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

Figure 2

Illustration of our G-Net. i and j on X^i,j represents the scale index and processing index. On the left-hand side is the feature pyramid structure with residual blocks X^i,j on each scale. For the skip connection, we propose 1 × 1 convolution layers to reduce the feature channels and sum with decoded features for coarse decoding. On the right-hand side is our proposed progressive fine decoding structure through HFE. Note that X^i,1 do not implement any processing, they are coarse decoded features from left. Moreover, we apply the Tanh activation function to the output from X^0,5 and the element-wise sum between output and input as global residual connection to obtain the final result.