Enhancing the YOLOv8 model for realtime object detection to ensure online platform safety

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Online platforms face challenges in detecting harmful objects like weapons and illicit substances.
• User safety necessitates robust content moderation and detection systems.
• Existing deep learning models require enhancement for complex detection scenarios.

Purpose of the Study:

• To introduce an Enhanced Object Detection (EOD) model for improved identification of harmful objects.
• To enhance the YOLOv8-m architecture for superior feature extraction and detection.
• To contribute to online user safety through advanced harmful object detection.

Main Methods:

• The study enhances the YOLOv8-m architecture by modifying cross-stage partial fusion blocks.
• Three additional convolutional blocks are incorporated into the model head for improved feature processing.
• A public dataset with six categories of harmful objects is utilized for training and evaluation.

Main Results:

• The EOD model achieved high performance metrics: 0.88 precision, 0.89 recall, and 0.92 mAP50 on standard test data.
• On challenging test cases, the model demonstrated strong results with 0.84 precision, 0.74 recall, and 0.82 mAP50.
• The EOD model surpasses existing deep learning approaches in harmful object detection accuracy.

Conclusions:

• The Enhanced Object Detection (EOD) model significantly improves the identification of harmful objects in complex online environments.
• Explainable AI techniques were used to validate the model's decision-making process and confidence.
• This research sets a new benchmark for harmful object detection, enhancing safety on digital platforms.