Video Objectionable Content Detection via Embedding Similarity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Video sharing services face challenges in detecting and quickly removing videos containing objectionable content, such as pornography, violence, and objectionable language, from their platforms.
Innovation Solution
A method using a neural network to predict an embedding for an uploaded video, which specifies its location in a multi-dimensional space based on content. This method identifies nearby videos and determines if the uploaded video is likely to contain objectionable content, triggering its blocking from the service.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional content detection methods are used, then the system is simpler to implement, but the speed and accuracy of detecting objectionable content is insufficient
Solution Approach 1:
The system pre-processes video content by extracting features and generating embeddings before actual detection is needed. Video clips are pre-analyzed for visual, audio, and metadata features, and these pre-computed embeddings are stored for rapid retrieval during detection operations, enabling fast identification of objectionable content without real-time heavy processing
Solution Approach 2:
The patent introduces an intermediary embedding space as a mediator between raw video data and detection algorithms. Instead of directly analyzing complex video streams, the system transforms videos into simplified embedding representations that capture essential content characteristics, allowing faster and more accurate detection through similarity search in this intermediate representation space
2Measurement precision
If manual verification is used for all uploaded videos, then the accuracy of content moderation is high, but the time required for processing is excessive
Solution Approach 1:
The system applies partial automation by using AI-based detection to handle the majority of videos (those with high confidence predictions), while reserving manual verification only for edge cases where the AI system is uncertain or conflicts arise. This partial application of automated action achieves high accuracy for most content while significantly reducing overall processing time compared to full manual verification
Solution Approach 2:
The system incorporates feedback mechanisms where manual verification results are fed back into the AI training process. Verified human judgments are used to refine and retrain the detection models, continuously improving accuracy over time. This feedback loop allows the system to learn from human decisions and become more accurate without requiring manual verification of every single case
3Reliability
If AI-based detection is implemented, then the speed and accuracy of detecting objectionable content improves, but the computational resources required increase
Solution Approach 1:
The detection process is segmented into multiple stages with varying computational intensity. The system first performs quick filtering using lightweight feature extraction and similarity search to identify potential candidates, then applies more computationally intensive analysis only to videos that pass the initial filter. This segmentation of processing tasks reduces overall computational resource consumption while maintaining high detection accuracy
Data Source
AI summary
Techniques are disclosed for identifying videos containing objectionable content. An example method comprises identifying, by a computing system, a video uploaded to a video sharing service, generating an embedding for the video using a neural network, wherein the embedding specifies a location of the video in a multi-dimensional space where a plurality of videos are located based on content of the videos, identifying from the videos a plurality of associated videos that each have an associated embedding that is located within a predetermined distance of the embedding, determining whether the video is likely to include a particular type of objectionable content by at least determining at least a predetermined amount of the associated videos that contain the particular type of objectionable content, and responsive to determining that the video is likely to include the particular type of objectionable content, causing the video to be blocked from the video sharing service.