ApsaraVideo Media Processing (MPS) provides video AI features such as video production, media review, media fingerprinting, and smart tagging. This topic provides an overview of these features.

Scenarios

The video AI features help you identify, analyze, and understand audio and video content. These features allow you to perform the following operations:
  • Analyze video content to intelligently generate thumbnails, perform image matting, and blur logos.
  • Detect prohibited content on short video platforms, live streaming platforms, and media platforms.
  • Identify duplicate and similar clips to prevent plagiarism or duplication, perform quick content moderation, calculate ad revenue, and trace the sources of videos.
  • Generate multi-dimensional tags for videos to facilitate video searching, personalized recommendation, and intelligent advertising.

Features

  • Media review: This feature supports content moderation and content quality analysis. It adopts deep learning algorithms trained by massive amounts of tagged data to detect prohibited content in audio, text, and images of a video.
    • Content moderation: This feature reviews the content, title, description, and thumbnail of a video and detects prohibited content in audio, text, and images of a video.
      • Pornography detection: integrates neural network algorithms with a sample library that contains millions of data entries to detect pornography in audio, text, and images.
      • Terrorist content detection: integrates deep learning algorithms with a real-time updated sample library to detect sensitive content, such as the content on weapons, bloody scenes, specific costumes, smoke and light scenes, specific logos, crowding, and parades.
      • Ad violation detection: intelligently detects different forms of ads, such as advertising text, watermarks, illegal ads, QR codes, and mini program codes.
      • Logo detection: uses object detection technologies to detect logos in a video, including logos of TV stations, trademarks, and watermarks. This helps with copyright protection.
      • Undesirable scene detection: integrates user behavior analysis with time series comparison technologies to identify undesirable scenes, such as picture-in-picture (PiP), smoking scenes, scenes of live broadcasting while driving, and meaningless images.
      • Audio anti-spam: adopts advanced acoustic models and language models to identify pornographic, terrorist, politically sensitive, and abusive content in English and Chinese.
    • Content quality analysis: This feature analyzes the quality of a video and identifies quality issues about images and audio in a video.
      • Image quality analysis: identifies quality issues about images in a video, such as ghosting, blurring, underexposure, overexposure, black screens, white screens, too much noise, mosaics, TV snow, frame freezing, frame skips, and reproducing.
      • Audio quality analysis: identifies quality issues about audio in a video, such as stuttering, muting, and lack of audio tracks.
      • Image aesthetic analysis: analyzes the aesthetic of images in a video and gives a score.
  • Media fingerprinting: This feature is implemented based on video recognition technologies developed by Alibaba Cloud. It uses a binary string to identify a video and supports comparison between images or audio of multiple videos. This feature applies to scenarios such as queries of duplicate videos and the tracing of video sources.
    • Plagiarism identification: This feature identifies whether a video plagiarizes other content products. This provides technical support for copyright protection, and therefore helps build a healthy video commercialization ecosystem where the benefits of video producers and short video platforms are protected.
    • Identification of duplicate videos: This feature creates a media fingerprint library to compare the fingerprints of a video with those in the library. This helps identify duplicate videos or clips. This also prevents excessive duplicate or similar videos from being pushed to users and affecting user experience during content delivery and personalized recommendation.
    • Quick content moderation: This feature maintains a video blacklist to compare the fingerprints of a video with those in the blacklist to determine whether the video contains prohibited content. Compared with traditional video moderation methods, this method provides higher efficiency, higher accuracy, and lower costs in processing large numbers of videos.
    • Ad revenue sharing: This feature uses the media fingerprint technology to monitor and identify specific ads from an ad library. This helps with ad revenue sharing in real time and provides an easier method to specify the time and frequency to deliver ads. This way, the benefits of ad producers and advertising platforms are protected. This also applies to scenarios in which you need to efficiently identify and replace specific ads or you need to sell ad placements.
    • Video source tracing: This feature retrieves a video from the media fingerprint library to determine the propagation path of the video. This helps trace video sources, analyze the propagation path of videos, and explore relationships between media resources.

References