Image super-resolution via enhanced multi-scale residual network

Abstract Recently, a very deep convolutional neural network (CNN) has achieved impressive results in image super-resolution (SR). In particular, residual learning techniques are widely used. However, the previously proposed block can only extract one single-level semantic feature maps of single receptive field. Therefore, it is necessary to stack blocks higher-level maps, which will significantly deepen network. While hard train and limits representation for reconstructing hierarchical information. Based on block, we propose an enhanced multi-scale (EMRN) take advantage features via dense connected (EMRBs). Specifically, newly (EMRB) capable constructing multi-level by two-branch inception. The inception our EMRB consists 2 layers 4 each branch respectively, therefore have different ranges fields within EMRB. Meanwhile, local fusion (LFF) used every adaptively fuse extracted Furthermore, global (GFF) EMRN then obtain abundant useful from previous EMRBs subsequent ones holistic manner. Experiments benchmark datasets suggest that performs favorably over state-of-the-art methods further superior (SR) images.