A Smart Visual Sensing Concept Involving Deep Learning for a Robust Optical Character Recognition under Hard...

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Document image distortions like blur, shadows, and low contrast severely degrade optical character recognition (OCR) system performance.
• Existing OCR models show limited success in recognizing text from degraded documents, especially under harsh real-world conditions.

Purpose of the Study:

• To develop a robust OCR model capable of high-accuracy text recognition from severely distorted document images.
• To introduce a novel deep-learning architecture that significantly outperforms previous OCR methods on challenging datasets.

Main Methods:

• Proposed a new OCR model architecture combining convolutional neural networks (CNNs) and recurrent neural networks (RNNs).
• Developed and utilized a new dataset featuring diverse and representative real-world distortion scenarios.
• Evaluated the model's performance against existing OCR approaches on degraded document images.

Main Results:

• The proposed CNN-RNN model demonstrates superior performance compared to existing related works.
• Achieved up to 97.5% accuracy in recognizing text from document images previously considered unrecognizable.
• The new dataset effectively showcases the model's robustness under various harsh distortion conditions.

Conclusions:

• The novel deep-learning-based OCR model offers a significant advancement in recognizing text from distorted documents.
• The developed model provides a robust and accurate solution for challenging OCR tasks, particularly in adverse imaging conditions.
• This research addresses the critical need for reliable OCR systems in real-world applications with image quality issues.