Improved Algorithms for Zero Shot Image Super-Resolution with Parametric Rectifiers

Abstract

Recently, a novel Zero-Shot Super-Resolution (ZSSR) method is proposed to generate high-resolution (HR) images from their low-resolution (LR) counterparts. ZSSR employs a convolutional neural network (CNN) to represent transformations from LR images to HR images and is trained on a single image. ZSSR achieves state-of-the-art performance on both real low-resolution images (i.e., historic images, and images taken with a mobile phone) and several benchmark datasets (e.g., Set 5 and Set 14 to name a few). However, the training of the CNN network of ZSSR is not stable since rectifier is used as the activation function and a custom learning rate adjustment policy is proposed in ZSSR. In this paper, we use parametric rectifier as the activation function and present an improved algorithm for the training of ZSSR. Experimental results demonstrate that the proposed method outperforms ZSSR in terms of both reconstruction accuracy and speed on two benchmark datasets: Set 5 and Set 14, respectively.