Super‐resolution using deep residual network with spectral normalization

Abstract

Abstract In this letter, the authors present a single‐image super‐resolution method based on introducing a novel spectral normalization to the convolution of a deep residual network. Moreover, the authors construct a new residual block (RB) and assemble it in a cascade form. The new RB was restructured by the spectrally normalized convolution layers and activated function. In addition, the RB allows the spectral normalization introduced on the trained network to update the additional weights. Furthermore, it minimizes pixel loss, thereby helping to obtain enhanced reconstruction results, limiting the number of parameters, and facilitating a low computational cost. The experimental results demonstrate that the proposed model shows the superior performance over that of state‐of‐the‐art methods in terms of visual quality metrics such as UQI and PIQE.

Keywords

• artificial intelligence
• computer vision
• image processing