The Research Of No-reference Image Quality Assessment Based On Statistic Of Natural Image Information

Image information is a main method of visual information received of human beings, and also plays an important role in communications. However, during the procedures of acquisition, processing, coding, transmission and reproduction, digital image may produce distortion in visual quality due to physical devices and image processing algorithms, so how to effectively assess image quality is very important. This paper in analyzing the reasons of distortion and the features of distorted images, we proposed:(1) A no-reference image quality assessment based on wavelet. Firstly, as we all known, human visual system has a character of multi-channel, so we can make use of multi-resolution in wavelet transformation to simulate human eyes. Secondly, different image region for human eyes has a different subjective perception, such as edge region, texture region, and smooth region. We need to design a method to extract the important region for human eyes. Thirdly, for natural scene statistic model we know that wavelet sub-band energy has a nearly linear principle in logarithm domain. So we can use the least square method, linear prediction and energy compensation to predict the original image sub-band energy. Finally, the quality metric could be constructed by real distorted image sub-band energy and predicted original image sub-band. Experimental result shows that this algorithm can assess a range of distortion types and be consistent with subjective evaluation.(2) A no-reference image quality assessment based on Contourlet transforms. First we use Contourlet transform to obtain image sub-band, and then according to the human visual system (HVS) to extract its visual interesting region. At last the metric would be obtained with measuring the distortion features in the natural scene statistics model. Because JPEG distortion is not suitable for this model, we using energy rate based on block outer and inner to expand our algorithm. Experimental result shows that this method can assess a range of distortions and be consistent with subjective evaluation.(3) A no-reference image quality assessment in blur and noise. According to the influence of local pixel in blur and noise distortions. Firstly, we can use local maximal deviation to get noise factor. Secondly, using some de-noise algorithm to decrease the effect of noise. Thirdly, the blur factor can be obtained by computing the deviation of edge region in blur and re-blurred image. Because of the mask effect in human eyes, the blur factor should be compensated in order to restore the original distortion intension. At last, the metric would be constructed by blur factor and noise factor. Experimental result shows that this metric be consistent with subjective evaluation.