Related Experiment Video
Updated: Oct 11, 2025
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
682
Scene Text Recognition Based on Bidirectional LSTM and Deep Neural Network
Mvv Prasad Kantipudi1, Sandeep Kumar2, Ashish Kumar Jha3
1Department of E&TC, Symbiosis Institute of Technology, Symbiosis International (Deemed University), Pune 412115, India.
Computational Intelligence and Neuroscience
|December 3, 2021
Summary
This study introduces a new deep learning method for scene text recognition, combining deep convolution neural networks (CNN) and bidirectional LSTM (Bi-LSTM). The approach achieves high accuracy, improving upon existing algorithms for image-based text identification.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Machine Learning
Background:
- Scene text recognition is crucial for many AI applications.
- Existing algorithms struggle with accuracy and speed in complex visual scenes.
- Deep learning offers potential for improved text recognition capabilities.
Purpose of the Study:
- To propose a novel approach for enhanced scene text recognition.
- To improve the accuracy and efficiency of text identification in images.
- To integrate deep convolution neural networks (CNN) and bidirectional Long Short-Term Memory (Bi-LSTM) for scene text recognition.
Main Methods:
- A contour-based image processing technique is applied to identify image contours.
- Deep convolution neural networks (CNN) extract sequential features from the contoured images.
- Bidirectional Long Short-Term Memory (Bi-LSTM) networks encode these features for recognition.
Main Results:
- The proposed method achieved high accuracy rates across multiple datasets: MSRATD 50 (95.22%), SVHN (92.25%), vehicle number plate (96.69%), SVT (94.58%), and random datasets (98.12%).
- Quantitative and qualitative analyses confirm the approach's superior performance.
- The integration of CNN and Bi-LSTM with contour-based input enhances recognition speed and accuracy.
Conclusions:
- The novel deep learning approach significantly improves scene text recognition.
- The combined CNN and Bi-LSTM model offers a promising solution for accurate and efficient text identification in diverse image scenarios.
- This method represents a significant advancement over existing scene text recognition techniques.
