<p>The multi-scale characteristic of natural images is an important cue for various vision tasks. Recently, some deep learning-based methods also investigated the multi-scale features (MSF) for image super-resolution (SR), in which implicit MSF and explicit MSF are two most widely studied categories. However, most existing multi-scale SR networks merely explored one type of MSF, which resulted in insufficient exploration of MSF and encountered inherent inferiority of both implicit and explicit MSF. To mitigate the problem, combining implicit and explicit MSF, we proposed a Multi-Scale and Multi-Dense Network (MSMDN) to leverage complementary superiority of implicit and explicit MSF for image SR. Specifically, two MSF extraction branch are designed in the building block of MSMDN to extract implicit and explicit MSF, respectively. In the implicit MSF extraction branch (iMSFEB), dense connections are utilized to extract implicit MSF, of which the resolution is the same as that of LR images. In the explicit MSF extraction branch (eMSFEB), input features of the building block are projected to spaces of different resolutions for acquisition of explicit MSF. For adequate utilization explicit MSF, dense connections are also adopted in eMSFEB to grasp features from spaces of different resolutions and then to extract features in a space of a new resolution. Besides, we introduce a multi-scale supervision and reconstruction strategy to enforce structurally multi-scale information in eMSFEB as well as to simultaneously super-resolve images with <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times 2\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mo>×</mo> <mn>2</mn> </mrow> </math></EquationSource> </InlineEquation>, <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\times 3\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mo>×</mo> <mn>3</mn> </mrow> </math></EquationSource> </InlineEquation>, and <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\times 4\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mo>×</mo> <mn>4</mn> </mrow> </math></EquationSource> </InlineEquation> factors. Ablation study shows the effectiveness of implicit, explicit MSF, dense connections in eMSFEB, <i>and</i> the multi-scale supervision and reconstruction strategy. Comparison with state-of-the-art methods also indicates the superiority of the proposed network on SR performance.</p>

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

A Multi-scale and Multi-dense Network for Image Super-Resolution

  • Feiqiang Liu,
  • Aiwen Jiang,
  • Beibei Wang,
  • Lihui Chen

摘要

The multi-scale characteristic of natural images is an important cue for various vision tasks. Recently, some deep learning-based methods also investigated the multi-scale features (MSF) for image super-resolution (SR), in which implicit MSF and explicit MSF are two most widely studied categories. However, most existing multi-scale SR networks merely explored one type of MSF, which resulted in insufficient exploration of MSF and encountered inherent inferiority of both implicit and explicit MSF. To mitigate the problem, combining implicit and explicit MSF, we proposed a Multi-Scale and Multi-Dense Network (MSMDN) to leverage complementary superiority of implicit and explicit MSF for image SR. Specifically, two MSF extraction branch are designed in the building block of MSMDN to extract implicit and explicit MSF, respectively. In the implicit MSF extraction branch (iMSFEB), dense connections are utilized to extract implicit MSF, of which the resolution is the same as that of LR images. In the explicit MSF extraction branch (eMSFEB), input features of the building block are projected to spaces of different resolutions for acquisition of explicit MSF. For adequate utilization explicit MSF, dense connections are also adopted in eMSFEB to grasp features from spaces of different resolutions and then to extract features in a space of a new resolution. Besides, we introduce a multi-scale supervision and reconstruction strategy to enforce structurally multi-scale information in eMSFEB as well as to simultaneously super-resolve images with \(\times 2\) × 2 , \(\times 3\) × 3 , and \(\times 4\) × 4 factors. Ablation study shows the effectiveness of implicit, explicit MSF, dense connections in eMSFEB, and the multi-scale supervision and reconstruction strategy. Comparison with state-of-the-art methods also indicates the superiority of the proposed network on SR performance.