SLIDE 1
1/8
◭◭ ◮◮ ◭ ◮
Back Close
2/8
◭◭ ◮◮ ◭ ◮
Back Close
Introduction
Character templates: A = A0
- {κ},
E = {ea | a ∈ A}, w(ea) = wa Image and result of its recognition:
¯ s = "and___has___t_hou___sl_ai_n" ¯ c = "and has thou slain"
Template size learning for text recognition problem Bogdan - - PowerPoint PPT Presentation
Template size learning for text recognition problem Bogdan - - PowerPoint PPT Presentation
1/8 Template size learning for text recognition problem Bogdan Savchynskyy, Sergii Olefirenko IRTC ITS, Kiev www.irtc.org.ua/image Back Close 2/8 Introduction Character templates: A = A 0 { } , E = { e a |
SLIDE 2
SLIDE 3
3/8
◭◭ ◮◮ ◭ ◮
Back Close
The problem of image x recognition: ¯ s∗ = arg min
¯ s f(x, ¯
s, E) Set of recognition parameters: { ¯ w, E} = {wa, Ea | a ∈ A} Parameters learning problem in general form: ( ¯ w∗, E∗) = arg max
¯ w,E P(x, ¯
c, ¯ w, E) ↓
- 1. Template learning: x, ¯
c, ¯ w0 → E.
- 2. Image recognition: x, ¯
c, E → ¯ s∗.
- 3. Width learning: x, ¯
c, ¯ s∗, ¯ w0 → ¯ w.
SLIDE 4
4/8
◭◭ ◮◮ ◭ ◮
Back Close
Width learning problem formulation
a) set of possible widths: w(L,R)
a
= w0
a (L,R) ± i,
i = 0, n b) ¯ s∗, ¯ w → ¯ s construction c) ¯ s → E construction — by averaging Problem formulation: ¯ w∗ = arg min
¯ w∈ ¯ W f(x, ¯
s(¯ s∗, ¯ w), E(¯ s))
SLIDE 5
5/8
◭◭ ◮◮ ◭ ◮
Back Close
Problem solution
Labeling problem:
- set of vertices — V = {aL, aR | a ∈ A0}
- labels — width variations
- ∀{an, an+1} ∈ ¯
c ∃ ε (aR
n → aL n+1)
SLIDE 6
6/8
◭◭ ◮◮ ◭ ◮
Back Close
Label and edge penalties: Labeling penalty: G(¯ k) =
- v∈V
qv(k(v)) +
- v,v′∈V
gvv′(k(v), k′(v′)) G(¯ k) = f(x, ¯ s(¯ s∗, ¯ w(¯ k)), E(¯ s)) Problem formulation: ¯ w∗ = ¯ k∗ = arg min
¯ k G(¯
k). This is submodular (min, +) problem and it can be solved by MIN- CUT algorithm.
SLIDE 7
7/8
◭◭ ◮◮ ◭ ◮
Back Close
Example 1. Error rate — 0.7%
SLIDE 8
8/8
◭◭ ◮◮ ◭ ◮
Back Close