Statistical Evaluation Of No-reference Quality Evaluation Metrics Of Remote Sensing Images

Digital imaging has become an indispensable part of human daily life with the rapid development of digital imaging technology.The quality of digital images will be affected by various degraded sources in the process of collection,transmission,storage and reconstruction,which will result in digital images having different degrees of quality degradation.Therefore,the demand for image quality evaluation becomes more and more urgent.The evaluation methods for image quality can be divided into two categories:one is the subjective quality assessment method,and the other is the objective quality assessment method.According to the different conditions required by the assessment method,objective quality assessment can be divided into three categories:Full Reference Image Quality Assessment(FR-IQA),and Reduced Reference Image Quality Assessment(RR-IQA)and No Reference Image Quality Assessment(NR-IQA).The raw image information is difficult to obtain in many practical applications.For image applications such as image de-smoothing,de-blurring,and fusion,few reference images can be used to compare with the enhanced image.Therefore,the no-reference image quality evaluation method is more practical than the first two methods.In recent years,many no-reference image quality assessment metrics(IQM)have been proposed to evaluate digital image quality.However,most of the non-reference metrics are designed for grayscale or color images.Are they suitable for remote sensing multispectral imaging?Still unknown.In this article,we first selected 21 commonly used quality metrics without reference images.Then,we used the average filter,Gaussian white noise,and linear motion to degrade the high-resolution Quickbird remote sensing images including the three different contents of cities,rural areas,and ports respectively.Each degradation method included 40 different levels of degradation.Next,we applied 21 quality evaluation metrics to the processed remote sensing images,and calculated the corresponding values for each quality evaluation index.Then we evaluate whether these image quality assessment metrics are robust based on the accuracy of prediction,the criteria of predictive monotony and predictive consistency.Finally,we use the multivariate statistical method.We have selected those image quality evaluation metrics that are considered to be robust,and divided them into several groups for factor analysis.Through quantitative calculation,we select the image quality evaluation index with the highest load factor as the group's representative.The experimental results show that the images of different contents and the results of different types of degradation suffer from different performance.Through experiments,we found that only 7 of the 21 commonly used image quality assessment metrics met the requirements for robustness.It is recommended to use Edge Intensity(EI)and Significant Distortion(JND)to evaluate the quality of the image that has been degraded by the averaging filter.Using an edge intensity(EI),the anisotropic blind image quality assessment(BIQAA)and the mean index(MM)were used to evaluate the quality of images that experienced degradation of Gaussian white noise.Laplace's derivative(LD),apparent distortion(JND)and standard deviation(SD)are recommended to evaluate the quality of a linearly moving image.Finally,it is recommended that edge intensity(EI)be used to assess image quality for images that do not know which type of degradation is specifically affected.