User Recognition Via Facial Parameters With Occlusion Using CNN

The prime essence of occlusion based face detection has been derived across the world with a simple agenda to tackle the rise of frauds that security reinforcement or surveillance techniques provide concrete solution with bio metric database. Though the prime solution is by bio metric authentication via smart phones through facial recognition of users; still the misuse is a challenge as the user’s face is often partially covered or occluded. Therefore, face detection or rather authentication even when occluded has become very important to prevent cyber frauds. Traditional approaches in terms of facial bio metric authentication typically have comprised of image processing or machine learning steps: feature extraction, pooling, segmentation, flattening and recognition. The researchers of this paper propose a facial occlusion detection network utilizing Convolutional Neural Networks (CNN). The result showed that accuracy in our developed model is in the range of 88 to 91 percent which in competence with the reported algorithms till occlusion of face in terms of percentage is restricted up to 50 percent of the focus area and 40 percent of the complete face.

I. Introduction

In a persons life a human can not only remember but identify various faces without difficulty in life; yet when we talk about modern machines made to do human job faces tremendous difficulties especially in the case of human face detection atomization. The limitation of any machine is evident instantaneously as the task done by a human takes hardly fraction of seconds whereas for a machine algorithm time constraints are enormous based on resource, area and cost limitations. The complexities are just not restricted to resources but on the matter even the surrounding environment conditions like that of light, or angle of vision for a human and in that context angle of capture of image for a machine show machines lacking behind human by a huge gap. Facial recognition has been under research since 1960’s but the success of algorithms is tested recently with the advent of mask usage because of pandemic. In current research taking information from surveillance systems is not just restricted to gender identification or human traffic monitoring rather it is required especially for bio metric authentication [1–5]. In bio metric authentication systems facial recognition is an important category that extracts the pinpoints of a face under test by performing certain image processing mathematical operations required during training which have a unique id or authentication key associated with it. The algorithm for testing on the other hand is required to compare information extracted from a live feed with the ones of pre-trained network so as to bring about identification success. Smartphones have made testing of such computerized systems a tremendous easy but necessary option for any login requirements on platforms like that of Aadhar, interstate border crossing, commercial shopping, banking etc. One such application that was restricted in primary phases was that of app purchases made by apple users where the platform was restricted by only the quality index users that could afford apple smartphones. With the open source boom android users also could feel the ease of life when Google Cloud Vision API came as a savior for android users. This machine learning technology is slowly getting adapted by various domain like that of its usage in amazon go shopping marts to reduce the billing constraints or rather the sports industry where third umpire decisions are automated [6–8].