Research On Objective Quality Assessment Metrics For Image Restoration

Recently,image restoration technology has attracted much attention for its wide application in many areas,such as remote sensing imaging,outdoor monitoring and public security,etc.A large number of image restoration algorithms have been proposed for restoring latent clear images from degraded images.However,there is no image restoration algorithm that works for all visual contents.The performance of restoration algorithms and whether the technology can be successfully applied in the real-world environment are directly determined by the quality of restored images.Therefore,it is of great significance to study the objective quality evaluation for image restoration.Based on the summary and analysis of the research actuality and development trends,this thesis makes an in-depth study of the key problems in this field and proposes the corresponding solutions.Specifically,this paper mainly focuses on the study of the quality assessment of restorted image and the performance evaluation of image restoration algorithms.The research achievements of this thesis can be used for automatic parameter selection of image restoration algorithms,selection of image restoration algorithms and guiding the design of the image restoration algorithms,which have important theoretical research significance and application value.The main work and contributions of this thesis are as follows:(1)A quality metric of Compressive Sensing(CS)recovered image based on the measurement of local and global distortions is proposed by exploring the effect of the distortions introduced during the CS image recovery process on the CS recovered image quality.Considering that there is no public database available for the performance testing of quality metrics,a CS Recovered Image Database(CSRID)is first built,where ten reference images and ten state-of-the-art CS image recovery algorithms with three different sensing rates are used to build the database.Then the subjective quality scores of the CS recovered images are obtained by the subjective experiment.The CS recovered images are typically contaminated by multiple distortions,particularly at low sampling rates.However,most of the existing image quality metrics cannot effectively evaluate multiple distortions,so they are limited in predicting the quality of CS recovered images.Based on the above analysis,this paper presents a quality metric based on the measurement of local and global distortions.The local features consist of a local phase coherence based edge sharpness measure and a gray level co-occurrence matrix based texture measure.Two kinds of features based on natural scene statistics are extracted to describe the global distortions.One type is obtained by calculating image naturalness parameters,and the other type is computed based on the statistics of singular value decomposition coefficients.Finally,support vector regression is employed to do model training and the subsequent quality prediction.Experimental results conducted on the CSRID database demonstrate the advantages of the proposed method.As an application,the proposed metric is used for automatic parameter selection for CS image recovery algorithms.(2)A quality metric of motion-deblurred images based on the measurement of noise,ringing and residual blur is proposed by exploring the multi-distortion characteristics in the motion-deblurred images.Motion deblurring usually introduces the ringing artefacts due to the inaccurate point spread functions estimation.Meanwhile,it is difficult to completely removed motion blurring,resulting in residual motion blurring in the motion-deblurred images.However,most of the existing image quality metrics are designed for the traditional distortions,which are difficult to measure ringing effect effectively.Therefore,they cannot effectively evaluate the quality of motion-deblurred images.Based on these,this paper presents a quality metric by measuring noise,ringing and residual blur.For the noise,we measure it separately in three channels and then calculate their average value.For the ringing effect,an evaluation method based on visual saliency is proposed by taking the motion-blur image as a reference.A reblurring-based blur assessment method is proposed for evaluating the residual blur.Finally,the overall quality score of a motion-deblurred image is obtained by pooling the scores of noise,ringing and blur.Experimental results conducted on a motion deblurring database demonstrate that the proposed metric significantly outperforms the existing quality metrics.In addition,the experimental results show that the proposed metric has an excellent performance in improving the performance of the existing general-purpose no-reference quality metrics.(3)A general no-reference framework for evaluating the performances of image restoration algorithms is proposed by exploring the relative quality ranking of restored images.The metrics proposed by work(1)and(2)are both designed for specific restoration application,so their generalization ability is limited.For the performance evaluation of image restoration algorithms,the ranking of restored images that are generated via various algorithms is the most heavily considered factor.Inspired by this,this paper presents a pairwise-comparison-based rank learning framework for the performance evaluation of image restoration algorithms.Under the proposed framework,this paper further proposes a general image restoration quality metric by integrating quality-aware features in both the spatial and frequency domains.The proposed metric exhibits good generalization performance and it is applicable to various restoration applications.The results of extensive experiments that are conducted on eight public databases of five restoration scenarios demonstrate the superior performance of the proposed method over the existing quality metrics.Moreover,the proposed framework can be used to improve the performance of the existing quality metrics in the performance evaluation task of image restoration algorithms.(4)A general image restoration quality metric based on convolutional neural network is proposed by exploring the relative quality difference between the Degraded Input and the corresponding Restored Image(DIRI).The metric proposed in work(3)establishes the model based on the quality ranking relation among restored images,and it cannot predict the relative quality difference between the DIRI pair,which is another index of the performance evaluation for image restoration algorithms.In addition,the method proposed in work(3)is designed based on hand-crafted features,so it is difficult to describe image restoration distortions comprehensively.For the performance evaluation of image restoration algorithms,quantifying the relative quality difference of the DIRI pair is one of the most reliable ways.Inspired by this,we propose a Siamese-network-based relative Quality Difference Learning(SQDL)model.Firstly,the DIRI pairs are constructed from the individual degraded inputs and the corresponding restored images.Then the DIRI pairs and the associated quality labels are used to train the end-to-end SQDL model.Finally,the performance ranking of these algorithms is determined by the predicted relative quality differences directly.To address the problem of the limited training data,a large-scale image quality database is built without laborious human labeling and then used to pre-train the SQDL model.The experimental results on four public databases of two restoration scenarios show that the proposed metric is superior to the state-of-the-arts.Moreover,the cross-database tests demonstrate the superiority of the proposed metric in terms of the generalization ability.This thesis consists of 42 images,43 tables and 198 references.