Improving the Neural Segmentation of Blurry Serial SEM Images by Blind Deblurring

Figure 4

Illustration of model training structure. Firstly, the supervised image z_s and unsupervised image z_u are fed into the generator. For the supervised learning, perceptual loss, adversarial loss, and content loss are calculated through VGG19, discriminator, and the supervised image pair, respectively. For the unsupervised learning, we use the measurement score from the discriminator to minimize the differences between and . Note that all the images that are fed into D-Net are processed with differentiable augmentation.

(a)

(b)