Study On Learning Blind Image Quality Assessment

Objective image quality assessment(IQA) aims at designing computational models such that it can automatically and precisely estimate perceived image quality. Objective IQA metrics can be utilized as the criteria for parameter optimation, performance evaluation, and quality of service(Qo S) monitoring of image processing algorithms or systems. Currently, objective IQA has become an active topic in the field of image processing. Thereinto, blind image quality assessment(BIQA), aims at precisely predicting an arbitrary test image without knowing any information about its undistorted version(i.e., the corresponding reference image). Since it is genreally impossile or difficult to acess the undistorted version of the test image in practical scenarios, BIQA is of great significance.This thesis presents a systematical study on BIQA, which aims at improving the representiveness of image features, the quality prediction precision, and the learning effeciency via machine learning techniques. This thesis focuses on four critical issues about BIQA, i.e. image feature extraction, BIQA model construction, learning framework, and performance verification via subjective study. The relevant work is supported by the National Natural Science Foundation of China(Grant Nos. 61125204, 61432014, and 61172146), Program for Changjiang Scholars and Innovative Research Team in University(No.IRT13088), and Shaanxi Innovative Research Team for Key Science and Technology(No.2012KCT-02). The main content of this dissertation is summarized as follows.The first part focuses on BIQA via multiple kernel learning(MKL). We first propose some new statistical features and then construct two BIQA models by incorporating these features with MKL techniques. In particular, we propose to extract features based on the non-Gaussianity, local dependency, and exponential decay characteristic of natural images, respectively. By analyzing how different distortions affect these statistical properties, we construct new universal BIQA methods using these features and incorporating the heterogeneous property of MKL. Specially, we present two universal BIQA models using the global scheme and the two-step scheme, respectively. Thorough experimental results on standard databases demonstrate that these features are highly correlated with perceived image quality, and that both of the proposed metrics are in remarkably high consistency with human perception.The second part involves the BIQA via active learning. An active feature learning framework is proposed and utilized for image quality prediction. We introduce the methodology of active learning into unsupervised feature learning in order to improve the discriminative ability of the learned image representation. In this framework, patches extracted from training images are normalizated and whitened to obtain the local discriptors. Afterwards, based on the methodology of active learning, a dictionary is constructed by selecting the most informative local discriptors according to both the representativeness and divergency of each local discriptor. Given a test image, the corresponding local discriptors are first calculated and then encoed on the dictionary. Afterwards, all the encoding coefficients are pooled together to represent the test image. Finally, we utilize the learned image representation for quality prediction. Thorough experiments on the standard database illustrate that when the feature vector is of low dimension, the proposed method outperforms the methods based on unsupervised feature learning by 8%.The third part concentrates on BIQA via learning to rank. Inspired by the methodology of learning to rank, we propose a novel framework where BIQA models are learned from the quality preference information between images. State-of-the-art learning based BIQA methods typically require subjects to score a large number of images to train a robust model. However, subjective quality scores are imprecise, biased, and time-consuming, thus the prediction precision and generalization ability of these BIQA methods are limited. To combat these limitations, we present a framework where the preference information between images are exploited for training a robust BIQA model. We first reduced the problem of learning quality prediction function to the pereference learning problem. In particular, we investigate the utilization of MKL and feature fusion to provide a solution. Afterwards, a simple but effective function to estimate perceptual image quality scores is then presented based on the voting strategy. Experimental results show that the proposed BIQA method is highly effective and can be easily extended to new distortion categories.The fourth part puts the emphasis on subjective study of image quality preference. An improved paired comparision method is proposed and adopted in the subjective experimetns for constructing an image quality preference database. To combat the limitations of traditional paired comparison methods, we introduce a slacking strategy in order to improve both the efficiency of subjective study and the reliability of the reported data. Afterwards, we construct a subjective study through constructing experiment evirenment, preparing materials, training observers, and so on. A statistical analysis of the labeled data demonstrates that the proposed paired comparision method makes sense. In addition, we test the performance of BIQA models learned from the labeled data and verify the reliability of this subjective study.In summary, this thesis mainly studies the critical issues of learning BIQA, and improves the precision, effeciency, and generalization ability of learned BIQA models. The research results provide a technical support for the develepment and application of BIQA metrics. In addition, these results show some novel methodologies which may benefit the research of visual quality assessment.