Abstract
The single image super-resolution (SISR) is a challenging problem due to its ill-posed nature. The main aim of SISR methods is to generate a high-resolution image from a given low-resolution image. Recently, learning methods of SISR based gained popularity due to advanced convolution neural networks (CNN). These networks have been trained using classical loss functions such as absolute and mean square error, resulting in improper network training due to over-fitting. Thus, a robust and better loss function is essential for reduced over-fitting of the network. The proposed method presents a local variance-based loss function to train CNN with reduced loss. The proposed loss function is proved to be more effective than the existing state of art methods. The effectiveness of the proposed loss function has been evaluated using subjective and objective evaluation metrics.
•A loss function based on the local variance has been presented in this work.•The smoothness of the proposed loss function is increased using L2 regularization.•An end-to-end CNN is trained using the proposed loss function.