<![CDATA[The increasing prevalence of online gambling advertisements on social media has led to the use of covert strategies, such as embedding visual watermarks and employing euphemistic language, to bypass traditional detection methods, rendering manual moderation ineffective. This study proposes an AI-based automated detection system designed to identify both explicit and obfuscated gambling content. The system operates in three stages: (1) Object detection: Faster R-CNN, using a ResNet-50 backbone and Feature Pyramid Network (FPN), detects gambling-related visual elements, such as watermarks and logos; (2) Text extraction: A Transformer-based Optical Character Recognition (TrOCR) model is employed to extract textual content from images and video frames, even in the presence of visual distortions; and (3) Text classification: A BERT-based Natural Language Processing (NLP) model is used to identify gambling-related language within the extracted text. The dataset, manually collected and annotated, was augmented with Roboflow to improve model robustness and generalization. Experimental results show that the Faster R-CNN model achieved an average precision of 98.1%, TrOCR demonstrated a Character Error Rate (CER) of 4.6% and a Word Error Rate (WER) of 29%, while the BERT classifier reached an impressive 99% accuracy with high precision and recall. The system was integrated into a Flask-based web application that allows real-time analysis of both image and video inputs. This system presents strong potential to support automated content moderation and curb the spread of online gambling advertisements on digital platforms, contributing to safer online spaces.]]>
Copyrights © 2025
