The common loss function in the field of hyper Division

1 Reconstruction losses

Directly compare the differences between pixels , In the field of hyperfractionation, usually L1 Loss function ratio L2 Clearer loss function 、 Convergence is faster .

class ReconstructionLoss(nn.Module):
    def __init__(self, type='l1'):
        super(ReconstructionLoss, self).__init__()
        if (type == 'l1'):
            self.loss = nn.L1Loss()
        elif (type == 'l2'):
            self.loss = nn.MSELoss()
        else:
            raise SystemExit('Error: no such type of ReconstructionLoss!')

    def forward(self, sr, hr):
        return self.loss(sr, hr)

2 Against the loss

GAN yes 2014 The unsupervised learning network that emerged in .GAN It is mainly the iterative optimization of the generation model and the discrimination model . The optimization goal is ：

  Maximize the probability estimation of the discriminant model for real samples , Minimize the estimation of the generated samples by the discriminant model .

3 Loss of perception

Perceived loss unlike MSE In that way, the pixels are calculated in the image space , Instead, the difference is made in the feature space .

  As can be seen from the above formula , Perceived loss is actually related to MSE The losses are very similar , It's just transforming the computing space from image space to feature space .