Study On No-Reference Quality Assessment Of Deblurred Images

Blur is one of the most common distortion types in the acquisition,transmission and processing of digital images.Image deblurring is an effective technique to restore a latent clean image from its blurred version.In the past decades,image deblurring has been extensively studied and significant achievements have been made.However,very little work has been dedicated to the quality assessment of deblurred images,which may hinder further development of more advanced image deblurring techniques.To study effective image quality assessment for deblurred images is in an urgent need for this purpose.In this thesis,we conduct both subjective and objective studies for defocus deblurred images.For objective assessment,two no-reference(NR)quality metrics for defocus deblurred images are proposed.The main contributions of this thesis are as follows.1.Image database is of great importance for the study of image quality evaluation methods.To explore how the human subjectively perceive the quality of the deblurred images,there must be a deblurred image database.To this end,a defocus deblurred image database(DDID)is first built using 8 state-of-the-art image defocus deblurring algorithms by processing 30 blurred images with different blur levels,producing 240 deblurred images.Then subjective experiment is conducted to collect human quality ratings.Finally,a systematic evaluation of performance rankings of the deblurring algorithms is conducted with the subjective test results.This sduty has very important guiding significance for the selection of image deblurring algorithms.2.In addition to residual blurring,deblurring algorithms are very likely to incur additional distortion of texture irregularity in the deblurred images.It has been proven that gray level co-occurrence matrix(GLCM)can effectively capture the texture characteristics.Motivated by these observations,a unified framework is proposed for NR defocus deblurred image quality evaluation based on GLCM.The framework consists of a texture irregularity module and a blur evaluation module.For the texture irregularity module,a preprocessing step is first conducted by computing the mean subtracted contrast normalized(MSCN)coefficients.Then the GLCM is computed,from which two statistics(entropy and homogeneity)are further calculated as effective texture features to measure the texture irregularity.The blur evaluation module can be achieved using the existing NR image blur/sharpness metrics or general-purpose NR image quality metrics.Finally,the overall quality score of a defocus deblurred image is obtained by linear weighting.Experimental results demonstrate that the proposed metric can evaluate the deblurred image quality accurately and generate quality scores highly consistent with humanperception.It can also be used to improve the performance of the existing NR image quality metrics.3.It has been proven that natural scene statistics(NSS)are excellent indicators of the degree of quality degradation in distorted images.Since image deblurring is a typical ill-posed inverse problem,it is common that the deblurred images exhibit unnaturalness.Motivated by this,a NR quality metric is proposed for defocus deblured images based on NSS.We propose to evaluate the quality of defocus deblurred images by capturing effective NSS features from both global and local perspectives.This is consistent with the working mechanism of the human visual system(HVS).Specifically,the spatial domain NSS features are naturalness factors of gradient distribution prior to characterize the global naturalness,and the frequency domain NSS features are distributions of log-Gabor filters response coefficients to portray the local structural distortions in different scales and orientations.All features are combined to train a support vector regression(SVR)model for quality prediction of defocus deblurred images.The experimental results demonstrate that the proposed metric can evaluate the deblurred image quality accurately and generate quality scores highly consistent with human perception.It also outperforms the relevant state-of-the-arts.As an application,the proposed metric is further used for benchmarking deblurring algorithms and very encouraging results are achieved.