Research On Image Quality Assessment Method Based On Shearlet Transform

The distortion can make the image lose information and degrade its quality,and the subjective visual experience of users and subsequent image processing are affected.For the image processing system,the recognition and quantization of the image quality is an indispensable task.In this thesis,the problem of the image quality assessment(IQA)is deeply investigated from four aspects based on characteristics of the shearlet transform(ST)and perceptual features of the human visual system(HVS).The main work of this thesis is given as follows:1.A full-reference natural image quality evaluation method based on the multiscale and multidirectional perceptual error(MMPE)in the shearlet domain is proposed,which can deal with the problem of the unconspicuous directional features in existing methods.This method adopts the ST and multiple lowlevel psychophysical characteristics of the HVS,and can improve the predictive accuracy.Firstly,to simulate the multi-channel mechanism of the HVS,the ST is used to decompose the image into multiple subbands at different scales and orientations.Secondly,the local directional band-limited contrast and the visual just noticeable difference threshold model in the ST subbands are defined.To handle the visual masking problem,both the contrast masking effect and the entropy masking effect are considered.Finally,perceptual errors of all subbands are pooled together to yield the quality assessment result of the distorted image.Experimental results demonstrate that the proposed method has better performance compared to the state-of-the-art IQA methods.2.A full-reference natural image quality assessment method based on shearlet transform subband structural similarity(STSS)is proposed,which can deal with the problem that spatial structural features in the images with severe distortion are often broken.In this method,firstly,the ST is used to decompose the reference and distorted images into multiple subbands at different scales and directions;secondly,the structural similarity of each subband between the reference and distorted images is computed;finally,the different weights are assigned to the structural similarity of subbands at different scales and directions based on perceptual characteristics of the HVS,and the weighting sum of the structural similarity of all subbands is calculated as the objective quality assessment result of the distorted image.Experimental results demonstrate that the proposed method achieves the better performance for the images with severe distortion.3.A reduced-reference natural image quality assessment method based on the statistical features of divisive normalization transform coefficients in the shearlet domain(SFDS)is proposed,which can deal with the problem that subband coefficients of the linear transforms in existing methods have the stronger statistical correlation.Firstly,the DNT is applied to ST subband coefficients of the image,and the purpose of this operation is to reduce the statistical correlation between subband coefficients of the shearlet transform.Secondly,the Gaussian distribution is used to approximate the statistical distribution of the normalized DNT coefficients and the difference between the Gaussian distribution and the actual distribution is calculated.This difference and the statistical features of the actual distribution are used as the features of the image subbands.Finally,different weights are assigned to the features of subbands at different scales and directions based on the characteristics of the HVS,and the similarities of weighted subband features between reference and distorted images are combined together as the quality assessment result of the distorted image.Experimental results show that the proposed method can achieve better quality assessment performance.4.A regionalized structural features based evaluator(RSFE)is proposed to implement the no-reference screen content image(SCI)quality assessment,which can deal with the problem that textual regions and pictorial regions in the SCI have completely different characteristics.Firstly,the shearlet local binary pattern is used to extract the multiscale and multidirectional texture features in the shearlet domain,and meanwhile,the local derivative pattern is used to extract the texture features in the spatial domain.For pictorial regions,the texture feature is adopted as their structural feature and meanwhile,the luminance feature is used as their auxiliary feature.For textual regions,improved histograms of oriented gradients extracted from multi-order derivatives are used as their structural features.Secondly,to derive the quality assessment values of textual and pictorial regions,the features of these two regions are supplied separately to the support vector regression.Finally,to obtain the quality assessment value of the whole content of the SCI,an activity weighting strategy is used to combine the quality values of textual and pictorial regions.Experimental results show that the proposed method achieves better performance than the state-of-the-art SCI quality assessment methods.