattern recognition including security applications in terms of face recognition have been an important topic and also required for research programs (Wu, et al., 2021). Face recognition requires only a least amount of cooperation users, rather than other accessible biometric cooperative subjects. Face recognition has been providing a high security system, however the frontal faces images can be constrained under normal lighting, and so, the accuracy is affected (Jeevan, et al., 2022). Some applications have a tendency to bring severe conflicts to face recognition, e.g. occlusion is considered to an important problem in application of face recognition techniques. Human faces are recognized with high accuracy in restricted environment. But, there can be obstacles to recognize faces in practical environment such as occlusion, variation in poses changes in illumination (Mishra, et al., 2021). Therefore, a robust recognition technique is implemented to use in practical environments. In the case of pose changes, it includes dissimilarities in feature spaces and in creating pose-independent recognition complications. Although, pose-independent face recognition is realized after merging with certain pose dependent classifiers (Ren, et al., 2021).
In practical environments, human faces get occluded with scarves, sunglasses, and other stuffs, and thus, to overcome this occlusion, robust classification is important (Yang, et al., 2013). The shadows on faces due to illumination changes are also considered as occlusion. As a result, the face recognition accuracy in practical environment is improved with robustness. Due to occlusion, the face images are often vulnerable and cause numerous issues in face recognition (Ding, et al., 2015). The degradation in face recognition is a result of the occluded parts in the face images, and the robust algorithm is essential to real applications for solving the occlusion. Until now, in literature, lots of approaches that can deal with the occlusion are proposed (Wen, et al., 2016). Day by day, the face recognition applications are developed for the benefit of security purpose in public places. However, the face recognition has gained more efficiency, and accuracy (Mi, et al., 2020). Due to the few variations like illumination variation, differences in facial expression, effects of aging, and variation in pose, the application of face recognition is restricted (Fu, et al., 2017).
Deep learners have reorganized the face recognition knowledge, where deep learning classifiers design most of the well-executing face recognition models (Masud, et al., 2020) (Shah, et al., 2016). In the environment of biometric recognition, the CNN (He, et al., 2019) has performed well to learn a large set of features as it has a series of layers. CNN classifiers contain most of the accomplished face recognition algorithms (Alrjebi, et al., 2017). Larger scale inputs are trained with CNN classifiers to get efficient values, the issues can be overcome with the accessibility of large-scale datasets (Chandrakala & Devi, 2021). Recently, the borrowed feature embedding ideas in deep learning schemes and a “Robust LSTM-Auto encoders (RLA)” model is projected, in addition of an LSTM encoder (like feature extraction) and an LSTM decoder (Li, et al., 2007) (like error coding). On comparing with shallow methods (Yang, et al., 2013), the limitations of RLA have seen all kinds of facial occlusion in training stage, and its never-seen occlusion needs to be tested. Well-controlled images like pose, illumination, and variation without occlusion are trained and tested by previous works (Yang, et al., 2015). Hence, the deep learning-based face recognition model is developed here for solving the occlusion problem along with optimal pattern extraction.
The major contribution of the designed model is given here.