Bridging classical and neural methods for improved segmentation in mathematical text based images

Sakshi ¹, Chetan Sharma ², Vivek Bhardwaj ³

⁰Amity Institute of Information Technology, Amity University, Noida, India.

Scientific Reports +

|

December 13, 2025

View abstract on PubMed

Summary

This study introduces an optimal neural network for segmenting handwritten mathematical expressions, significantly improving recognition accuracy. The new method outperforms traditional techniques, enhancing computer vision applications.

Area Of Science

Computer Vision
Image Processing
Artificial Intelligence

Background

Mathematical expression recognition is hindered by segmentation challenges.
Existing research prioritizes recognition over segmentation, particularly in computer vision and image processing.
Handwritten mathematical text and expression recognition requires robust segmentation.

Purpose Of The Study

To address the critical segmentation problem in handwritten mathematical expression recognition.
To analyze and develop an optimal segmentation solution for mathematical expressions.
To improve the accuracy and robustness of mathematical expression recognition systems.

Main Methods

Exploration, classification, and testing of classical segmentation methods.
Comparative case analyses on diverse mathematical expression datasets.
Development and proposal of an optimal neural network-based segmentation approach.

Main Results

The proposed neural network model achieved competitive mean Intersection over Union (IOU) scores across multiple datasets (CROHME, Aidapearson, HasyV).
Performance metrics included 79.4% (CROHME 2014), 83.5% (CROHME 2016), 81.3% (CROHME 2019), 74.6% (Aidapearson), and 79.6% (HasyV).
The neural network approach demonstrated superior performance compared to traditional segmentation methods.

Conclusions

The neural network-based segmentation effectively overcomes limitations of classical methods.
The proposed solution shows significant potential for enhancing mathematical expression recognition systems.
This work advances segmentation techniques crucial for computer vision and AI in mathematical contexts.

Keywords:

CROHME dataset Handwritten Mathematical expression Neural network Segmentation