Image Quality Assessment Based On Dictionary Learning And Sparse Coding

In the procedure of image acquisition,storage,transmission and processing,the visual quality of the image declines inevitably due to various factors.For applications related to image quality,designing an effective computational model to automatically estimate image quality is a very important focus.In addition,objective image quality assessment has also been the focus of the academic area.Image degradation often results in subjective visual differences of images,accompanied by changes in the structure of the image content.In recent years,based on this observation,many techniques for extracting the intrinsic structure of images have been successfully applied to the field of image quality assessment.Among them,the technology of dictionary learning and sparse coding has received extensive attention due to the existence of its biological neuroscience foundations related to the human visual system and computational basis.The current image quality assessment methods based on sparse representation focus on utilizing the sparse coefficients to construct image features,while the role of the dictionary can be further exploited and utilized.In addition,the existing methods use the linear reconstruction model of the signal to characterize the image,where the linear model has certain limitations since that the image data in the real world is more complex.In the view of above problems,this paper intends to explore and study from two aspects of dictionary and sparse coefficient based on the sparse representation theory and related technologies.First,for blurred and compressed images,the degradation will lead to the loss of high-frequency details.Through the analysis of the problem and the observation of the corresponding relationship between two dictionaries of the reference image and the degraded version,we find that the difference of dictionaries can be used to quantify the degree of degradation when the coding coefficient is fixed.Therefore,a quality evaluation method based on hybrid dictionary learning for these types of images is proposed.Second,in view of the fact that complex image data often lies in a low-dimensional sub-manifold in a high-dimensional space,this paper introduces kernel sparse coding to linearize the nonlinearity in image data and thus proposes an image quality evaluation metric based on kernel sparse coding.At the same time,considering the physical meaning of the sparse coding,the resulting sparse coefficients and reconstructed residuals are comprehensively utilized by the proposed method for estimating the quality score,which be more likely compatible with the human visual system.To validate the two proposed methods,this paper separately designs and carries out a series of experiments on the commonly used image quality evaluation benchmark datasets.The experimental results demonstrate the effectiveness and superiority of the proposed methods.