Download presentation
Presentation is loading. Please wait.
Published byRudolf Kelly Modified over 9 years ago
1
Data-Driven Handwriting Synthesis in a Conjoined Manner Hsin-I Chen, Tse-Ju Lin, Xiao-Feng Jian, I-Chao Shen, Bing-Yu Chen National Taiwan University, University of British Columbia 1 1 1 2 1 1 2
2
Handwriting Font is Popular Personal Website Invitation letter Personal Decoration
3
Handwriting Synthesis “The grown-ups are certainly very, very old”, he said to himself, as he continued on his journey. “The grown-ups are certainly very, very old”, he said to himself, as he continued on his journey Segoe print Lucida Handwriting
4
Different shape of the same character he little prince he tippler
5
Previous work - Handwriting Style From forensic science viewpoint, Important factors of handwriting style: Elements of styles Elements of execution Natural variance between each writing Other effects Alcohol, emotion, etc.
6
Previous work - Handwriting Synthesis Handwriting Synthesized HandwritingSynthesized Handwriting Synthesized Wang et al., IJDAR 2005Chang and Shin, IJDAR 2012 Lin and Wang, PR 2007
7
Contribution A statistical learning approach to synthesize non- existent paragraph: A data-analysis stage A character grouping method A data-driven optimization framework
8
Approach pipeline Data CollectionParameterizationShape Model Paragraph Synthesis Word Synthesis Character Synthesis
9
Approach pipeline Data CollectionParameterization Shape Model Paragraph Synthesis Word SynthesisCharacter Synthesis
10
Data Collection We should collect at least two instances for each letter. We should cover more commonly used letter pairs(“aa”, “ab”, “ac”,…). The collection sheet should not over be constrained. 10
11
Data Collection
12
Approach pipeline – Parameterization Data collectionParameterizationShape Model Paragraph Synthesis Word SynthesisCharacter Synthesis
13
Character Parameterization HandwrittenReconstructed Control points
14
Approach pipeline – Shape Model Data collectionParameterizationShape Model Paragraph Synthesis Word SynthesisCharacter Synthesis
15
Shape Model Shape coefficientDisplacement
16
Approach pipeline – Character Synthesis Data collectionParameterizationShape Model Paragraph Synthesis Word SynthesisCharacter Synthesis
17
Collected dataSynthesized data
18
Approach pipeline – Word Synthesis Data collectionParameterizationShape Model Paragraph Synthesis Word SynthesisCharacter Synthesis
19
Character Grouping We group letters to obtain more conjoining information Example: Criteria Synthesis target: Data set: How to connect ‘r’ and ‘i’ ?
20
Character Grouping The probability we connect two neighboring letters Ending group Starting group
21
Word Synthesis Smoothness term Data term Boundary constraint
22
Word Synthesis Synthesized Result Connected
23
Approach pipeline – Paragraph synthesis Data collectionParameterizationShape Model Paragraph Synthesis Word SynthesisCharacter Synthesis
24
Paragraph Synthesis Line angle, word height and word angle
25
User Study : Visual discrimination
27
65 Subjects
28
User Study : Similarity test Handwritten sample Our result
29
User Study : Similarity test Handwritten paragraph Synthesized paragraph without layout info Synthesized paragraph with layout info
30
Visual Result Handwritten paragraph Our synthesized paragraph
31
Visual Result Style 1 Style 2 Style 3
32
Comparison II Handwritten samples Synthesized result by Lin et al., PR2007 Handwritten samples Our result
33
Comparison I Original writingWang et al., IJDAR 2005Ours
34
Conclusion We present a data-driven optimization approach to synthesize non-existed paragraph : We analyzing cursiveness property in the data. A novel trajectory optimization for synthesizing conjoining character. The user study and comparison results show that our approach successfully imitate one’s handwriting style.
35
Q & A Thank you http://graphics.csie.ntu.edu.tw/~fensi/
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.