Research On Efficient No-reference Image Quality Assessment Technology

With the rapid development of multimedia intelligent terminal and network information technology,digital image acquisition and sharing become more and more convenient.At the same time,due to different shooting equipment,complex environment and transmission processing factors,the quality of digital images cannot be guaranteed,that is,the quality of digital images is vulnerable to damages from the process of production to the final arrival at consumers.Therefore,evaluating the image quality accurately and monitoring the image quality strictly in each link to ensure the image quality to the maximum extent are of great importance to the application of digital images.As the image is ultimately observed by people,the most accurate quality evaluation method is subjective quality evaluation,namely by the observers to judge the quality of the image.However,this approach requires a lot of human resources and is unable to operate in practice.Therefore,the objective quality assessment,which through the designed algorithm to evaluate the quality of the image is particularly important.Toward this end,this paper carries out an in-depth study on objective image quality evaluation methods.The main work and innovation points of this paper are summarized as follows:On the aspect of no-reference image quality assessment,the existing methods mostly belong to supervised method,this kind of methods trains a quality evaluation model with the subjective ratings and then uses the obtained model to forecast the image quality.However,such methods are usually restricted to the used training set.Therefore,changing the dataset will affect the algorithm performance by a certain degree.Namely,the generalization ability of these algorithms is very limited.By comparison,the unsupervised method has strong generalization ability without the intervention of the subjective score.Therefore,we focus our attention on the unsupervised methods.Through exploring the representing effect of statistical features of natural images and visual saliency on image quality,we propose an unsupervised no-reference method based on natural image statistics and visual saliency.The experimental results verify the high consistency between the proposed method and subjective quality evaluation.Contrast distortion is one of the common types of image distortions due to improper exposure.For example,overexposure or underexposure will results in too bright or too dark images.Most existing no-reference evaluation methods usually give a quality score,but cannot describe the image distortion in detail.For example,the information such as which region of the image has distortion or what degree the distortion is.Such information can be used to guide image contrast enhancement and thereby conveys practical value.Therefore,in this paper,we propose a method to describe the contrast-distorted image.A local distortion mapping model is learned through the convolutional neural network,and the model is used to describe the contrast distortion area and the degree of distortion accurately.Experimental results show that the proposed method can predict the information of the contrast distortion in the image precisely.