Algorithm 1: Compressed Poisson noise reduction based on the SCENet |
Input: degraded image z and trained parameters aL=1,...,20, F=1,2,3, bL=1,...,20 |
Output: restored image y |
1: Compute DC, AC(1,0), AC(0,1) by T{z}. |
2: Obtain S = {SDC, SAC(1,0), SAC(0,1)} by merging the each coefficient. |
3: for L = 1, …, 20 do |
4: Stabilize using VST by f{S}. |
5: Apply convolution with trained parameters and then destabilize it by f −1{aL,F* f{S}}. |
6: Appy BN and ReLU by max(BN{f −1{aL,F* f{S}}}, 0). |
7: end for |
8: Obtain Sout = {SDC,out, SAC(1,0),out, SAC(0,1),out} by applying a fully-connected layer. |
9: Estimate Lout by T−1{Sout}. |
10: Obtain H by z −T−1{S}. |
11: Estimate Hout by running steps 3−8 above with H and bL=1,...,20 instead of S and aL,F. |
12: Estimate final restored image by y = Lout + Hout. |