| The quality assessment of digital image is a hot research topic in the field of computer vision.Affected by various factors in life,the types of image distortion are various.Facing multiple types of image distortion,people gradually find that the subjective assessment method has the defects of complicated process,time-consuming and laborious.It inherently the objective image quality assessment comes into being.However,the existing objective image quality assessment can only deal with the problem of single distortion type image.In this thesis,aiming at the above problems,the research of blind image quality assessment for non-single distortion type is carried out,and then the blind image denoising by modeling the image noise is realized.The main research contents and innovations are as follows:(1)Blind image quality assessment for non-single distortion typeIn order to solve the problem of non-single distortion blind image quality assessment,a multi-task hierarchical deep neural network(MH-DNN)is proposed in this thesis.The network is composed of two subnetworks: distortion identification network and quality prediction network.Image quality assessment including various distortion types is achieved by use of the early sharing layer,and the generalized division regularization(GDN)inspired by biology as the activation function.Experimental results show that the algorithm proposed in this thesis is superior to DIIVINE and BRESQUE algorithms in PLCC,SROCC and other indicators,and can solve the problem of image quality assessment including many types of distortion.(2)Blind Image quality assessment and denoising based on noise modelingIn the research process of image noise modeling and denoising,according to the variational inference theory,an algorithm is proposed by integrating noise estimation and image denoising into a unique Bayesian framework in this thesis.Firstly,a denoising convolution network is trained by clean/corrupted pairs of images to obtain the parameters of the posterior probabilities,and then the noise distribution image is obtained.Secondly,the denoising network is constructed according to the noise distribution images,and a denoising MS-SSIM loss function.Finally,the blind image denoising is realized.Experimental results show that the algorithm has good performance in noise estimation and removal of real scenes(Gaussian noise,Poisson noise,salt and pepper noise),which can provide a new idea for blind image denoising. |
