No Reference Image Quality Assessment Based On Pyramid Transform

With the continuous development of shooting equipment,digital images have been applied in many fields in recent years.The requirements put forward by people have gradually improved as the attention that scientists focus on has become how to assess the image quality accurately.According to different evaluation subjects,IQA can be divided into two kinds,one is subjective assessment and the other is objective assessment.Subjective assessment is greatly limited to practical applications with the disadvantage of high time cost,high expenses,tedious operation.As for objective assessment,it has processed algorithm in the computer in advance and it can predict the quality of images by extracting feature information of distorted images after images are input into the computer.It is more practical compared to the subjective assessment.Due to the various distorted types,diverse distorted manifestations the effective method of objective image quality assessment for all distorted types is still to be developed.Early Fourier transforms indicate that all the curve waveforms can be superimposed by sine and cosine shifts,which opens up a new direction for image processing.However,the shortcomings of the localization of the Fourier transform itself,limited its application in image processing.Pyramid decomposition is the best method of high dimensional information processing.It has the advantages of translation invariance and direction controllability.It is based on human visual perception,Its structural characteristics are considered as the model of the primary treatment form in the visual nerve.The multi-directionality is consistent with the different sensitivity of human vision to different directions.The multi-scale is consistent with the multi-resolution of human vision,and it has strong ability to express local features,so the image pyramid decomposition is used to deal with various fields.This paper proposes three objective image quality assessment based on Pyramid decomposition combined with Pyramid decomposition theory.The main content is summarized as follow.1.Based on the research of directional control pyramid,the stereo image quality evaluation algorithm of directional control pyramid is proposed.First,the minimum energy difference method is used to get the left and right view of the disparity map,and then the left view,the right view and the disparity map were 4-scale 12-direction controllable pyramid(Steerable Pyramid)decomposition,getting three high-frequency sub-band 114 direction sub-band.The binary generalized Gaussian distribution fitting is performed on the 48 directional sub-bands corresponding to the decomposition of the left and right views,extracting its shape parameters and scale parameters.And the feature information such as cross-scale correlation and spatial correlation is extracted from all direction sub-bands.Finally,these features were inputted into support vector regression(SVR)training prediction to obtain stereo image quality score.2.According to the principle that the image distortion will change the high frequency information,a stereo image quality evaluation algorithm based on double complex wavelet transform is proposed.First,the parallax is extracted from the left and right views according to the minimum energy error principle,then,double complex wavelet decomposition are done for the left and right views,set the low frequency information to zero,merge the high frequency with the low frequency,setting to 0,getting the fused image containing only high frequency information,Then,the asymmetric generalized Gaussian distribution(AGGD)and the relative gradient direction are extracted from the left and right fusion images and the disparity map.Finally,obtain the objective quality of stereo image,through Ada-Boosting BP neural network training.3.This paper proposed the Laplacian transform of Pyramid in the gradient direction,that transform the gradient direction of the gradient filter in four directions based on Laplacian Pyramid,filtering sub,each layer of the Laplacian gold tower,has four images of different gradient direction,and the probability distributions of these filter sub-bands are counted,get the shape parameters and variance,as well as,energy characteristics and localized binarized LBP features of grayscale co-occurrence matrices.Use Ada-Boosting by BP neural network regression model to predict image quality.The experimental results showed that the SROCC and CC of the LIVE library were both up to 0.94 and the results were accurate and consistent with the subjective evaluation.