Abstract
To guarantee a satisfying Quality of Experience (QoE) for consumers, it isrequired to measure image quality efficiently and reliably. The neglect of thehigh-level semantic information may result in predicting a clear blue sky asbad quality, which is inconsistent with human perception. Therefore, in thispaper, we tackle this problem by exploiting the high-level semantics andpropose a novel no-reference image quality assessment method for realistic blurimages. Firstly, the whole image is divided into multiple overlapping patches.Secondly, each patch is represented by the high-level feature extracted fromthe pre-trained deep convolutional neural network model. Thirdly, threedifferent kinds of statistical structures are adopted to aggregate theinformation from different patches, which mainly contain some common statistics(i.e., the mean\&standard deviation, quantiles and moments). Finally, theaggregated features are fed into a linear regression model to predict the imagequality. Experiments show that, compared with low-level features, high-levelfeatures indeed play a more critical role in resolving the aforementionedchallenging problem for quality estimation. Besides, the proposed methodsignificantly outperforms the state-of-the-art methods on two realistic blurimage databases and achieves comparable performance on two synthetic blur imagedatabases.