【24h】

SUPER-RESOLUTION OF TEXT IMAGES THROUGH NEIGHBOR EMBEDDING

机译:通过近邻嵌入对文本图像进行超分辨率

获取原文
获取原文并翻译 | 示例

摘要

Single-image super-resolution (SISR) is the problem ofrngenerating a high resolution (HR) image from a single lowrnresolution (LR) image, possibly with the help of a set ofrntraining images. The SISR technique known as neighborrnembedding (NE) is based on the assumption that correspondingrnsmall patches in low and high resolution versionsrnof an image form manifolds with similar localrngeometry. NE utilizes a training ensemble of pairs of lowrnand high resolution image patches, where the patches in arngiven pair represent the same image region. An input patchrnfrom a LR image is approximated by nearby LR trainingrnpatches and a HR patch estimate is constructed from therncorresponding combination of HR training patches. WhilernNE has shown good success for super-resolution of facernand scene imagery, little has been reported on NE forrnenhancing text images. We apply NE to enhance LR textrnimages, and achieve good HR estimates at 2x, 3x and 4xrnmagnification. Our experiments show that NE raisesrnPSNR found with bicubic interpolation (BI) by 68% andrn89% at 3x, and 4x magnification, respectively. We showrnhow to naturally extend the original NE luminance featuresrnto an arbitrary number, and achieve further improvementrnin PSNR by adding just one more feature.
机译:单图像超分辨率(SISR)是从单个低分辨率(LR)图像中生成高分辨率(HR)图像的问题,可能需要借助一组训练图像。 SISR技术被称为邻居嵌入(NE),是基于这样的假设:低分辨率和高分辨率版本中相应的小补丁会形成具有相似局部几何形状的图像的流形。 NE利用成对的低分辨率和高分辨率图像块的训练合奏,其中安吉文对中的块代表相同的图像区域。来自LR图像的输入补丁通过附近的LR训练补丁来近似,并且HR补丁估计是根据HR训练补丁的相应组合来构造的。虽然NE在面部和场景图像的超分辨率方面显示出了巨大的成功,但关于NE增强文本图像的报道很少。我们应用NE来增强LR文本图像,并以2倍,3倍和4倍放大率获得良好的HR估计。我们的实验表明,NE将使用三次三次插值(BI)发现的rnPSNR分别放大了3倍和4倍,分别提高了68%和rn89%。我们展示了如何自然地将原始NE亮度特征扩展到任意数量,并通过仅添加一个特征来进一步提高PSNR。

著录项

相似文献

  • 外文文献
  • 中文文献
  • 专利
获取原文

客服邮箱:kefu@zhangqiaokeyan.com

京公网安备:11010802029741号 ICP备案号:京ICP备15016152号-6 六维联合信息科技 (北京) 有限公司©版权所有
  • 客服微信

  • 服务号