Abstract:
The evolution of surveillance techniques in the modern era has powered AI to develop algorithms that provide numerous opportunities to improve security and well being. Surveillance cameras monitoring a wide field of view can result in capturing indiscernible faces which makes human activity monitoring and recognition impossible. Single Image Super-resolution (SISR) provides a feasible solution to enhance such faces in order to recognise the individuals. However, lack of ground truth high-resolution methods make it infeasible to develop such devised methods to develop a standard SISR model. In this research, I explore the challenge of building a SISR system from a limited set of unaligned pairs of Low-res and High-res images. Though none of the models could produce substantial results, each one of those unravel problems in an unsupervised setting.
I propose a model that leverages literature from Style Transfer to achieve the reconstruction of noisy low-res face images, but in an extremely specific domain. I further conclude the research for exploration of models and domains in Deep Learning where we might find the possible solution to the problem.