Inverse Scene Text Removal

2025-06-26

AI Summary

This work addresses the potential misuse of scene text removal (STR) techniques by introducing, for the first time, the inverse scene text removal (ISTR) task—aimed at detecting whether an image has undergone STR and precisely localizing the erased text regions. Methodologically, we propose an end-to-end deep learning framework integrating a synthetic-data-driven binary detector with a tampering localization module, and further explore content recovery via text recognition models. Experiments demonstrate state-of-the-art performance in both detection accuracy and localization IoU, significantly outperforming baseline methods; however, text content recovery remains challenging, exposing the fragility of STR-induced steganographic artifacts. This study establishes the first systematic framework for security assessment and digital forensics of STR technologies, providing both a principled methodology and a benchmark suite for ISTR research.

Technology Category

• Computer Vision: Adversarial Attacks & Robustness
• Natural Language Processing: Safety and Robustness
• Intelligent Robots: Localization, Mapping, and Navigation

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Security and Privacy: Large-scale security measurements
• Web Mining and Content Analysis: Normalization, clustering, classification, and summarization of Web text

Abstract

Scene text removal (STR) aims to erase textual elements from images. It was originally intended for removing privacy-sensitiveor undesired texts from natural scene images, but is now also appliedto typographic images. STR typically detects text regions and theninpaints them. Although STR has advanced through neural networksand synthetic data, misuse risks have increased. This paper investi-gates Inverse STR (ISTR), which analyzes STR-processed images andfocuses on binary classification (detecting whether an image has un-dergone STR) and localizing removed text regions. We demonstrate inexperiments that these tasks are achievable with high accuracies, en-abling detection of potential misuse and improving STR. We also at-tempt to recover the removed text content by training a text recognizerto understand its difficulty.

Problem

• Detect if an image underwent text removal
• Locate regions where text was removed
• Attempt to recover the removed text content

Innovation

• Inverse STR for detecting text removal
• Binary classification of STR-processed images
• Localizing removed text regions accurately