An Introduction to Age Invariant Face Recognition

Hi everyone, Greetings and Welcome to you all to my first medium article. This article will discuss one of the most popular and challenging computer vision applications: Age Invariant Face Recognition (AIFR).

Since this is a huge topic to discuss and this article won’t be enough to complete everything, I would like to just introduce AIFR, its challenges, motivation, and its various applications.

But before that, I will just briefly define what is Face Recognition. It is identifying or verifying an individual's identity using their face and it is of the two types.

(i) Face Identification: is a classification problem that classifies a face image

(ii) Face Verification: is determining whether given two face images belong to the same identity or not. Now coming back to the definition of AIFR. As they say, a picture is worth a thousand words.

Figure 1. Example showing pictures of Cristiano Ronaldo at different age

You would have correctly guessed this famous Footballer by now, yes all these pictures are of the Footballer Cristiano Ronaldo. So that sums up what is Age Invariant Face Recognition, it simply means recognizing faces across different ages while the person’s face has undergone several changes.

The human face is a complex object and, modeling the various changes that it undergoes as its ages is well-nigh impossible. Changes due to shape and texture variations occurring in one person across different age groups result in more intra-class variation than inter-class variation. Sounds confusing right! the following figures might clear your doubts.

Figure 2. Example depicting More Intraclass Variation than the Interclass Variation

Yes, the first pair of photos look like the same person compared to the second pair even though it is from a different person, which means more intraclass variation or difference than the inter-class variation.

Besides that, Face Recognition faces many challenges such as different poses of the person, the image quality of the input data, lightning condition of the image, and partial or complete occlusion.

Figure 3 Challenges of Face Recognition

AIFR provides a wide range of applications that are used in practical, day-to-day life scenarios which sets the tone for my motivation behind this blog.

Motivation and Applications

The motivation for the problem comes from numerous real-life applications. The following is the list of some of the notable applications of AIFR in real-world scenarios.

Biometrics: It is uniquely identifying one person using one or more distinctive biological features. The unique features include fingerprints, retina, and iris patterns. Moreover, improving recognition of the face during aging also boosts the robustness of the model. During this period, a face may have undergone significant change. It is an example where the person’s face will compare with the image taken long before.

Forensics: AIFR perhaps finds maximum use in forensics. Law enforcement agencies often need to recognize and identify wanted criminals by comparing suspects’ faces to mugshots that could have been taken years before. There may be cases when the forensics experts require to change the age of the criminal’s face. Hence, this demands a robust face recognition model in the presence of aging.

Medicine: It is used for diagnosing the disease by discovering the improper premature aging of a person. Identifying the disease early will help the treatment of a person effectively than discovering later in the future.

Because of the sensitive nature of these applications, robust and accurate algorithms must be developed to recognize faces across ages.

Approaches

There are two approaches to tackle this problem.

One of them is the generative approach where we attempt to construct the 2D or 3D model to compensate for the age variation caused by the aging process. This approach requires assuming several parameters and requires accurate age labels to tackle the problem. Due to the complexity of modeling the exact aging process, this approach becomes very expensive and is not at all suitable for real-world applications.

Another approach is the discriminative approach where the methods aim to model a suitable feature representation and an efficient matching framework. The problem with this approach is that the feature still consists of age information, hence is not suitable for AIFR.

Deep Learning-based AIFR: Recently, Deep Learning (DL) has gained momentum in tackling this problem because of huge data availability, computational power, etc. The ability and strength to learn and represent hidden features from the given input data have made deep learning models successful in achieving better results in contrast to the traditional conventional approach.

Figure 4. Deep Learning Neural Network

Conclusion and Takeaways

This article briefly discussed Face Recognition, Age Invariant Face Recognition, its various challenges, the motivation behind this problem, and its various real-life applications. It was so wonderful writing and sharing this post. I hope you also enjoyed reading it. In my next article, I will discuss the various datasets and Deep Learning algorithms for solving this problem. Till then Stay Safe and Keep Learning :)