Nowadays, deep learning is widely used in various fields, and deep learning has also been widely used in the field of IQA. To do any kind of visual task, corresponding evaluation indicators are needed to judge the effect of the algorithm. The following is my summary of the field of image quality evaluation. evaluation index.
1. SROCC
The full name is Spearman rank -order correlation coefficient ( SROCC ), and some documents are also written as SRCC or " Spearman rank correlation coefficient ".
SROCC performs a linear correlation analysis on the rank size of two target arrays (the rank correlation of two sets of data), which is often considered as the Pearson linear correlation coefficient after the two objects are arranged separately, and has a wide range of applications. The value range is 0~1, and when the performance value is equal to 1, it indicates that the two sets of data are completely consistent (it is impossible to achieve in the field of image quality evaluation...).
N represents the number of samples, and di represents the difference between the i-th image’s subjective quality score ranking and objective quality score ranking (assuming that the array of subjective evaluation values is x, and the array of objective evaluation values is y, firstly divide x and y in order from large to small Arrange in order, record xi, yi as the positions of the i-th image in the respective arrays after sorting, then di = xi - yi).
2. PLCC
The full name of Pearson linear correlation coefficient (Pearson linear correlation coefficient, PLCC ) is used to evaluate the accuracy of IQA model prediction , and some literature is written as linear correlation coefficient (Linear correlation coefficient, LCC ).
PLCC describes the linear correlation between two sets of data, and its value range is -1~1. When the value of PLCC is zero, it means that the two sets of data are completely uncorrelated (the objective and subjective quality scores of the image are very different); when the value of PLCC is 1 or -1, it means that the two sets of data are completely correlated (the objective quality score of the image Quality Score is the same as Subjective Quality Score). PLCC describes the correlation between the objective evaluation score of the algorithm and the subjective score of the human eye, and measures the accuracy of the prediction of the IQA algorithm.
Among them, Q represents the original objective quality score, β1, β2, β3, β4, and β5 are model parameters (parameters that need to be fitted), and p is the objective quality score after regression operation. Formula 2 calculates PLCC, si and pi represent the subjective quality score and objective quality score of the i-th image respectively, sˉ and pˉ represent the average value of subjective quality score and objective quality score respectively.
3. KROCC
The full name of the Kendall rank-order correlation coefficient (KROCC), written as "Kendall rank correlation coefficient" in some literature, is used to measure the monotonicity of the prediction results of the IQA model just like SROCC.
The larger the value of KROCC, the better the correlation between the two signal data, and the smaller the value, the worse the correlation.
Where N represents the number of samples, and Nc is the number of consistent pairs in the data set (also known as "harmonious pairs", which refer to two sample observations with the same order of variable size, that is, the order of high and low levels of x is the same as that of y, otherwise called "discordant" or "discordant"), and Nd is the number of discordant pairs in the data set. Any pair of data (xi, yi) and (xj, yj) in the two data sequences, when xi > xj and yi > yj or xi < xj and yi < yj, the data pair is consistent (consistent high and low order); when xi > xj and yi < yj or xi < xj and yi > yj, the data pairs are inconsistent (inconsistent high and low order); when xi = xj or yi = yj, it is neither consistent nor inconsistent.
The above are commonly used rating indicators for image quality evaluation, and others include RMSE (root mean square error), SSIM, PSNR, etc. Readers who are interested can explore it by themselves, and these are introduced here.