PA VIF: A passive aggressive visual information fidelity for full reference image quality assessment

摘要

Measurement of visual quality is of significant importance to many image processing tasks. The target of image quality assessment (IQA) is to design effective computational models in order to automatically predict the quality of images in a perceptual consistent manner. We propose a full reference (FR) IQA metric based on information-theoretic IQA framework and passive aggressive learning algorithm. We model the image distortion process with a signal attenuation matrix and an additive noise matrix in wavelet domain. The ridge regression is employed to get the initial attenuation matrix. The passive aggressive algorithm is then utilized in order to learn the particular attenuation matrix of each image block. At last we measure the perceptual quality by quantifying the loss of visual information to the proposed distortion model through an information-theoretic IQA framework. Our proposed metric is learning-free because the learning algorithms are adopted just in order to learn the attenuation matrix of each image block, and no training data is needed in the proposed metric. Experimental results demonstrate that our proposed IQA metric is competitive with existing learning-free IQA metrics; moreover, crossdatabase validation shows that the proposed IQA metric has a satisfying generalization ability.