Blind Image Quality Assessment: Exploiting New Evaluation and Design Methodologies

Abstract

The great content diversity of real-world digital images poses a grand challenge to automatically and accurately assess their perceptual quality in a timely manner. In this thesis, we focus on blind image quality assessment (BIQA), which predicts image quality with no access to its pristine quality counterpart. We first establish a large-scale IQA database---the Waterloo Exploration Database. It contains 4,744 pristine natural and 94,880 distorted images, the largest in the IQA field. Instead of collecting subjective opinions for each image, which is extremely difficult, we present three test criteria for evaluating objective BIQA models: pristine/distorted image discriminability test (D-test), listwise ranking consistency test (L-test), and pairwise preference consistency test (P-test). Moreover, we propose a general psychophysical methodology, which we name the group MAximum Differentiation (gMAD) competition method, for comparing computational models of perceptually discriminable quantities. We apply gMAD to the field of IQA and compare 16 objective IQA models of diverse properties. Careful investigations of selected stimuli shed light on how to improve existing models and how to develop next-generation IQA models. The gMAD framework is extensible, allowing future IQA models to be added to the competition.

We explore novel approaches for BIQA from two different perspectives. First, we show that a vast amount of reliable training data in the form of quality-discriminable image pairs (DIPs) can be obtained automatically at low cost. We extend a pairwise learning-to-rank (L2R) algorithm to learn BIQA models from millions of DIPs. Second, we propose a multi-task deep neural network for BIQA. It consists of two sub-networks---a distortion identification network and a quality prediction network---sharing the early layers. In the first stage, we train the distortion identification sub-network, for which large-scale training samples are readily available. In the second stage, starting from the pre-trained early layers and the outputs of the first sub-network, we train the quality prediction sub-network using a variant of stochastic gradient descent. Extensive experiments on four benchmark IQA databases demonstrate the proposed two approaches outperform state-of-the-art BIQA models. The robustness of learned models is also significantly improved as confirmed by the gMAD competition methodology.