Pashto script and graphics detection in camera captured Pashto document images using deep learning model.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Layout analysis is crucial for Document Image Analysis (DIA) but underexplored for Pashto documents.
• Existing DIA systems lack specific models for Pashto text and graphics.
• The unique visual characteristics of Pashto documents present a significant challenge.

Purpose of the Study:

• To develop a deep learning-based classifier for detecting Pashto text and graphics in document images.
• To create a novel, real-world dataset of Pashto documents for research and development.
• To establish a baseline for Pashto document layout analysis.

Main Methods:

• Development of a deep learning classifier utilizing a variant of Faster R-CNN, specifically the Single-Shot Detector (SSD).
• Creation of a new dataset comprising over 1,000 camera-captured Pashto document images.
• Application of Convolutional Neural Networks (CNNs) for feature extraction and classification.

Main Results:

• Achieved a mean average precision (mAP) of 84.90% on a test set of 300 Pashto document images.
• Successfully demonstrated the ability to differentiate between Pashto text and graphical elements.
• Validated the effectiveness of the proposed SSD-based deep learning model for Pashto document layout analysis.

Conclusions:

• The proposed deep learning model effectively performs layout analysis on Pashto documents, a previously unexplored area.
• The newly created Pashto document dataset is a valuable resource for future research in DIA.
• This work significantly contributes to the field of document image analysis for low-resource languages.