Pornography classification: The hidden clues in video space-time.

Area of Science:

• Computer Science
• Artificial Intelligence
• Multimedia Security

Background:

• Automatic pornography detection is crucial for online child safety.
• Traditional methods using still images miss vital motion information.
• Existing approaches lack efficiency and accuracy in video classification.

Purpose of the Study:

• To develop an efficient and accurate video-pornography classification method.
• To introduce a novel space-temporal feature descriptor, Temporal Robust Features (TRoF).
• To evaluate TRoF's performance against commercial and existing scientific solutions.

Main Methods:

• Developed Temporal Robust Features (TRoF) for space-temporal motion description.
• Utilized Fisher Vectors for aggregating local TRoF information into a mid-level representation.
• Evaluated performance on the new Pornography-2k dataset (2000 videos, 140h).

Main Results:

• The best TRoF approach achieved a 79% reduction in classification error compared to commercial classifiers.
• A sparse TRoF description reduced errors by over 69% with 19x less memory.
• The method demonstrated high classification accuracy and low false negative rates.

Conclusions:

• TRoF offers a significant advancement in video-pornography detection.
• The proposed method is both effective and efficient, suitable for real-time applications.
• This work contributes a robust solution for enhancing online safety, particularly for children.