Page 164 - tech fest 2025 ב״ש חוברת תקצירים

164
Beyond Age Groups: Continuous Age Prediction from Handwriting Using Deep Learning
SE-A-11
Lior Abergel; liorabergel@gmail.com Maor Merling; maormerling@walla.com
Advisors: Dr. Irina Rabaev1, Dr. Marina Litvak1 1SCE - Shamoon College of Engineering, Be’er Sheva
Estimating age from handwriting has applications in biometrics, forensics, psychology, and historical analysis. Traditional classification methods impose discrete age groups, limiting precision. This study introduces a deep-learning framework for continuous age estimation using regression. We evaluate ResNet50, InceptionV3, DenseNet121, InceptionResNetV2, and EfficientNetV2M, together with ensemble strategies to enhance accuracy. Performance is assessed using multiple metrics, including MAE, RMSE, R2 score, MAPE, and threshold-based accuracy within two and five years of actual age. Our results demonstrate the superiority of regression for handwriting-based age estimation, offering finer granularity and improved reliability. To the best of our knowledge, this is the first study to apply regression to the task of age estimation from handwriting, providing a more precise solution for real- world applications.
Keywords: deep learning, document image processing, ensemble of models, handwriting analysis, regression
Popular Science Article Generation with Text Simplification
SE-A-12
Eitay Alter; alter1eitai@gmail.com Jonatan Cohen; jonicohen97@gmail.com
Advisors: Dr. Irina Rabaev1, Dr. Marina Litvak1 1SCE - Shamoon College of Engineering, Be’er-Sheva
The number of published scientific articles has increased dramatically recently. However, these articles often contain complex terminology and specialized language, making them difficult for non- expert readers to understand. This reduces knowledge sharing and collaboration between different fields. Our goal is to improve the accessibility of scientific content by simplifying it using Transformer models. We developed SimplifiSci, which consists of a two-part pipeline: First, Transformer-based models, Grammarly-CoEdit and Google-FLAN-T5, simplify the text, then we identify scientific terms and provide definitions based on the article’s subject. We compared SimplifiSci against ChatGPT4.0 and DeepSeek-R1. The results show that SimplifiSci preserves the article’s structure while simplifying key terms and concepts. ChatGPT generated the most readable output, whereas DeepSeek produced text with the most diversity.
Keywords: ai models, machine learning, science, simplification, transformers