计算机科学与探索 ›› 2014, Vol. 8 ›› Issue (12): 1517-1524.DOI: 10.3778/j.issn.1673-9418.1405036

• 人工智能与模式识别 • 上一篇    下一篇

多核学习纹理特征的无参考图像质量评价

严大卫+,桑庆兵   

  1. 江南大学 物联网工程学院,江苏 无锡 214122
  • 出版日期:2014-12-01 发布日期:2014-12-08

Blind Image Quality Assessment with Texture Feature via Multiple Kernel Learning

YAN Dawei+, SANG Qingbing   

  1. School of Internet of Things Engineering, Jiangnan University, Wuxi, Jiangsu 214122, China
  • Online:2014-12-01 Published:2014-12-08

摘要: 由于多种失真类型灰度-共生矩阵特征的不规则性,单核方法无法取得理想结果,从而提出了一种基于多核学习,针对多种失真类型的无参考图像质量评价方法。首先对灰度图像进行相位一致和结构张量变换,得到相位一致图和结构张量图,然后分别对它们提取灰度-梯度共生矩阵二次统计特征,最后将提取的特征输入到高效的分层多核学习机进行训练学习,预测得到图像的质量评分。多图像库多次随机实验结果表明,新方法结果与主观评价值有较好的一致性,并具有较好的推广性。

关键词: 无参考图像质量评价, 多核学习, 灰度-梯度共生矩阵, 结构张量, 相位一致

Abstract: Because of the irregular features extracted from images with various types of distortion, the single kernel method cannot get the ideal result, this paper presents a non-reference image quality evaluation method based on multiple kernel learning for various learning types of distortion. Firstly this paper does a conversion of the gray scale image with the structure tensor and the phase congruency, then extracts the secondary statistical features of gray level-gradient co-occurrence matrix from them, finally inputs these features into hierarchical multiple kernel learning machine for training and gets the quality score. The random experiment results on multiple image library show that the new method results are consistent with the subjective evaluation values, and have better generalization.

Key words: non-reference image quality assessment, multiple kernel learning, gray level-gradient co-occurrence matrix, structure tensor, phase congruency