Algorithm 1: Global spatial level adaptation

Input: iris data $X$, pixel label $Y, \mathrm{segmentation} \mathrm{network} G$, training epochs $T$, ratio coefficient $\gamma$

Initialize: binary discriminator $D$

For t = 1,…,T do

Unfreeze the D and freeze the G

$\mathrm{Compute} \mathrm{the} \mathrm{logit} \mathrm{maps}: P=softmax\left(G\left(X\right)\right)$

$\mathrm{Compute} \mathrm{the} \mathrm{entropy}: E^{\left(h,w\right)}\left(X\right)=-{\displaystyle \sum }_c^K-P^{\left(h,w,c\right)}\cdot \log \left(P^{\left(h,w,c\right)}\right)$

Sort the logit maps $P$ based the $S=\sqrt[h\cdot w]{{\displaystyle \prod }_{i=1}^h{\displaystyle \prod }_{j=1}^w{E}^{\left(i,j\right)}}$

Split the iris data $X$ corresponding to sorted $P$ into $X_{hard}$ and $X_{easy}$ by ratio coefficient $\gamma$

Compute the D’s loss: $L_{global}^{adv}\left(X_{easy},X_{hard}\right)=\log \left(1-D\left(E_{hard}\right)\right)+\log \left(D\left(E_{easy}\right)\right)$

$\underset{D}{\min }{L}_{global}^{adv}$ to train $D$

Unfreeze the G and freeze the D

$\mathrm{Compute} \mathrm{the} \mathrm{cross}-\mathrm{entropy}: L_{cross}^{seg}\left(X\right)=-{\displaystyle \sum }_h^H{\displaystyle \sum }_w^W{\displaystyle \sum }_c^K{Y}^{\left(h,w,c\right)}\cdot \log \left(P^{\left(h,w,c\right)}\right)$

$\underset{G}{\min }{L}_{cross}^{seg}+\underset{G}{\max }{L}_{global}^{adv}$ to train $G$