Viewpoint Invariant Disguised Facial Recognition using Deep Networks
Hi! I am Sareya Qayyum and I am currently doing Masters in Computer Science from COMSATS University Islamabad (CUI) Lahore Campus, Lahore Pakistan. My Major is Computer Vision And I aspire to become a Data Scientist. Currently I am working on one of its domain: Facial Recognition. My Research topic is Facial Recognition under Disguise Variations. A short introduction about my Research is given below.
1. Problem Statement: Viewpoint Invariant Disguised Facial Recognition using Deep Networks.
2. Objective: For the past couple of decades research in face recognition has seen remarkable progress. Many algorithms work excellently in performing face recognition in constrained environments, the current algorithms achieve high performance in unconstrained environments too. But these algorithms fail to capture the complexities and challenges of face recognition under disguise variations. In the real world, an individual may use disguise with intentional or unintentional purposes so that his identity may be concealed or he might try to impersonate another person’s identity. Facial recognition systems should be capable of identifying such disguises and classify the individuals as disguised or impersonators. Many deep architectures have been presented to address this problem. A novel deep learning architecture called Capsule Network will be used to address face recognition under disguise. According to the literature review Face recognition with disguise has not been addressed using Capsule Network until now.
Facial Recognition can be classified into two types:
Facial Recognition in Constrained Environments: Recognition systems that require an upfront image of a person with little or no PIE (pose, illumination, expression) variations are termed as constrained Facial Recognition.
Facial Recognition in Unconstrained Environments: Recognition systems that require images taken in the wild meaning they have PIE (pose, illumination, expression) variation. A person need not be in an upfront position for the recognition system to work properly. Main resources for unconstrained facial recognition system are surveillance videos where a person is unaware of the security camera.
3. Disguised Facial Recognition:
Disguised Facial Recognition is one of the problems of unconstrained environments and researchers are trying to investigate this provocative problem. For facial recognition in unconstrained environments like disguised facial recognition, surveillance is the main resource of data procurement. Thus systems working in DFR require viewpoint invariance in order to be able to recognize a face and verify a person even if the images are prone to different views of the same person.
3.1 Applications of Disguised Facial Recognition:
Some important applications of Disguised face verification are:
Felony identification and in justice systems. In crime scenes, a face could be at any angle and can contain many noise factors leading to the obfuscation of facial features
Picture indexing or tagging for social media applications. Faces could appear with poses and orientations so it is desired to detect and tag each person’s face in a picture correctly
3.2 Challenges in Disguise Facial Recognition:
The word Disguise means ‘obscure’ or ‘obfuscate’. Factors that affect Facial recognition are called disguises. Disguises can be of two types:
Intentional Disguises: Intentional disguise means a person is actively trying to conceal his identity with different disguises and may impersonate another’s identity in order to gain access to certain security facilities. It is the responsibility of the system to detect such frauds and correctly classify the person as disguised or impersonator.
Unintentional Disguises: Un-Intentional disguises mean that a person is not trying to evade the facial recognition systems deliberately but it happens coincidently or by chance.
For Intentional and Unintentional disguises many factors collaborate to make the disguised facial recognition a challenging problem. Some of them are:
Wearing Hats
Sunglasses
Makeup
Wigs
Tattoo inscription
Different facial orientation and positions
Light and Camera effects
Age Factor
Constrained Facial Recognition Systems are incapable of dealing with such vast variations. Or they may need computationally expensive pre-processing steps in order to normalize these variations. Inter-class variations are so minimum in disguised faces between different subjects which makes it a challenging and interesting problem as well.
4. Methodology:
The approach to be used in solving the problem of disguised facial recognition is a novel deep learning architecture called Capsule Network. The concept was presented by Hinton et al along with two published papers exploring a new algorithm called ‘routing by agreement’ to address the limitations of deep convolutional neural networks. Capsule Network claims to provide viewpoint invariance thus reducing the need for data augmentation used to increase the size of dataset for convolutional neural networks. This claim was proved by Hinton et al by achieving 99.75% test classification accuracy using MNIST dataset.
4.1 Basic Achitecture of Capsule Network:
5. Outcome: For a given pair of image, system will output a genuine match if both images are similar, else it will produce an output saying the pair is a mismatch/imposter.