HMER论文系列
1、论文阅读和分析:When Counting Meets HMER Counting-Aware Network for HMER_KPer_Yang的博客-CSDN博客
2、论文阅读和分析:Syntax-Aware Network for Handwritten Mathematical Expression Recognition_KPer_Yang的博客-CSDN博客
3、论文阅读和分析:A Tree-Structured Decoder for Image-to-Markup Generation_KPer_Yang的博客-CSDN博客
4、 论文阅读和分析:Watch, attend and parse An end-to-end neural network based approach to HMER_KPer_Yang的博客-CSDN博客
5、 论文阅读和分析:Multi-Scale Attention with Dense Encoder for Handwritten Mathematical Expression Recognition_KPer_Yang的博客-CSDN博客
6、 论文阅读和分析:Mathematical formula recognition using graph grammar_KPer_Yang的博客-CSDN博客
7、 论文阅读和分析:Hybrid Mathematical Symbol Recognition using Support Vector Machines_KPer_Yang的博客-CSDN博客
8、论文阅读和分析:HMM-BASED HANDWRITTEN SYMBOL RECOGNITION USING ON-LINE AND OFF-LINE FEATURES_KPer_Yang的博客-CSDN博客
1.主要内容:
(1、提出创新的树结构解码器来表示树、输出树、优化基于注意力的编解码框架;
(2、设计一个问题说明特别是在复杂结构时字符解码失败的背后原因,图示解释为什么树结构在解码过程中使得更优的解码能力;
(3、证明在化学式识别和数学公式识别上,树解码器的有效性;
2.树解码器
Figure 5. Illustration of tree decoder, including parent decoder part, child decoder part, memory attention part and an optional relation prediction part. “Pred” is short for “prediction”
树解码器使用GRU和注意力机制生成一系列的子树结构:
(
o
1
c
,
o
1
p
)
,
(
o
2
c
,
o
2
p
)
,
…
,
(
o
T
c
,
o
T
p
)
.
(o_{1}^{\mathrm{c}},o_{1}^{\mathrm{p}}),(o_{2}^{\mathrm{c}},o_{2}^{\mathrm{p}}),\ldots,(o_{T}^{\mathrm{c}},o_{T}^{\mathrm{p}}).
(o1c,o1p),(o2c,o2p),…,(oTc,oTp).
需要添加一些规则,限制:
(1、每个孩子节点必须有一个父节点,因此不存在孤立节点;
(2、父节点必须是一个现存的节点;
2.1、parent解码器和child解码器
使用循环神经网络、注意力机制:
2.2、基于记忆的注意力
生成中间父亲节点序列和为训练父亲解码器的目标函数
d
t
j
m
e
m
=
tanh
(
W
m
e
m
s
t
p
+
U
m
e
m
b
e
r
s
s
j
m
e
m
)
\mathbf{d}_{tj}^{\mathrm{m}em}=\tanh(\mathbf{W}_{\mathrm{mem}}\mathbf{s}_t^{\mathrm{p}}+\mathbf{U}_{\mathrm{members}}\mathbf{s}_{j}^{\mathrm{mem}})
dtjmem=tanh(Wmemstp+Umemberssjmem)
G
t
j
m
e
m
=
σ
(
ν
m
e
m
T
d
t
j
m
e
m
)
G^{\mathrm{m}em}_{tj}=\sigma(\mathbf{\nu}^{\mathrm{T}}_{\mathrm{mem}}\mathbf{d}^{\mathrm{mem}}_{tj})
Gtjmem=σ(νmemTdtjmem)
s t p \mathbf{s}_t^p stp:父亲解码器状态;
s j m e m \mathbf{s}_j^{mem} sjmem:孩子解码器状态 s t c \mathbf{s}_t^c stc存储,作为key;
父亲节点二分类训练损失:
L
p
=
−
∑
t
∑
j
[
G
ˉ
t
j
m
e
m
log
(
G
t
j
m
e
m
)
+
(
1
−
G
ˉ
t
j
m
e
m
)
log
(
1
−
G
t
j
mem
)
]
\begin{aligned}\mathcal{L}_\mathrm{p}=-\sum_t\sum_j[\bar{G}_{tj}^\mathrm{mem}\log(G_{tj}^{\mathrm{mem}})\\ +(1-\bar{G}_{t j}^\mathrm{mem})\log(1-G_{tj}^\text{mem})]\end{aligned}
Lp=−t∑j∑[Gˉtjmemlog(Gtjmem)+(1−Gˉtjmem)log(1−Gtjmem)]
G
ˉ
t
j
m
e
m
\bar{G}_{tj}^\mathrm{mem}
