Image Quality Assessment Based On Human Visual Perception

One picture is worth ten thousand words.With the rapid development of the Internet,digital images and videos are gradually becoming the main carriers of information for their intuitive form of expression.As the final receiver of various electronic equipments,consumers always wish to watch high definition and high resolution images.Unfortunately,the images are inevitable subjected to a mass of noises during the stages of acquisition and processing in practical situations.As a result,the images are distorted and the visual experience of consumers are impaired.To measure the image distortions,researchers put forward the image quality assessment(IQA)metrics to automatically evaluate the image quality for computers.Based on the accurate evaluation of image quality,researchers can better optimize and guide the image processing algorithms and systems,which then further modify the image quality.Therefore,an effective IQA method is crucial to improve the visual quality of experience(Qo E)and even the quality of life.According to the amount of information obtained from the reference image,the existing IQA metrics are generally classified into three types: full-reference(FR),reduced-reference(RR),and blind/no-reference(NR).This dissertation does not study the FR IQA metrics since they require the whole access to the reference image and can easily achieve excellent performance with low computation complexity.Due to the lack of enough reference information,it is difficult to design the RR and NR IQA metrics.There is still much space to improve the performance of RR and NR metrics,which are our research focus.Through analyzing the perceptual characteristics of human visual system(HVS)and exploring the effect of distortion types and intensities on image quality,this dissertation studies the evaluation metrics from RR to NR based on the required reference information.Aiming at the disadvantages of the existing IQA metrics,we make certain improvements.The main contributions are summarized as follows.(1)A RR IQA metric based on the entropy differences in discrete Cosine transform(DCT)domain is proposed.The existing RR IQA metrics usually confront the contradiction between the data rate from reference image and the evaluation performance.The metrics may perform excellently with a large amount of reference data,while the performance will decline as the reference data reduces.For this problem,the authors propose a RR IQA metric based on the entropy differences in DCT domain.The first step is to decompose the distorted and reference images by the block-based DCT.The DCT coefficients are then reorganized and merged according to the frequency bands.The second step is to compute the entropies of coefficients with different bands for distorted and reference images,respectively.The third step is to calculate the differences between entropies of distorted and reference images at each band.The quality of distorted image is finally obtained by the weighted sum of entropy differences based on the sensitivity of human eyes on different frequency information.Through measuring the distortion at each frequency band individually and considering the visual property of human eyes,the proposed RR IQA method is able to achieve good performance with only eight reference data.(2)A NR IQA metric with improved natural scene statistics(NSS)model is proposed.Recently,the NSS-based NR IQA metrics assume the distorted images fit the natural statistical properties as the natural images do.And the distortions are measured based on the variations of model parameters.However,the distorted images may not fit the NSS model well when the distortions are serious.Thus,the fitting error between the real distribution and fitted model appears,which causes loss of evaluation accuracy.To solve this problem,the authors propose a NR IQA metric with improved NSS model.The proposed frame fits the statistical distribution of images with NSS models.And the fitting parameters,fitting errors,and likelihood probabilities of the models are computed as image features.Given the training set with pre-known subjective quality scores,the relationships between image features and perceptual qualities as well as the distortion types are learned using the support vector machine(SVM).For the test image,its features are extracted and the quality is predicted based on the training model.The proposed method analyzes the effect of fitting error on image statistical properties and overcome the inferiority of the existing NSS-based NR IQA metrics.The performance of the proposed metric correlates well with human visual perception.(3)A bag-of-words feature representation for NR IQA metric with local quantized pattern is proposed.Many NR IQA metrics usually extract image structural information as image features.However,the existing feature representation methods can not reflect the spatial relation well,and the distortion intensity may be evaluated inaccurately.Considering this problem,the authors propose a bag-of-words feature representation for NR IQA metric with local quantized pattern.Firstly,given a natural image set,the local quantized pattern of all images are calculated and the visual dictionary is constructed using the bag-of-words model.Then,given the training set containing distorted images with the pre-known subjective quality scores,the features of each distorted image are computed based on the visual dictionary.Next,the mapping from the image features to subjective quality scores is learned via support vector regression.Lastly,the feature histogram of test image is extracted and the quality is predicted based on the regression model.The proposed method utilizes the local structural information and spatial relationship for feature extraction,which can effectively reflect the image distortion intensity.The proposed metric performs highly consistent with the human perception.