Related Experiment Video
Updated: Jun 26, 2025
05:56
Objectification of Tongue Diagnosis in Traditional Medicine, Data Analysis, and Study Application
Published on: April 14, 2023
2.5K
DB-EAC and LSTR: DBnet based seal text detection and Lightweight Seal Text Recognition.
Baohua Huang1, Aokun Bai1, Yuqiong Wu2
1School of Computer and Electronic Information, Guangxi University, Nanning, China.
Plos One
|May 16, 2024
Summary
This study introduces improved models for Chinese seal text recognition, enhancing efficiency in document processing. The new DB-ECA and LSTR models achieve high accuracy despite limited data and image challenges.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Document Analysis
Background:
- Chinese seal text recognition is crucial for efficient document processing.
- Low accuracy is attributed to image blurring, occlusion, and limited datasets.
- Existing methods struggle with these challenges.
Purpose of the Study:
- To develop robust models for accurate Chinese seal text detection and recognition.
- To address limitations of existing methods, particularly with small datasets and image degradation.
- To improve overall efficiency in administrative document workflows.
Main Methods:
- Improved Differentiable Binarization (DBnet) model (DB-ECA) incorporating efficient channel attention (ECA) and delayed downsampling for text detection.
- Lightweight Seal Text Recognition (LSTR) model utilizing a lightweight CNN, self-attention, and Connectionist Temporal Classification (CTC) for text recognition.
- Developed a novel homemade dataset for experimental validation.
Main Results:
- DB-ECA outperformed five common detection models with precision (90.29%), recall (85.17%), and F-measure (87.65%).
- LSTR achieved the highest accuracy (91.29%) compared to five recent recognition models.
- LSTR demonstrated advantages in parameter efficiency and inference speed.
Conclusions:
- The proposed DB-ECA and LSTR models significantly enhance Chinese seal text recognition accuracy and efficiency.
- These models offer effective solutions for real-world scenarios with data scarcity and image quality issues.
- The developed models represent a notable advancement in automated document processing and administrative efficiency.