Gˉtjmem:父亲节点的ground-truth;如果
j
−
t
h
j-th
j−th孩子节点被存储在内存中是step
t
t
t的父亲节点则是1,否则是0;
在测试阶段,选择作为父亲节点;
o
j
^
c
,
j
^
=
argmax
(
G
t
j
mem
)
o^c_{\hat{j}},\hat{j}=\operatorname{argmax}(\mathbf{G}_{tj}^{\text{mem}})
oj^c,j^=argmax(Gtjmem)
2.4、关系预测
正如父亲上下文向量和孩子上下文向量包含空间信息,加上父亲节点和孩子节点的内容信息,可以计算关系:
p
r
e
(
o
t
r
e
)
=
s
o
f
t
m
a
x
(
W
o
u
t
r
e
(
c
t
p
,
c
c
)
)
p^{\mathrm{re}}(o_t^{\mathrm{re}})=\mathrm{softmax}\left(\mathbf{W}^{\mathrm{re}}_{\mathrm{out}}(\mathbf{c}_t^{\mathrm{p}},\mathbf{c}^{\mathrm{c}})\right)
pre(otre)=softmax(Woutre(ctp,cc))
损失函数:
L
r
e
=
−
∑
t
log
p
r
e
(
v
t
)
\mathcal{L}_{\mathrm{re}}=-\sum_t\log p^{\mathrm{re}}(v_t)
Lre=−t∑logpre(vt)
2.5、实现在父亲注意力和孩子注意力的正则化
在不同时间步中的孩子节点,可能有相同的父亲节点,这时不同时间步的孩子节点的父亲节点的注意力概率是相似的。
L
r
e
g
=
−
∑
t
α
^
t
p
log
α
^
t
p
α
t
P
\mathcal{L}_{\mathrm{reg}}=-\sum_t\hat{\alpha}_t^{\mathrm{p}}\log\dfrac{\hat{\alpha}_t^\mathrm{p}}{\alpha_t^\mathrm{P}}\quad
Lreg=−t∑α^tplogαtPα^tp
3、损失函数
O = λ 1 L c + λ 2 L p + λ 3 L r e + λ 4 L r e g O=\lambda_1\mathcal{L_c}+\lambda_2\mathcal{L_p}+\lambda_3\mathcal{L_\mathfrak{re}}+\lambda_4\mathcal{L_{\mathfrak{reg}}} O=λ1Lc+λ2Lp+λ3Lre+λ4Lreg
实验经验上: λ 1 = λ 2 = 1 ; λ 4 = 0.1 ; \lambda_1=\lambda_2=1;\lambda_4=0.1; λ1=λ2=1;λ4=0.1;如果是数学公式识别 λ 3 = 1 \lambda_3=1 λ3=1,如果是化学式识别 λ 3 = 0 \lambda_3=0 λ3=0;
4、结论:
Table 1. Evaluation of math formula recognition systems on CROHME 2014, CROHME 2016 and CROHME 2019 test sets (in %). “ExpRate”, “≤ 1 s.error” and “≤ 1 s.error” means expression recognition rate when 0 to 2 symbol or structural level errors can be tolerated, “StruRate” means structure recognition rate.
Figure 8. Split the SMILES test set into four sub-sets (“Easy”, “Normal”, “Hard”, “Massive”) based on the length of testing SMILES strings.
Table 3. Recognition rate comparison (in %) between string decoder and tree decoder on SMILES dataset. “Easy”, “Normal”, “Hard”, “Massive” denote the four sub-sets of test set with different length of SMILES string, “All” means the overall recognition rate on the whole test set (in %). “SD” and “TD” refer to string decoder and tree decoder based approaches, respectively.文章来源:https://www.toymoban.com/news/detail-403684.html
参考:
A Tree-Structured Decoder for Image-to-Markup Generation (ustc.edu.cn)文章来源地址https://www.toymoban.com/news/detail-403684.html
到了这里,关于论文阅读和分析:A Tree-Structured Decoder for Image-to-Markup Generation的文章就介绍完了。如果您还想了解更多内容,请在右上角搜索TOY模板网以前的文章或继续浏览下面的相关文章,希望大家以后多多支持TOY模板网!