Generative Adversarial Network for Heritage Image Super Resolution

Abstract

In this paper we have proposed Generative Adversarial Network (GAN) based Super resolution reconstruction (SRR) method for the estimation of high resolution (HR) heritage images. The proposed SRR method via the GAN model (SRRGAN) is modeled as a min-max optimization process by integrating some new loss functions which will enable to maintain spatial and textural homogeneity in the estimated HR images. Initially, we have divided the input test image into Textured and Non-Textured patches based on their inherent characteristic followed by the optimization of the GAN model via loss functions lossTeXP and lossNonT eXP respectively to yield textured and non-textured patches of the HR image. Loss function lossTeXP helps to preserve realistic and diverse textural content of heritage images along multi-scale and multiorientations. Whereas, lossNonT eXP efficiently preserves the complete high-frequency map of the heritage images. Consequently, the estimated HR images preserve sufficient image details with realistic textural content and reduced artifacts compared with the existing state-of-the-art SRR methods.