A Study on Super Resolution Method using Encoder-Decoder with Residual Learning

DongWoo Lee , SangHun Lee , HyunHo Han

In this paper, image Super Resolution method using Encoder-Decoder network structure with Residual Learning was proposed. The CNN(Convolutional Neural Network) based Super Resolution method using the existing Bicubic Interpolation uses convolution with a fixed output size of the feature map to extract unnecessary and distorted features, resulting in blurred or distorted contours. We used Encoder step that extracts features to solve these problems, and Encoder-Decoder structures that performs Super Resolution in Decoder step so as to restore image of the extracted feature map. The Residual Block process was configured between Encoder and Decoder to delete unnecessary information from the feature map extracted by Encoder and emphasize high frequency characteristics. In addition, more efficient learning was through Residual Learning using Skip Connection and Long Skip Connection. Experiments had improved contour distortion and blurring over existing BicubicInterpolation and VDSR(Very Deep Super Resolution) methods.

Volume 11 | 05-Special Issue

Pages: 2426-2433