Research On No-reference Image Blur Assessment Methods

Every part of imaging chain, image processing algorithms and transmission has great influence on image quality, and how to assess image quality is one of the basic and challenging problems in the field of image processing. Blur assessment is one kind of image quality assessment, which is used to estimate the perceived sharpness or blurriness of images output by the imaging systems or processing algorithms. Blur assessment has numerous uses in practice, and also plays a central role in shaping many image acquisition, processing, analyzing,autofocus algorithms and systems, as well as their implementation, optimization and testing.The goal of this thesis is to study no-reference blur assessment methods. Based on the characteristics of visual processing of human visual system (HVS) such as hierarchy and self learning, the process of no-reference blur assessment is treated as a matching process between the features of the test image and the simulated common representations of clear or unclear images stored in memory. Four kinds of blur assessment methods based on different levels of features and different reference models are proposed in this thesis, and the main research works are as follows:(1)A novel 3D parameter structure tensor based blur assessment method is proposed.Natural images have statistical properties such as edge domain and high order singularity.The state of the art 2D local structure tensor based blur metrics use eigenvalues of edge patches to measure blurriness, and they are sensitive to noises.Based on the 2D local structure tensor,a novel 3D parameter structure tensor representation of a single image is proposed, which uses the difference between two blurred versions of an image as the third dimension.Eigenvalues of the 3D parameter structure tensor of each image block is computed to be a patch blur measuring parameter, and the weighted sum of the blur measuring parameters and visual attention weights is used to measure the blurriness of an image. The proposed method uses local structures to measure the burriness at the microscopic level, and it is sensitive to blur variation and robust to noises.Performance comparison with the state of the art algorithms shows that the proposed method outperforms the compared methods.Of all evaluation metrics, Pearson linear correlation coefficient and Spearman rank-order correlation coefficient of the proposed method are improved poorly, but they have been increased at least 1.44 and 0.17 percent respectively on average.(2) A novel salient texture features based blur assessment method is proposed.Texture is one of the most important characteristics used in identifying objects or regions of interest in an image. In this thesis, gray level co-occurrence matrix (GLCM) is used to measure the spatial distribution of gray levels in an image. Salient texture features are proposed to be extracted from GLCM of an image patch and used as visual features to measure image blurriness. The key idea of the proposed method is to blur the test image with two smoothing filters of different parameters firstly, and then to extract patch salient features based on GLCM, and finally to compute the blur metric which is modeled as a function of the total salient texture features difference,and the total salient texture features difference is defined as a weighted sum of the salient texture features’ difference of each patch between the two blurred versions of the input image.The proposed method uses statistical distributions of local structures to measure the burriness at the macroscopic level and it correlates well with the HVS, but it is less sensitive to blur variation for using macroscopic scale features.Performance comparison with the state of the art algorithms shows that the proposed method outperforms the compared methods.Of all evaluation metrics,Pearson linear correlation coefficient ,Spearman rank-order correlation coefficient and outlier ratio of the proposed method have been increased at least 1.63,0.40 and 17.5 percent respectively on average.(3) A novel image sparse representation based blur assessment method is proposed.Studies have shown the reason that the HVS can predict the blurriness of an image accurately without any reference is that there is an existing high quality model obtained by long term learning in the human brain.Studies have also shown that natural images can be sparsely coded by unsupervised learning, and their basis functions are similar to the particular shapes of V1 simple-cell receptive fields which are spatially localized, oriented and bandpass.In this thesis, basis learned from clear and sharp images is considered as what the sharp image should be, and sparse coefficients of an image based on the learned basis are used to measure blurriness.Firstly,the dictionary used to sparsely code the image is learnt from clear and sharp image samples by unsupervised learning methods,and then sparse coefficient vectors and visual attention weights of each patch are computed,and finally the blur measure is computed based on the linear superposition of the p-norm of sparse coefficient vectors and visual attention weights of each patch.For considering the self-learning and sparse coding abilitiy of the human visual sytem, the proposed method correlates better with the HVS. Performance comparison with the state of the art algorithms shows that the proposed method outperforms the compared methods.Of all evaluation metrics, Pearson linear correlation coefficient and Spearman rank-order correlation coefficient of the proposed method are improved poorly, but they have been increased at least 1.59 and 0.17 percent respectively on average.(4)A novel image sparse representation and pLSA based blur assessment method is proposed.This method is based on the hypothesis that images have latent characteristics which can be used to measure sharpness, and the fact that human brain can learn unsupervisedly. Probabilistic latent semantic analysis (pLSA) model is used to abstract the sparse codes of images and discover meaningful topics that are latent in large corpora of sharp natural images and the test image, and the similarity between the average topics of sharp images and the topics of the test image is used to measure blurriness. Firstly, the dictionary is constructed from clear sample images;and then clear sample images are sparsly coded and latent topics are extracted by using pLSA; and finally the latent topics of the test image is extracted,and the correlation coefficient between the topics of the test image and the average topics of sharp images is used to measure blurriness. The proposed method uses abstract feaures of sparse codes of an image to measure the blurriness, and incorporate the unsupvised learning ability and hierarchical feature extraction process of the human visual system into the process of blur assessment, therefore, it closely correlates with human perception, but it has higher complexity for the steps of sparse coding and topic extraction.Performance comparison with the state of the art algorithms shows that the proposed method outperforms the compared methods.Of all evaluation metrics, Pearson linear correlation coefficient and Spearman rank-order correlation coefficient of the proposed method are improved poorly, but they have been increased at least 1.71 and 0.54 percent respectively on average.To test and compare the performance of the proposed blur assessment methods and the state of the art algorithms, experiments were conducted on synthetic images and public image quality databases.The same test images and evaluation metrics were used to evaluate the above methods.The experimental results show that the proposed blur assessment methods outperform the compared methods on monotonicity and antinoise ability, and have better or similiar performance on VQEG suggested evaluation metrics.