计算机科学与探索 ›› 2017, Vol. 11 ›› Issue (1): 144-154.DOI: 10.3778/j.issn.1673-9418.1512082

• 人工智能与模式识别 • 上一篇    下一篇

稀疏表示的无参考图像质量评价方法

桑庆兵+,程大宇   

  1. 江南大学 物联网工程学院,江苏 无锡 214122
  • 出版日期:2017-01-01 发布日期:2017-01-10

Blind Image Quality Assessment via Sparse Representation

SANG Qingbing+, CHENG Dayu   

  1. School of Internet of Things Engineering, Jiangnan University, Wuxi, Jiangsu 214122, China
  • Online:2017-01-01 Published:2017-01-10

摘要: 现有的无参考图像质量评价算法多采用支持向量回归、神经网络等作为映射,训练过程需要大量样本,且泛化性能差(即在一个数据集上的训练识别效果好,在另一个数据集上可能很差),从而提出了基于稀疏表示的无参考图像质量评价算法。利用梯度幅值与拉普拉斯变换图像的联合统计信息和小波变换子带相关性组成特征字典,并对测试图像特征进行稀疏表示,最后综合稀疏系数与字典图像DMOS值获得预测质量得分。多数据库中大量实验结果表明,新算法在少量训练样本条件下即可获得优良而稳定的结果,且具有更好的泛化性能和稳定性。

关键词: 无参考图像质量评价, 稀疏表示, 统计特征, 小波交换

Abstract: Existing blind image quality assessment (BIQA) algorithms usually build a mapping model by support vector regression (SVR) or neural network. These algorithms need a large number of samples; they are weak in generalization (perform well when training and testing on the same database, but poorly on different one). In order to overcome these deficiencies, this paper proposes a BIQA algorithm based on sparse representation. It utilizes the joint statistics information of gradient magnitude (GM) map and Laplacian of Gaussian (LOG) response and the correlation in the wavelet subbands to compose a feature dictionary, then represents the feature of a test image via sparse coding, finally estimates the differential mean opinion scores (DMOS) of dictionary images and sparse coefficient to get the predicted quality scores. The experimental results on multiple databases show that, compared with other algorithms, the proposed algorithm performs more accurately and stably under low ratio of training samples, and possesses better generalization ability and stability.

Key words:  blind image quality assessment, sparse representation, statistic feature, wavelet transform