Edge loss interpretation

The paper ： https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8461664

In image hyperfraction , We often use MSE loss, But it has some disadvantages , Such as ：
MSE There is no good correlation between the evaluation index and people's subjective perception of the image , Because it considers the average , The influence of some salient features is not well considered , The significant structure and texture information are highly correlated with human subjective perception .

This paper presents an edge loss (edge loss) function , High frequency texture structure information can be well considered , Improve the detail performance of hyperimage .

The function is as follows ：

among X and Y They are respectively the hyperfractional image output from the network and the real hyperfractional image ,E Is the edge feature obtained .
In fact, the formula is easy to understand , |X-Y| The absolute error between the hyper image and the real image is expressed , But we are more concerned about the information on the edge , So use E De weighting , The proportion of the edge part will become larger , So this loss To better optimize edge information .