Algorithm 1. TRAINING INTERATION PROCEDURE

Require:$B$ is the batch size, $\beta$ is a batch of data with size $B$, and $\epsilon$ donates each sample in a batch. $L_{\mathit{score}}$ is score loss, $L_{\mathit{boundary}}$ is boundary loss, $\eta$ is a hyperparameter, $\theta$ denotes the model parameters, $\tilde{\theta }$ denotes the updated model parameters. $\alpha$ is the learning rate.

1: $\beta =\mathit{RandomSample}\left(\left\{{\epsilon }_1,···,{\epsilon }_N\right\},B\right)$

2: for${\epsilon }^j\in \beta$

3:  $L_{\mathit{score}}+={\sum }_{t=0}^j\mathit{CEH}({\mathit{score}}_t,s_t)$

4:  $L_{\mathit{boundary}}+={\sum }_{t=0}^j\left(SmoothL1\right(2*\sigma \left(\mathrm{\Delta }{y}_{1,t}\right),2*\sigma (t-gt{b}_{left,t}))+SmoothL1(2*\sigma \left(\mathrm{\Delta }{y}_{2,t}\right),2*\sigma (gt{b}_{right,t}-t)\left)\right)*s_t$

5:  $L+=\eta *L_{\mathit{score}}+(1-\eta )*L_{\mathit{boundary}}$

6: end for

7: $\tilde{\theta }=\theta -\alpha {\nabla }_{\theta }L$

8: return$\tilde{\theta }$