Image Super-resolution by Residual Attention Network with Multi-skip Connection

Abstract

Deep convolutional neural networks (Deep CNNs) are difficult to train as they become deeper.Moreover,in image super-resolution,channel-wise features and inputs of the low-resolution (LR) image are treated equally between different channels,resulting in the deficiency of the representational ability of the CNNs.To resolve these issues,residual attention network with multi-skip Connection (RANMC) is proposed for single-image super resolution (SISR),which employs residual in multi-skip connection (RIMC) structure,then a very deep network is formulated with serval residual groups.Each residual group (RG) contains a certain number of short skip connections (SSC) and multi-skip connections (MC).Based on RIMC,rich low-frequency (LF) information is allowed to be bypassed through multi-skip connection,and high-frequency (HF) information is focused on learning by the principal network.Furthermore,considering interdependencies in channel and spatial dimension,attention mechanism block(AMBlock) is proposed to focus on the location of the information and adaptively readjust channel-wise features,where the spatial attention (SA) mechanism and channel attention (CA) mechanism are taken in the approach.Experiments indicate that RANMC can not only recover image details better,but also obtain higher image quality and network performance.

Keywords

• image super-resolution|attention mechanism block|residual network|residual in multi-skip connection|skip connection