Deep learning for image super-resolution and its application in urban security

Abstract

Improving the image quality of surveillance cameras is highly demanded in urban security, which could greatly ease the workload of policies in identifying criminals. Deep learning, especially for generative adversarial networks (GAN), had been used in increasing image quality by translating low-resolution images into high-resolution images. The technology that converting low-resolution images into high-resolution images is called image super-resolution. GAN consists of a generator and a discriminator in which the discriminator is to discriminate real and fake high-resolution images while the generator is to convert low-resolution images into fake high-resolution images that can not be differentiated by the discriminator. The generator and discriminator work together in a competitive manner that tries to improve each other. This paper aims to explore GAN-based methods for image super-resolution, which is important for urban security. We first provide readers with theories about GAN. Moreover, we introduce four popular methods based on GAN for image super-resolution including super-resolution GAN, enhanced super-resolution GAN, residual channel attention GAN, and super-resolution GAN with ranker. We also list some challenges about applying image super-resolution to urban security and provide the corresponding solution.