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 { κ } , E = { e a | a ∈ A } , w ( e a ) = w a Image and result of its recognition: s = "and___has___t_hou___sl_ai_n" ¯ c = "and has thou slain" ¯ Widths tuning: ◭◭ ◮◮ ◭ ◮ Back Close
The problem of image x recognition: 3/8 s ∗ = arg min ¯ s f ( x, ¯ s, E ) ¯ Set of recognition parameters: { ¯ w, E } = { w a , E a | a ∈ A } Parameters learning problem in general form: w ∗ , E ∗ ) = arg max ( ¯ w,E P ( x, ¯ c, ¯ w, E ) ¯ ↓ w 0 → E . 1. Template learning: x, ¯ c, ¯ s ∗ . 2. Image recognition: x, ¯ c, E → ¯ w 0 → ¯ s ∗ , ¯ 3. Width learning: x, ¯ c, ¯ w . ◭◭ ◮◮ ◭ ◮ Back Close
4/8 Width learning problem formulation a) set of possible widths: s ∗ , ¯ b) ¯ w → ¯ s construction ( L,R ) ± i, w ( L,R ) = w 0 i = 0 , n a a c) ¯ s → E construction — by averaging ◭◭ Problem formulation: ◮◮ w ∗ = arg min ◭ s ∗ , ¯ ¯ W f ( x, ¯ s (¯ w ) , E (¯ s )) w ∈ ¯ ◮ ¯ Back Close
5/8 Problem solution Labeling problem: • set of vertices — V = { a L , a R | a ∈ A 0 } • labels — width variations ∃ ε ( a R n → a L • ∀{ a n , a n +1 } ∈ ¯ c n +1 ) ◭◭ ◮◮ ◭ ◮ Back Close
Label and edge penalties: 6/8 Labeling penalty: G (¯ � � g vv ′ ( k ( v ) , k ′ ( v ′ )) k ) = q v ( k ( v )) + v ∈ V v,v ′ ∈ V G (¯ w (¯ s ∗ , ¯ k ) = f ( x, ¯ s (¯ 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. Back Close
Example 1. Error rate — 0 . 7% 7/8 ◭◭ ◮◮ ◭ ◮ Back Close
Example 2. Error rate — 0 . 8% 8/8 ◭◭ ◮◮ ◭ ◮ Back Close
Recommend
More recommend