问题描述:
英语翻译
For each test image from validation set,a quantitative measure
\2p,called patch to patch mean error (henceforth P2PME),is applied
to evaluate the reconstruction quality.By averaging all distances
between reconstructed and true high-resolution patches,the
P2PME \2p is formulated as follows:
where bYi denotes the ith patch recovered and Yi its corresponding
ground truth.bYiðjÞ and YiðjÞ are the jth points in the respective
patches with S \2 S points.P denotes the total number of patches.
An advantage using Eq.(6) is that \2p works on a patch scale rather
than pixel scale,so that it can reflect more exactly about the recovery
quality of patch based super-resolution,before the overlapping
patches are averaged and smoothed and finally turned into dependent
pixels.
The \2p values for each test image taken from the first six validation
images are shown in Fig.3a.The denotations of the legend are
as follows:1.First-order gradient; 2.second-order gradient; 3.
norm luminance; al:Weighting factor a between combination of
1 and 3.And combination of such notes means concatenation of
the corresponding features.For example,13 means concatenation
of first-order gradient and norm luminance as feature vector.
As Fig.3 demonstrates,features concerning the combination of
first-order gradients and norm luminance with and without a (denoted
by 13 and 13al) have relatively lower P2PME than any other
single feature and combination for the first six validation images
when these images are used for test.When a ¼ 4,especially,feature
13al achieves best P2PME,significantly lower than most of
the other features.First-order gradient feature provides acceptable
errors as well,but we will show that its visual results are of blurry
quality (see Fig.7).The variances of patch to patch errors of the
13al are also lowest,suggesting the most stable performance for
the reconstruction.
For each test image from validation set,a quantitative measure
\2p,called patch to patch mean error (henceforth P2PME),is applied
to evaluate the reconstruction quality.By averaging all distances
between reconstructed and true high-resolution patches,the
P2PME \2p is formulated as follows:
where bYi denotes the ith patch recovered and Yi its corresponding
ground truth.bYiðjÞ and YiðjÞ are the jth points in the respective
patches with S \2 S points.P denotes the total number of patches.
An advantage using Eq.(6) is that \2p works on a patch scale rather
than pixel scale,so that it can reflect more exactly about the recovery
quality of patch based super-resolution,before the overlapping
patches are averaged and smoothed and finally turned into dependent
pixels.
The \2p values for each test image taken from the first six validation
images are shown in Fig.3a.The denotations of the legend are
as follows:1.First-order gradient; 2.second-order gradient; 3.
norm luminance; al:Weighting factor a between combination of
1 and 3.And combination of such notes means concatenation of
the corresponding features.For example,13 means concatenation
of first-order gradient and norm luminance as feature vector.
As Fig.3 demonstrates,features concerning the combination of
first-order gradients and norm luminance with and without a (denoted
by 13 and 13al) have relatively lower P2PME than any other
single feature and combination for the first six validation images
when these images are used for test.When a ¼ 4,especially,feature
13al achieves best P2PME,significantly lower than most of
the other features.First-order gradient feature provides acceptable
errors as well,but we will show that its visual results are of blurry
quality (see Fig.7).The variances of patch to patch errors of the
13al are also lowest,suggesting the most stable performance for
the reconstruction.
问题解答:
我来补答展开全文阅读