Efficient Pyramid-Structured Networks Using Channel Separating Gate

Convolutional neural networks (CNNs) have been brought noticeable improvement on numerous image processing applications. In particular, spatial pyramid structures have been widely used, but the number of required parameters increases exponentially as the network deepens. Therefore, we propose a channel separating gate (CSG) that can determine the number of channels in convolutional layers according to the spatial level. For a given spatial level, the CSG separates the channels of feature maps for use at the smaller level. Consequently, the standard non- pooling model can be altered to an efficient pyramid structure by performing the proposed method repeatedly. For single image super-resolution (SR) and deblurring, we demonstrate that our method improves the baseline network in terms of peak signal- to-noise ratio (PSNR) and floating-point operations (FLOPs).