eduzhai > Applied Sciences > Engineering >

OverNet Lightweight Multi-Scale Super-Resolution with Overscaling Network

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 10 pages

Abstract: Super-resolution (SR) has achieved great success due to the development ofdeep convolutional neural networks (CNNs). However, as the depth and width ofthe networks increase, CNN-based SR methods have been faced with the challengeof computational complexity in practice. Moreover, most of them train adedicated model for each target resolution, losing generality and increasingmemory requirements. To address these limitations we introduce OverNet, a deepbut lightweight convolutional network to solve SISR at arbitrary scale factorswith a single model. We make the following contributions: first, we introduce alightweight recursive feature extractor that enforces efficient reuse ofinformation through a novel recursive structure of skip and dense connections.Second, to maximize the performance of the feature extractor we propose areconstruction module that generates accurate high-resolution images fromoverscaled feature maps and can be independently used to improve existingarchitectures. Third, we introduce a multi-scale loss function to achievegeneralization across scales. Through extensive experiments, we demonstratethat our network outperforms previous state-of-the-art results in standardbenchmarks while using fewer parameters than previous approaches.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×