Hybrid CNN Architectures for Image Super-resolution

2 Proposed Method Whole network architectures are based on recent methods [1] as shown in Fig. 1. Our networks are different from existing methods in block architectures. Figures 2 and 3 show proposed blocks. Following the great success of DPN [2], we combine the residual network and the densely connected network. DBN has two branches in DB blocks to combine the residual network and the densely connected network. Since previous studies showed accuracy improvements, we employ grouped convolutions with 1×1 convolutions to the residual branch. EDPAN is an enhanced DPN [2] for SISR by our observations. We remove 1×1 convolutions of DPN and introduce the weight normalization (WN). Moreover, we add the channel attention (CA) mechanism which rescales feature maps by channel-wise relationships and the significance [5].