Online handwriting refers to any text created with a computer, smartphone, or similar digital device with a digital pen, so called digital writing. With the evolution of several different technologies, the development of handwriting recognition is becoming increasingly significant. The target group of this project is the children of the 1-3 grades. The main aim is to create a model able to recognize the writing of children. In this project, two different handwriting recognition models are developed, both with deep learning for online handwriting and the other one for offline handwriting. The accuracy of the models is compared against the targetted >= 80%. Different methods of preprocessing and segmentation are developed to analyze the online and offline handwriting data. Different layers, optimizers and batch sizes have been experimented with in this project. The model with the highest score for IAM data will be selected to predict children's data. The two different machine learning models to child handwriting is discussed. Finally, a simple application to recognize the handwriting is built.
Machine Learning, Deep Lerning, Beam Search, Language Model, Jaro Winkler, Levenshtein distance, N-Gram
For teachers, the manual reading of handwritten German texts of second to fourth grade elementary school students is a tedious task. If children lack the necessary motor skills in handwriting, they produce poorly legible writing and teachers struggle even more to read their texts. The use of online handwriting recognition tools could help teachers to spend more time on feedback rather than on manual handwriting recognition. The purpose of this project is to construct two different machine learning models of writing recognition in the German language.
With the problem from above in mind, there are some goals that are defined in this project.
1. Develop two different handwriting recognition model with IAM handwriting database. One with deep learning for online handwriting and the other one for the offline handwriting
2. Use different methods of preprocessing or segmentation if necessary for analyze the online and offline handwriting data
3. Comparison of the two different machine learning model for the children handwriting
Both models have shown a good result for recognizing the handwriting in the IAM-Database. Due to the lack of the children's data, it is difficult to prove wheter the model can recognize the handwriting of the children. There are 53 children's data, which are used in this project. The online model shows a less than 25% error and the offline model shows a 50% error rate.
| Duratrion: | 25 Wochen |
| Time: | 720 hours |
| Teamgrösse: | 1 Person |
Susanne Grassmann
edu-RD GmbH
Industriestrasse 20
5200 Brugg
fhnw.ch
Haocen Wang
Prof. Dr. Martin Melchior martin.melchior@fhnw.ch
