Skip to content
GitLab
Menu
Projects
Groups
Snippets
Loading...
Help
Help
Support
Community forum
Keyboard shortcuts
?
Submit feedback
Contribute to GitLab
Sign in / Register
Toggle navigation
Menu
Open sidebar
wangsen
paddle_dbnet
Commits
78d9efcf
Commit
78d9efcf
authored
May 06, 2022
by
tink2123
Browse files
rename svtr_lcnet
parent
13e929b1
Changes
1
Hide whitespace changes
Inline
Side-by-side
Showing
1 changed file
with
7 additions
and
7 deletions
+7
-7
doc/doc_ch/PP-OCRv3_introduction.md
doc/doc_ch/PP-OCRv3_introduction.md
+7
-7
No files found.
doc/doc_ch/PP-OCRv3_introduction.md
View file @
78d9efcf
...
@@ -70,7 +70,7 @@ LKPAN(Large Kernel PAN)是一个具有更大感受野的轻量级[PAN](https://a
...
@@ -70,7 +70,7 @@ LKPAN(Large Kernel PAN)是一个具有更大感受野的轻量级[PAN](https://a
<a
name=
"3"
></a>
<a
name=
"3"
></a>
## 3. 识别优化
## 3. 识别优化
PP-OCRv3 识别模型在 PP-OCRv2 的基础上从8个策略上进一步优化,整体 pipelin
n
e 如下图所示:
PP-OCRv3 识别模型在 PP-OCRv2 的基础上从8个策略上进一步优化,整体 pipeline 如下图所示:
<img
src=
"../ppocr_v3/v3_rec_pipeline.png"
width=
800
>
<img
src=
"../ppocr_v3/v3_rec_pipeline.png"
width=
800
>
...
@@ -87,7 +87,7 @@ PP-OCRv3 识别模型在 PP-OCRv2 的基础上从8个策略上进一步优化,
...
@@ -87,7 +87,7 @@ PP-OCRv3 识别模型在 PP-OCRv2 的基础上从8个策略上进一步优化,
|-----|-----|--------|----| --- |
|-----|-----|--------|----| --- |
| 01 | PP-OCRv2 | 8M | 74.8% | 8.54ms |
| 01 | PP-OCRv2 | 8M | 74.8% | 8.54ms |
| 02 | SVTR_Tiny | 21M | 80.1% | 97ms |
| 02 | SVTR_Tiny | 21M | 80.1% | 97ms |
| 03 |
PP-LCNet_SVTR
| 12M | 71.9% | 6.6ms |
| 03 |
SVTR_LCNet
| 12M | 71.9% | 6.6ms |
| 04 | + GTC | 12M | 75.8% | 7.6ms |
| 04 | + GTC | 12M | 75.8% | 7.6ms |
| 05 | + TextConAug | 12M | 76.3% | 7.6ms |
| 05 | + TextConAug | 12M | 76.3% | 7.6ms |
| 06 | + TextRotNet | 12M | 76.9% | 7.6ms |
| 06 | + TextRotNet | 12M | 76.9% | 7.6ms |
...
@@ -109,9 +109,9 @@ PP-OCRv3 期望在提升模型精度的同时,不带来额外的推理耗时
...
@@ -109,9 +109,9 @@ PP-OCRv3 期望在提升模型精度的同时,不带来额外的推理耗时
1.
将SVTR网络前半部分替换为PP-LCNet的前三个stage,保留4个 Global Mixing Block ,精度为76%,加速69%,网络结构如下所示:
1.
将SVTR网络前半部分替换为PP-LCNet的前三个stage,保留4个 Global Mixing Block ,精度为76%,加速69%,网络结构如下所示:
<img
src=
"../ppocr_v3/svtr_g4.png"
width=
800
>
<img
src=
"../ppocr_v3/svtr_g4.png"
width=
800
>
2.
将4个 Global
Attenntion
Block 减小到2个,精度为72.9%,加速69%,网络结构如下所示:
2.
将4个 Global
Mixing
Block 减小到2个,精度为72.9%,加速69%,网络结构如下所示:
<img
src=
"../ppocr_v3/svtr_g2.png"
width=
800
>
<img
src=
"../ppocr_v3/svtr_g2.png"
width=
800
>
3.
实验发现 Global
Attention
的预测速度与输入其特征的shape有关,因此后移Global Mixing Block的位置到池化层之后,精度下降为71.9%,速度超越 CNN-base 的PP-OCRv2-baseline 22%,网络结构如下所示:
3.
实验发现 Global
Mixing Block
的预测速度与输入其特征的shape有关,因此后移
Global Mixing Block
的位置到池化层之后,精度下降为71.9%,速度超越 CNN-base 的PP-OCRv2-baseline 22%,网络结构如下所示:
<img
src=
"../ppocr_v3/LCNet_SVTR.png"
width=
800
>
<img
src=
"../ppocr_v3/LCNet_SVTR.png"
width=
800
>
具体消融实验如下所示:
具体消融实验如下所示:
...
@@ -120,9 +120,9 @@ PP-OCRv3 期望在提升模型精度的同时,不带来额外的推理耗时
...
@@ -120,9 +120,9 @@ PP-OCRv3 期望在提升模型精度的同时,不带来额外的推理耗时
|-----|-----|--------|----| --- |
|-----|-----|--------|----| --- |
| 01 | PP-OCRv2-baseline | 8M | 69.3% | 8.54ms |
| 01 | PP-OCRv2-baseline | 8M | 69.3% | 8.54ms |
| 02 | SVTR_Tiny | 21M | 80.1% | 97ms |
| 02 | SVTR_Tiny | 21M | 80.1% | 97ms |
| 03 |
PP-LCNet_SVTR
(G4) | 9.2M | 76% | 30ms |
| 03 |
SVTR_LCNet
(G4) | 9.2M | 76% | 30ms |
| 04 |
PP-LCNet_SVTR
(G2) | 13M | 72.98% | 9.37ms |
| 04 |
SVTR_LCNet
(G2) | 13M | 72.98% | 9.37ms |
| 05 |
PP-LCNet_SVTR
| 12M | 71.9% | 6.6ms |
| 05 |
SVTR_LCNet
| 12M | 71.9% | 6.6ms |
注: 测试速度时,输入图片尺寸均为(3,32,320); PP-OCRv2-baseline 代表无蒸馏模型
注: 测试速度时,输入图片尺寸均为(3,32,320); PP-OCRv2-baseline 代表无蒸馏模型
...
...
Write
Preview
Markdown
is supported
0%
Try again
or
attach a new file
.
Attach a file
Cancel
You are about to add
0
people
to the discussion. Proceed with caution.
Finish editing this message first!
Cancel
Please
register
or
sign in
to comment