Densely Multi-path Network for Single Image Super-Resolution.

摘要: Recently, deep convolutional neural networks (CNNs) make many breakthroughs in accuracy and speed for single image super-resolution (SISR). However, we observe that the fusion of information on different receptive fields have not been fully exploited current SR methods. In this paper, propose a novel densely multi-path network (DMPN) SISR introduces blocks (DMPBs). A DMPB contains several subnets (MPSs) with dense skip connections, concatenates outputs MPSs are fed into next DMPB. MPS uses convolution kernels sizes each path, exchanges through cross-path connections. Such strategy allows to full use levels better adapt extracting high-frequency features. Quantitative qualitative experimental results indicate effectiveness proposed DMPN, which achieves restoration performance visual effects than state-of-the-art algorithms.