Lightweight Super-resolution Learning Model for Extremely Exposed Images

Video surveillance system adopting wireless sensor network (WSN) becomes more and more popular. To achieve energy efficiency and low transmitting bandwidth, low-cost and low-resolution video camera may be used. However, captured image/video with low resolution may cause information loss; for example, suspicious objects such as a bomb, and emergent events such as fire emergency. Moreover, it is getting deteriorated in case an extremely exposed scene is presented. In this paper, a lightweight learning-based super-resolution (LLBSR) image reconstruction algorithm is proposed for the control center of surveillance system to recover information details from low-resolution images with extremely exposed scenes. The captured video sequences were processed via a simplified difference residual network (DRN) to improve contrast first. Then the pre-processed video sequences were scaled up via a lightweight SR neural network (LSRNN). Experimental results show that the proposed algorithm can achieve a comparable PSNR performance using a simple neural network as compared with a famous prior work with very deep neural network