Single Image Super Resolution via a Refined Densely Connected Inception Network

Single image super resolution has achieved a significant breakthrough with the development of deep learning technology. Among these approaches based on deep learning, the mainstream method is to build a cascading network and attempt to add more learning layers. However, as the depth of the model increases, features far away from the reconstruction layer are less considered in the reconstruction process. In this paper, we propose a novel model based on a refined densely connected network for super-resolution reconstruction tasks. By utilizing densely connected paths in the model, we can significantly shorten the distance between the feature maps from different levels and the reconstruction layer. Besides, an inception-like structure is employed to replace the ordinary convolutional layer to take full advantage of the contextual information. Moreover, quantities of 1 ×1 filters are used to ensure an acceptable model size. Extensive experiments are conducted for demonstrating that the proposed method can achieve the state-of-the-art performance with smaller model size.

1. Introduction

With the popularity of deep learning, a lot of computer vision tasks are undergoing an unprecedented innovation. Single Image Super Resolution (SISR) is a classic but tricky problem on image reconstruction, which plays an important role in many industry fields such as medical imaging, satellite imaging, security and surveillance. It refers to a task which concentrates on recovering a high resolution (HR) image from its degraded image [1]. Considering that the single image can only provide limited context information for image recovering, it is a typical ill-posed inverse problem.