欧阳甜,江先阳.一种自学习的新型单幅图像超分辨率高质量重建算法[J]. 微电子学与计算机,2024,41(5):1-10. doi: 10.19304/J.ISSN1000-7180.2023.0216
引用本文: 欧阳甜,江先阳.一种自学习的新型单幅图像超分辨率高质量重建算法[J]. 微电子学与计算机,2024,41(5):1-10. doi: 10.19304/J.ISSN1000-7180.2023.0216
OUYANG T,JIANG X Y. A novel super-resolution reconstruction algorithm with high quality based on self-learning for a single image[J]. Microelectronics & Computer,2024,41(5):1-10. doi: 10.19304/J.ISSN1000-7180.2023.0216
Citation: OUYANG T,JIANG X Y. A novel super-resolution reconstruction algorithm with high quality based on self-learning for a single image[J]. Microelectronics & Computer,2024,41(5):1-10. doi: 10.19304/J.ISSN1000-7180.2023.0216

一种自学习的新型单幅图像超分辨率高质量重建算法

A novel super-resolution reconstruction algorithm with high quality based on self-learning for a single image

  • 摘要: 经典的基于稀疏表示和字典学习的超分辨率算法在图像重建质量和计算复杂度上都具备较好的表现。然而,基于外部样本训练得到的字典和待重建图像缺少相关性,会伴随算法鲁棒性较差的问题。为克服这一缺点,提出了一种基于自学习的新型单幅图像超分辨率高质量重建算法。该算法无需引入外部训练图像,即完全通过待重建图像自身构建的样本进行字典学习和图像重建;这一机制增强了训练字典与待重建图像的相关性。具体而言,在字典训练阶段,针对输入的待重建图像,基于二维经验模态分解进行高频修复预处理,以增强样本源的高频特征;随后构建训练样本集,使用K-奇异值分解算法获得自学习主字典和自学习残差字典,构成双字典。在图像重建阶段,将双字典结构与自学习相结合,先通过主字典实现主高频恢复,再进一步通过残差字典恢复图像的残差高频信息。实验结果表明,所提算法在重建图像的主观视觉效果以及专业质量评价指标上,相对于传统插值算法及经典的字典学习算法具有显著优势。

     

    Abstract: Classic super-resolution algorithms based on sparse representation and dictionary learning have good performance according to both the quality of reconstructed images and computational complexity. However, the dictionary trained by external samples is lack of correlation with the image to be reconstructed, which is likely to lead to poor robustness of these algorithms. To overcome this shortcoming, a novel single image super-resolution reconstruction algorithm with high quality based on self-learning is proposed. Rather than using external training pictures, the proposed algorithm can completely carry out dictionary learning and then reconstruct the image through its own samples, which enhances the correlation between the trained dictionary and the image to be reconstructed. Specifically, in the dictionary training stage, for the image to be reconstructed, the high-frequency repair preprocessing is implemented based on bidimensional empirical mode decomposition to enhance the high-frequency characteristics of training samples, the K-SVD (Singular Value Decomposition) algorithm is applied to train both self-learning main dictionary and self-learning residual dictionary to form a dual-dictionary. In the image reconstruction stage, the dual-dictionary structure and self-learning are combined to further recover the high-frequency information of the image through residual dictionary. Experimental results show that the proposed algorithm has significant advantages over traditional interpolation algorithm and classical dictionary learning algorithm in terms of subjective visual effect and the quality of the reconstructed image.

     

/

返回文章
返回