Algorithm 2: Adversarial Process of Parameter Agent Network (PAN) and Recognizer

Input: input image P, selected region A

Output: optimized ${\mathbf{PAN}}^{\prime }$

1: sample transformation state $s_1$  from predicted distribution:  $s_1\leftarrow \mathbf{PAN}\left(A\right)$

2: generate random transformation states from $s_2$ to state $s_{11}$, based on state $s_1$

3: all states including $s_1-s_{11}$ contain anchor points

4: obtain augmented images from $x_1$ and $x_{11}$ based on states $s_1$ to state $s_{11}$:

5: The following i all represent 1 to 11

6: $x_i\leftarrow \mathbf{PT}(P,s_i)$

7: identify images:  $\mathit{Reg}\leftarrow \mathbf{Recognizer}\left(P\right)$  ${\mathit{Reg}}_i\leftarrow \mathbf{Recognizer}\left(x_i\right)$

8: if  $\mathit{ID}\left[{\mathit{Reg}}_i\right]=\mathit{ID}\left[\mathit{Reg}\right]$ then

9:    test distance $\Delta$ between ${\mathit{Reg}}_i$ and $\mathit{Reg}$

10:    if $max\Delta ({\mathit{Reg}}_i,\mathit{Reg})$ then

11:      ${\mathbf{PAN}}^{\prime }$ = update $\mathbf{PAN}$ with $s_i$

12:    end if

13: else

14:   ${\mathbf{PAN}}^{\prime }$ = $\mathbf{PAN}$

15: end if