Rectification and Super-Resolution Enhancements for Forensic Text Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Text extraction from images is difficult due to low resolution and text orientation.
• Accurate text recognition is vital for identifying illegal content in environments like the Tor Darknet and Child Sexual Abuse (CSA) imagery.

Purpose of the Study:

• To evaluate and improve text recognition performance on challenging image datasets.
• To assess the effectiveness of combining text recognizers with rectification and super-resolution techniques.

Main Methods:

• Evaluated eight different text recognizers.
• Integrated recognizers with rectification networks and super-resolution algorithms.
• Tested the combined approaches on standard and custom datasets, including TOICO-1K and CSA-text.

Main Results:

• Achieved a 0.3170 score on TOICO-1K using Deep Convolutional Neural Networks (CNN) and rectification.
• Reached a 0.6960 score on the CSA-text dataset with resolution enhancements.
• Observed a 4.83% performance increase on ICDAR 2015 by combining MORAN recognizer with Residual Dense resolution.

Conclusions:

• Rectification methods generally outperform super-resolution when used independently.
• The combination of rectification and super-resolution yields the best overall performance improvements across datasets.
• These advanced techniques are crucial for enhancing text extraction in difficult real-world scenarios.