Reconstructing Obfuscated Human Faces with Conditional Adversarial Network

Abstract

In today’s era of advanced forensic and security technologies, the problem of identifying a human face from a low-quality image obtained from low-quality hardware or other reasons is a major challenge. Trying to extract meaningful information from these images is very difficult. These low-quality images have mainly two kinds of distortion in it, i.e. blurring and pixelation. Prior attempts have been done using different machine learning and deep learning techniques, but the desired high-quality images are not obtained. In this paper, we have used the conditional adversarial network to reconstruct highly obfuscated human faces. Various previous works on the conditional adversarial network have suggested it as a general-purpose solution for image-to-image translation problems. The conditional adversarial network is able to learn mapping from the provided input image to resulting output image. It is also able to learn the loss function to train the mapping. We have examined the result of this model using pixel loss function which gives the exact mapping of obtained high-quality human face with ground truth; furthermore, we have examined the capabilities of this model with very high-level obfuscated images.