Open junxnone opened 4 years ago
CA-FCN
Vatti clipping algo
Sequence label
Character label
hxwxc
c
hxwx1
hxwxN
N
H = Q * S
5 epchs
@SynthText - character level
1 epoch
@real images sequence-level
10^-3
10^-4
10^-5
64x256
Lc
Lo
Ls
Lm
λ
λl
=10
λo
λm
=0
=1
λ = 0.2 - γ = 2
λ = 0.2
γ = 2
English Dataset Test
Chinese Dataset Test
junxnone/tech-io#749
Reference
Brief
CA-FCN
Vatti clipping algo
- vati 1992Sequence label
-Character label
- Mutual-Supervision类别分支
几何分支
Attention Decoder vs Segmentation-based vs TextScanner
Arch
Class Branch
hxwxc
c
= all character classes + backgroundGeometry Branch
hxwx1
hxwxN
N
- 预定义的字符序列长度H = Q * S
-hxwxN
Mutual-Supervision
Word Formation
Training
5 epchs
pre-train@SynthText - character level
+1 epoch
fine-tuning@real images sequence-level
10^-3
->10^-4
->10^-5
64x256
Loss
Lc
- Localization mapLo
- Order segmentationLs
- Text segmentationLm
- Mutual supervisionλ
λl
-=10
λo
-=10
λm
=0
=1
Test
English Dataset Test
Chinese Dataset Test