结合稀疏表示和概率潜在语义的图像模糊度评价

目的 图像的模糊度评价是客观图像质量评价的一种,主要用来衡量图像信号经过成像系统或处理算法后的降质程度,其在图像获取、传输、分析以及图像处理系统或算法评价等领域有着广泛的应用。针对目前图像模糊度评价方法没有考虑人类视觉系统的无监督学习和层次化特征提取的特性,本文将图像稀疏表示和利用概率潜在语义提取图像主题相结合,提出基于稀疏表示和概率潜在语义的图像模糊度评价算法。方法 该算法在图像稀疏表示的基础上,通过概率潜在语义方法分别提取清晰训练图像和待测图像的主题,以待测图像潜在主题与清晰图像平均潜在主题之间的相似性作为模糊度评价的依据。主要过程分为3个阶段:词典构建阶段、训练学习阶段和模糊度评价阶段。词典构建阶段的目的是通过样本学习获得图像稀疏表示的词典;训练学习阶段的目的是采用概率潜在语义的方法获得训练图像的平均主题;模糊度评价阶段的目的是通过待测图像的潜在主题与训练图像的平均潜在主题的相关系数来计算图像的模糊程度。结果 在仿真图和公共测试数据库上与典型算法的比较实验表明:本文算法在单调性、抗噪性以及视频质量专家组制定的5个评价指标上都取得了较好的效果,其中Pearson相关系数和Spearman秩相关系数分别为0.995 6和0.993 4。结论 本文根据人类视觉系统具有无监督学习和层次化特征提取的特点,以待测图像潜在主题与清晰图像平均潜在主题之间的相似性作为模糊度评价的依据,提出了一种新的基于稀疏表示和概率潜在语义的图像模糊度评价方法。实验结果表明该方法能够对图像的模糊度进行较准确的评价,并且结果与人的主观评价结果一致。

Image blur evaluation based on sparse representation and probabilistic latent semantic analysis

Objective Blur evaluation is an image quality assessment process used to estimate the perceived sharpness or blurriness of images that imaging systems or processing algorithms output. Blur evaluation has numerous uses in practical applications and plays a central role in shaping image acquisition, transmission, analysis algorithms and systems, as well as the implementation, optimization, and testing of these algorithms and systems. An image sparse representation and probabilistic latent semantic analysis (pLSA)-based blur evaluation method is proposed to incorporate the unsupervised learning characteristic and the hierarchical feature abstraction model of the human visual system to the blur evaluation process. Method The proposed method is based on the hypothesis that images possess latent characteristics that can be used to measure image quality and the fact that the human brain can learn in an unsupervised manner. The pLSA model is used to identify meaningful topics that are latent in the sparse codes of natural images and the test image. The similarity of the latent topics between training images and the test image is used to measure blurriness. The proposed method has three crucial stages, which are the dictionary construction, learning, and blur metric computation stages. The dictionary construction stage develops the dictionary from clear sample images. The learning stage extracts awerage topics from the sparse codes of clear training images by using pLSA. The blur metric computation stage calculates the blur metric according to the correlation coefficient between the latent topics of the test image and the awerage topics of the clear training images. Result Experimental results on the synthetic images and public image quality databases show that the proposed method exhibits better performance than the state-of-the-art blur metrics in terms of monotonicity, anti-noise capability, and the suggested evaluation metrics of the Video Quality Experts Group. The Pearson correlation and Spearman rank-order correlation coefficients are approximately 0.995 6 and 0.993 4, respectively. Conclusion Based on the unsupervised learning characteristic and the hierarchical feature abstraction model of the human visual system, this study proposes a novel blur evaluation method. This method uses the similarity of latent topics between training images and the test image to measure blurriness. Experimental results show that the proposed blur evaluation method can evaluate the amount of blurriness in images with high accuracy, and it correlates well with the human visual system.