Research On Image Objective Quality Assessment Based On Visual Characteristics

Image is one of the most important mediums to access,deliver and store information for humans.The diversity and visualization of forms make it to play a key role in various fields such as daily life,production and military.In recent years,varieties of image processing technologies develop and have made remarkable achievements with the improvement of technology.However,Image transmission and processing will introduce distortions to the image.In order to quantify distortion degree and apply image in various fields better,many image quality assessment methods have been proposed.Image quality measures have been a critical component to improve human vision perception and promote the development of other image processing technology.In this dissertation,an issue about perceived quality assessment for optical natural images which are most widely used is carried out.Natural images have its unique properties,and distortions will make them deviate from the original “natural statistics state”.The forms of natural images have to be adjusted according to the statistics of themselves and human visual system to make it more conducive to transmission,storage and understanding of the information.Therefore,we start with the statistics of natural images,forms of image information and characteristics of human visual system,and a number of objective evaluation measures consistent with perception quality of an image are proposed.The main contributions can be summarized as follows:1.By exploring the relationships of among perceived quality of images,types of distortions and image contents,a full reference image quality metric based on image contents is proposed by incorporating the response characteristics of classical field in human visual pathways.We firstly employ Laplacian of Gaussian operator to extract the main structure information of images.And then,we analysis the type of distortion based on the difference of structure informations between distorted image and original image.Finally,according to the difference of sensitivity of human visual system to various types of distortions in different image area,we weight and merge the image distortions in each region and combined them with the change of luminance to obtain the objective score.The experimental results on subjective database show that the objective scores of the metric are well consist with subjective scores,and the performance excels exist full reference metrics.2.According to the relation between human visual mask and image structure regularity,two full reference image quality metrics based on the similarity of structure regularity are presented.For the first metric,we use the entropy of histogram of orientation gradient for local image to represent structure regularity of an image.Moreover,the similarity of structure regularity of distorted image is combined with the similarity of luminance to get quality map.Finally,the local variance is employed to highlight the impact of important region to perceived quality.The second metric normalizes histograms of orientation gradient of each region to improve the computing method of the entropy of the first method.Since histograms of orientation gradient use gradients as weights,the normalization can reflect the importance of different regions.Therefore,the weighting processing can be omitted.A large number of experiments confirm that the two metrics proposed outperform other metrics which quantify image quality by computing the similarity of structure.In addition,although the second metric simplify calculations of the first one,its performance for each distorted images gets improvement.3.On the basis of the hypothesis that object contour can represent the main information of an image,a universal no reference image quality metric is proposed by combining the statistics of natural images in the Contourlet domain and the properties of human visual system.We firstly extract 69 Contourlet features.And then,we utilize the features above to build natural scene statistic model and distorted image model.Finally,we regard the distance between natural scene statistic model and distorted image model as the objective score of the distorted image.Experimental results illustrate that the proposed metric outperforms the most recognized universal no reference image quality metric NIQE.In addition,since we only use natural images to build natural scene statistic model,the metric has better universality than the training metrics.4.According to the limitation on the transmission efficiency from the statistics of natural image,we employ variance normalization to optimize data expression mode,and a universal no reference image quality metric and a blurred image quality assessment metric are presented.For the first one,we operate variance normalization on the luminance in the spatial domain,and extract log-energy and standard deviation as features to indicate image quality.We calculate the difference between natural scene statistic model and distorted image model to assess the quality of various types of distorted images.For the second one,we carry out variance normalization on the DCT coefficients,and the probability density distribution and standard deviation represent the blurred degree of an image.The quality of blurred image is the distance between natural scene statistic model and distorted image model.The empirical studies on the subjective database demonstrate the two no reference metrics are consistent with subjective assessment.Moreover,since they use simple and effective features to represent the quality image,they have higher efficiency in the implementation,and are applicable in the application areas which need to assess image quality in real time.