A mid-level video representation based on binary descriptors: A case study for pornography detection

Carlos Caetano, Sandra Eliza Fontes de Avila, William Robson Schwartz, Silvio Jamil Ferzoli Guimarães, Arnaldo de Albuquerque Araújo

Introduction

Detecting pornographic content is a critical component of web filtering systems used in educational and professional environments. However, it remains a significant challenge in computer vision. Early attempts at automated detection relied heavily on skin-color modeling, which frequently produced false positives in non-pornographic contexts such as beach scenes, wrestling matches, or medical imagery. While more sophisticated local features like SIFT (Scale-Invariant Feature Transform) improved accuracy, they introduced high computational costs and high-dimensional data that hindered real-time processing.

The Limitations of Frame-Level Analysis

A common strategy in video classification is to analyze individual frames independently and then apply a "majority voting" scheme to determine the video's overall label. This approach is inherently flawed for pornography detection, as explicit content may only appear in a small fraction of a video's duration. Furthermore, noise or transient visual artifacts in a few frames can easily lead to a misclassification of the entire clip.

Majority voting often fails when explicit content is sparse within a video.

Efficient Feature Extraction with Binary Descriptors

To address the need for speed and accuracy, this research proposes a pipeline utilizing binary descriptors such as ORB, BRIEF, BRISK, FREAK, and BinBoost. Unlike real-valued descriptors (e.g., SIFT), binary descriptors are significantly faster to compute and require far less memory. The study found that descriptors like ORB and BRIEF were approximately 10 times faster at feature extraction than HueSIFT, making them ideal for large-scale video processing.

BossaNova and Video-Level Aggregation

The core contribution of this work is the use of the BossaNova mid-level representation combined with a novel video aggregation strategy. While standard Bag-of-Words (BoW) models simply count the frequency of visual "words," BossaNova preserves the distribution of distances between local features and codewords, maintaining a richer visual signal.

To capture the temporal context of a video, the researchers introduced the BossaNova Video Descriptor (BNVD). Instead of classifying frames and then voting, the BossaNova vectors of all keyframes are aggregated into a single video-level descriptor using pooling functions such as min, max, mean, or median.

The aggregation process pools frame-level histograms into a single video descriptor.

Experimental Evaluation

The proposed method was evaluated using the Pornography dataset, which contains 80 hours of video (400 pornographic and 400 non-pornographic). The non-pornographic set included "difficult" cases with high skin exposure to rigorously test the system's discriminative power.

Key Findings:

  • Superior Accuracy: The best configuration, using the BinBoost descriptor with BNVD, achieved an accuracy of 92.40%. This represents a 16% reduction in error compared to previous state-of-the-art methods.
  • Robustness of the Median: Among the pooling functions, the median performed best. Its robustness to outliers allows the system to ignore "noisy" frames that do not represent the core content of the video.
  • Binary Advantage: Surprisingly, binary descriptors—when paired with BossaNova—actually outperformed more complex real-valued descriptors, proving that computational efficiency does not have to come at the cost of accuracy.

Conclusion

By shifting the focus from frame-level voting to a unified video-level representation (BNVD) and leveraging the efficiency of binary descriptors, this study provides a highly effective framework for pornography detection. While some challenges remain—such as distinguishing explicit content from non-pornographic high-skin-exposure scenes like breastfeeding—the BNVD approach sets a new benchmark for both speed and precision in video content filtering.

Summary updated

Abstract

With the growing amount of inappropriate content on the Internet, such as pornography, arises the need to detect and filter such material. The reason for this is given by the fact that such content is often prohibited in certain environments (e.g., schools and workplaces) or for certain publics (e.g., children). In recent years, many works have been mainly focused on detecting pornographic images and videos based on visual content, particularly on the detection of skin color. Although these approaches provide good results, they generally have the disadvantage of a high false positive rate since not all images with large areas of skin exposure are necessarily pornographic images, such as people wearing swimsuits or images related to sports. Local feature based approaches with Bag-of-Words models (BoW) have been successfully applied to visual recognition tasks in the context of pornography detection. Even though existing methods provide promising results, they use local feature descriptors that require a high computational processing time yielding high-dimensional vectors. In this work, we propose an approach for pornography detection based on local binary feature extraction and BossaNova image representation, a BoW model extension that preserves more richly the visual information. Moreover, we propose two approaches for video description based on the combination of mid-level representations namely BossaNova Video Descriptor (BNVD) and BoW Video Descriptor (BoW-VD). The proposed techniques are promising, achieving an accuracy of 92.40%, thus reducing the classification error by 16% over the current state-of-the-art local features approach on the Pornography dataset.
BibTeX
@misc{caetano2016mid,
  author = {Carlos Caetano and Sandra Eliza Fontes de Avila and William Robson Schwartz and Silvio Jamil Ferzoli Guimarães and Arnaldo de Albuquerque Araújo},
  title = {{A mid-level video representation based on binary descriptors: A case study for pornography detection}},
  year = {2016},
  eprint = {1605.03804},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1605.03804},
}
Figure