Deep Learning Architecture for Off-Line Recognition of Handwritten Math Symbols
| Autor: | Yosra Meguebli, Afef Kacem Echi, Kawther Khazri Ayeb |
|---|---|
| Rok vydání: | 2021 |
| Předmět: |
Alphanumeric
business.industry Deep learning Context (language use) Pattern recognition 02 engineering and technology Convolutional neural network Symbol (chemistry) Field (computer science) 03 medical and health sciences 0302 clinical medicine Pattern recognition (psychology) 0202 electrical engineering electronic engineering information engineering 020201 artificial intelligence & image processing Artificial intelligence Transfer of learning business 030217 neurology & neurosurgery Mathematics |
| Zdroj: | Pattern Recognition and Artificial Intelligence ISBN: 9783030718039 MedPRAI |
| DOI: | 10.1007/978-3-030-71804-6_15 |
| Popis: | Real scientific challenge, handwritten math formula recognition is an attractive field of pattern recognition leading to practical applications. Hundreds of alphanumeric and math symbols need to be recognized, many are so similar in appearance that some use of context is necessary for disambiguation. Analysis of the spatial relationships between symbols is challenging. In this work, we focus on handwritten math symbols and propose to recognize them by a deep learning approach. The symbol images, used for train, validation, and test are generated from Competition on Recognition of Online Handwritten Mathematical Expressions dataset (CROHME) 2019’s online patterns of mathematical symbols. As the large dataset is crucial for the performance of the deep learning model and it is labor-intensive to obtain a large amount of labeled data in real applications, we first augmented the database. Standing on the transfer learning technique, we then tested and compared several pre-trained Convolutional Neural networks (CNNs) like VGGNet, SqueezeNet, DenseNet, and Xception network and we tuned them to better fit our data. An accurate classification of 91.88% (train), 88.82% (validation), and 83.68% (test) for 101 classes is achieved, using only off-line features of the symbols. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|