Surveillance based on Representation learning using Generative Adversarial Networks

Abstract

Face recognition is one of the most advancing topics in Deep Learning due to its wide applications/ Recently, great progress has been made in computer vision for security and biometric applications. However that it does not offer a high degree of pose variation—The major challenge in face recognition is the variance of face attributes. The face recognition also fails when the Face is covered with arbitrary masks. To overcome these challenges,we propose a surveillance system using Disentangled Representation learning- Generative Adversarial Network (DR-GAN) a encoder-decoder structure implementing Wasserstein loss, which can be grouped into two categories. First, we apply face implanting which reconstructs an image which is masked or noisy. Second, some work appliesface frontalization on the input image to generate pose-invariant faces, where traditional face-detection algorithms are applicable, or an identity representation can be obtained using the face rotation.Together they can be used to search/detect a human face with only masked or profile image as input.We use a self collected dataset of Indian Faces to improve accuracy on Indian Faces.
Keywords:Representation learning, generative adversarial network, face inplanting, surveillance pose-invariant face recognition, face rotation and frontalization.