Blind image quality assessment via semi-supervised learning and fuzzy inference

Blind image quality assessment (BIQA) is a challenging task due to the difficulties in extracting quality-aware features and modeling the relationship between the image features and the visual quality. Until now, most BIQA metrics available try to extract statistical features based on the natural scene statistics (NSS) and build mapping from the features to the quality score using the supervised machine learning technique based on a large amount of labeled images. Although several promising metrics have been proposed based on the above methodology, there are two drawbacks of these algorithms. First, only the labeled images are adopted for machine learning. However, it has been proved that using unlabeled data in the training stage can improve the learning performance. In addition, these metrics try to learn a direct mapping from the features to the quality score. However, subjective quality evaluation would be rather a fuzzy process than a distinctive one. Equally, human beings tend to evaluate the quality of a given image by first judging the extents it belongs to “excellent,” “good,” “fair,” “bad,” and “poor,” and estimating the quality score subsequently, rather than directly giving an exact subjective quality score. To overcome the aforementioned problems, we propose a semi-supervised and fuzzy framework for blind image quality assessment, S2F2, in this paper. In the proposed framework, (1) we formulate the fuzzy process of subjective quality assessment by using fuzzy inference. Specially, we model the membership relation between the subjective quality score and the truth values it belongs to “excellent,” “good,” “fair,” “bad,” and “poor” using a Gaussian function, respectively; and (2) we introduce the semi-supervised local linear embedding (SS-LLE) to learn the mapping function from the image features to the truth values using both the labeled and unlabeled images. In addition, we extract image features based on NSS since it has led to promising performances for image quality assessment. Experimental results on two benchmarking databases, i.e., the LIVE database II and the TID2008 database, demonstrate the effectiveness and promising performance of the proposed S2F2 algorithm for BIQA.