急求电子类英语高手帮忙翻译,不要软件翻译,急需!
急求电子类英语高手帮忙翻译,不要软件翻译,急需!
IQM Evaluation Criteria: The results presented in Tables III and IV are essentially identical, since they both are measuring the correlation of an IQM after an identical nonlinear fit.
However,RMSE is a more intuitive criterion of IQM comparison than the linear correlation coefficient because the latter is a nonuniform metric, and it is not easy to judge the relative performance of IQMs especially if the metrics are all doing quite well.The QA community is more accustomed to using the correlation coefficient,but RMSE gives a more intuitive feel of the relative improvement of one IQM over another.Nevertheless,one can derive identical conclusions from the two tables.
The Spearman rank order correlation coefficient merits further discussion.SROCC belongs to the family of nonparametric correlation measures,as it does not make any assumptions about the underlying statistics of the data.For example,SROCC is independent of the monotonic nonlinearity used to fit the IQM to DMOS.However,SROCC operates only on the rank of the data points,and ignores the relative distance between datapoints. For this reason,it is generally considered to be a less sensitive measure of correlation,and is typically used only when the number of datapoints is small[28] .In Table V,although the relative performance of IQMs is essentially the same as in Table III,there are some interesting exceptions.For example,one can see that on JPEG#2 dataset,Sarnoff JNDMetrix performs almost as good as SSIM(MS) or VIF[in fact,it is statistically indistinguishable from VIF[in fact,it is statistically indistinguishable from VIF or SSIM(MS)on this dataset],but the SROCC shows a different conclusion .that is,JND is much worse than SSIM(MS) or VIF. A further analysis of the data and the nonlinear fits reveals that the fits of the three methods are nearly identical(see Fig.6),and the linear correlation coefficient,or RMSE is indeed a good criterion for comparison.It is also easy to see why SROCC is not a good criterion for this dataset.One can note that many datapoints lie in that region of the graph where the DMOS is less than 20(high quality),and all of these image are high quality with approximately the same quality rank.The difference between DMOS of these images is essentially measurement noise.Since SROCC ignores the relative magnitude of the residuals,it treats this measurement noise with the same importance as datapoints in other parts of the graph,where the quality rank spans a larger range of the measurement scale.Thus,in this case,the nature of the data is such that SROCC is not a good measure of IQM performance ,and leads to misleading conclusions.
不要软件翻译