Font Size: a A A

Study On Some Key Issues Of Image Quality Assessment

Posted on:2017-01-15Degree:DoctorType:Dissertation
Country:ChinaCandidate:T H WangFull Text:PDF
GTID:1108330491962907Subject:Computer Science and Technology
Abstract/Summary:PDF Full Text Request
Images, as the foremost information carrier, play an important role in our daily life. In many image processing tasks (e.g., image acquisition, compression, restoration, transmission, etc.), it is necessary to assess the quality of the output image. The end-user of images is human, thus the subjective assessment is always the ultimate and the most reliable test. However, the subjective assessment is time-consuming, expensive and cannot be real-time. That is why objective methods mimicking human perception have been developed to assess the perceived quality automatically.This dissertation mainly focused on the full-reference image quality assessment (FRIQA) and no-reference image quality assessment (NR IQA) metrics. The main work included in this dissertation are as follows:(1) For the structrual similarity measures in FR IQA, the similarity map of which are generated in a local manner based on the specific features extracted from an image. The local window of them are usually fixed in size, this make them unable to reflect the difference among edge, texture and smoothness regions. While the human visual system is sensitive to these regions, thus the fixed-sized local window strategy correlates not very well with human judgements. To address these issues, we resort to perceptual grouping strategy to enhance the performance of IQA metrics. The proposed method firstly adopt the superpixel algorithms to conduct perceptual groupings, then compute the structrual similarity measures of the reference superpixel and distortion superpixel, and finally integrate all the superpixel-based similarity to yield a quality score. For consideration of efficiency, the author implement a simple linearity iterative clustering method, it has a good generalization ability to handle different kinds of images, including grayscale images, color images and medical images. Experimental results on several image databases show its effectiveness.(2) The engineering based FR IQA usually follow the two-step trend, namely, feature extraction with pooling strategy following behind. Thus great efforts should be focused on feature extraction step and pooling step. It matters what features we selected and what pooling strategy we adopted. The features should be human visual system sensitive and the pooling strategy also should be have a good connection with the selected features. It is noted that the contrast which is a low-level feature in the human visual system, plays an important role in understanding an image. The author assumes there is a certain connection between the contrast and standard deviation pooling strategy. The first perception for this two terms is that they are both the range indicator. The contrast reflects the change of luminance, while standard deviation is the range for indicating the distortion severity of an image. This phenomenon may indicate the internal connection between each other. To further explore the relationship of them, we resort to multi-scale technology. The reason is that human visual system usually has the characteristic of multi-resolution, namely it is changeable to the observation distance and sampling frequency. Furthermore, images are naturally multi-scale. The proposed method, namely, multi-scale contrast similarity deviation, firstly compute the contrast similarity map of the reference and distortion image at each scale, and then apply the standard deviation pooling to it to generate the similarity score, and finally pooling across different scales to yield the quality score. Experimental results show the effectiveness and efficiency of the proposed method.(3) The natural scene statistics based NR IQA metrics use a model to characterize the distribution of coefficients in certain domain, and the parameters of the model are treated as the features to capture the variation of the marginal distribution. In consideration of human visual system’s sensitive to edge and the edge can be well captured by gradient. It is also noted that the log probability characterizes the non-Gaussian nature of these probability distributions more accurately, and illustrates the nature of the tails more clearly. To this end, we do not focus on the raw probability but rather the log of the probability. Experiments show that the log histograms of the log directional derivatives are well fitted by general Laplacian model, and besides, different distortions alter the histogram in their own way. Thus the features of the proposed method are obtained from the log histograms of log horizontal, vertical, main-diagonal and secondary-diagonal derivatives in the gradient domain, along with variance, kurtosis, differential entropy and entropy. To capture the multiscale behavior of images, all those features are then computed at two scales. Thus we have 32 features for an image. With the features pre-extracted, a support vector machine regressor is utilized to map the feature vector to a quality score. Several natural image databases and a group of medical images demonstrate the effectiveness and efficiency of the proposed method.
Keywords/Search Tags:image quality assessment, perceptual grouping, pooling, general Laplacian model
PDF Full Text Request
Related items