Ghost Residual Networks for Image Companding

Image companding is a technique to convert a High Definition Resolution (HDR) image to Low Definition Resolution (LDR) image and convert it back to HDR image again. With the development of a convolutional neural network, the Ghost Substraction Additional Network (GSANet) is proposed as a method that can be used to perform image expanding or convert LDR image to HDR image. GSANet is created by utilizing the concept of residual learning to maintain its performance when the network deepens. In addition, ghost modules are used in GSANet, in order to produce a smaller number of parameters and computational costs. GSANet produces an average PSNR and average SSIM of 32.91 and 0.923 respectively which are better than the previous method.