Research And Application Of Image Quality Assessment Based On Probabilistic Inference

In image-applied domains such as multi-media,security monitoring,intelligent medicine,remote-sensing detection,deep-space and deep-ocean exploration,image quality is desired to be invulnerable during image acquisition,communication,processing and display systems.This ex-pectation brings the topic of image quality assessment(IQA)which has developed from time-consuming and cumbersome subjective assessment to algorithm-based objective evaluation.In this thesis,we built a new objective quality assessment framework based on probabilistic infer-ence.The framework consists of three paradigms called posterior likelihood,independent pos-terior likelihood,and prior likelihood.It extends the existing framework such that the posterior paradigm combines full-reference and reduced reference together,and the prior paradigm refers to no-reference.More importantly,the independent posterior paradigm is newly introduced by probabilistic inference.To enable IQA algorithms in the proposed framework to be perceptual,we proposed three key IQA techniques:1.In the posterior likelihood paradigm where undistorted images are available for evaluating the relevant distorted ones,a SVD-based IQA algorithm was developed.Compared with the state-of-the-art SVD-based IQA algorithms,the proposed one is superior for identifying an image content-dependent component and an image content-independent one based on a bivalve band-pass singular value curve.Then the two components were modulated by different human visual system(HVS)strategies.As a result,perceptual quality features were extracted,followed by being mapped into objective scores.Evaluation on 7 public-available natural image databases revealed that the proposed algorithm performed much better than state-of-the-art IQA algo-rithms,proving the proposal to be more perceptual.2.In the independent posterior likelihood paradigm where just a group of content-aligned and distortion-type-aligned images are available,we solved the marginal distribution of each image based on constant combined distribution of images' quality,and took the marginal probability as an objective score.This technique has been applied to medical imaging domain for optimizing reconstructive parameters of susceptibility weighted imaging(SWI).As a result,more visually acute SWI images can be selected for diagnosing cerebral stroke disease.3.In the prior likelihood paradigm where only a distorted image is available,a perceptual fea-ture extraction method was proposed based on domain-related prior knowledge.Our proposed method has been successfully applied to IQA for portable fundus photography in circumstance of remote medical diagnosis.After analyzing the prior knowledge composition,we summa-rized three items:luminance and color distortion,acuity,and contrast distortion.Then the feature extraction model was built by three HVS characteristics(multi-channel threshold sen-sation,just noticeable blur,and the contrast sensitivity function).These prior quality features were mapped into objective scores by learning-based classifiers(binary decision tree or sup-port vector machine).To acquire training sets and labels,a subjective assessment experiment was performed.Performances for detecting overall distortion,specified distortions,transferring ability of a trained model,and detecting noise artifacts were evaluated.The results revealed that the ability of detecting pool-quality fundus images achieved a sensitivity of 91.66%and a specificity of 87.45%,which indicated that the proposed method was a practical solution for fundus photography quality control in remote medical diagnosis.All these research works not only deepen the insights into applying probabilistic inference to visual scene processing tasks,but also build relatively concrete framework for mining perceptual quality information.IQA algorithms that are der:ived from this framework perform well on both natural images and medical imaging applications.