Journal of Computer Applications ›› 2017, Vol. 37 ›› Issue (9): 2636-2642.DOI: 10.11772/j.issn.1001-9081.2017.09.2636

Previous Articles     Next Articles

Enhanced self-learning super-resolution approach for single image

HUANG Feng1, WANG Xiaoming1,2   

  1. 1. School of Computer and Software Engineering, Xihua University, Chengdu Sichuan 610039, China;
    2. Robotics Research Center of Xihua University, Chengdu Sichuan 610039, China
  • Received:2017-03-30 Revised:2017-06-07 Online:2017-09-10 Published:2017-09-13
  • Supported by:
    This work is partially supported by the National Natural Science Foundation of China (61532009), the Chunhui Program of Ministry of Education (Z2015102), the Key Scientific Research Project of Sichuan Provincial Department of Education (11ZA004).

增强的单幅图像自学习超分辨方法

黄凤1, 王晓明1,2   

  1. 1. 西华大学 计算机与软件工程学院, 成都 610039;
    2. 西华大学机器人研究中心, 成都 610039
  • 通讯作者: 王晓明,wxmwm@aliyun.com
  • 作者简介:黄凤(1993-),女,四川南部人,硕士研究生,主要研究方向:图像超分辨率重建;王晓明(1977-),男,四川简阳人,副教授,博士,主要研究方向:模式识别、机器学习、图像处理、计算机视觉。
  • 基金资助:
    国家自然科学基金资助项目(61532009);教育部春晖计划项目(Z2015102);四川省教育厅自然科学重点项目(11ZA004)。

Abstract: Aiming at the main problem of the Sparse Representation (SR) coefficients of the image blocks in image super-resolution method, an enhanced self-learning super-resolution approach for single image was proposed by using the weighting idea. Firstly, the pyramid of high and low resolution images was established by self-learning. Then, the image block feature of low-resolution images and the central pixels of the corresponding high-resolution image blocks were extracted respectively. The role of the center pixel in constructing the image block sparse coefficient was emphasized by giving different weights of different pixels in the image blocks. Finally, the combination of SR theory and Support Vector Regression (SVR) technique was used to build the super-resolution image reconstruction model. The experimental results show that compared with the Self-Learning Super-Resolution method for single image (SLSR), the Peak Signal-to-Noise Ratio (PSNR) of the proposed method is increased by an average of 0.39 dB, the BRISQUE (Blind/Reference-less Image Spatial Quality Evaluator) score of no-reference image quality evaluation criteria is reduced by an average of 9.7. From the subjective perspective and objective values, it is proved that the proposed super resolution method is more effective.

Key words: digital image processing, single image super-resolution, Sparse Representation (SR), Support Vector Regression (SVR), weight coefficient

摘要: 针对图像超分辨率方法构建图像块的稀疏表示(SR)系数存在的主要问题,利用加权思想提出一种增强的单幅图像自学习超分辨方法。首先,通过自学习建立高低分辨率图像金字塔;然后,分别提取低分辨率图像的图像块特征和对应高分辨率图像块的中心像素,并给图像块中不同像素点赋予不同的权重,强调中心像素点在构建图像块稀疏系数时的作用;最后,结合SR理论和支持向量回归(SVR)技术建立超分辨率图像重建模型。实验结果表明,与单幅图像自学习超分辨率方法(SLSR)相比,所提方法的峰值信噪比(PSNR)平均提高了0.39 dB,无参考图像质量评价标准(BRISQUE)分数平均降低了9.7。从主观视角和客观数值证明了所提超分辨率方法更有效。

关键词: 数字图像处理, 单幅图像超分辨率, 稀疏表达, 支持向量回归, 权重系数

CLC Number: