8 minutes

The need

A marketing firm was interested in what their options were for automating the process of reviewing audio and videos posted to their online service. For example, if there was greater than x% of nudity the video should be flagged and not posted. Or if a video contained bad language it should not be posted.

The Solution

There are a number of tools available to “automate” the process of determining whether or not a written post should be deleted or flagged as unacceptable.

Automated censorship of videos can be divided into two tasks:

  • automated censorship of audio stream
  • automated censorship of video stream 

Automated censorship of video stream in its own turn could be reduced to the censorship of individual frames (with possible optimizations regarding human perception) thus it is the same as automated censorship of photos.

Audio stream area of monitoring could be bad (swearing) language.

Photos (video frames) area of monitoring could be for nudity, swear (bad) language drawings/prints, violence and other inappropriate material.

Audio Stream Monitoring

There are patented solutions for an input audio data stream comprising speech processing by an automatic censoring filter in either a real-time mode or a batch mode, producing censored speech that has been altered so that undesired words or phrases are either unintelligible or inaudible. For example, US patent 7437290 assigned to Microsoft Corporation.

Photos (Video Frames) Monitoring could be divided into a sub-set of tasks as:

  • nudity (pornography) detection
  • bad (swearing) language drawings/prints detection
  • violence detection

Detection for each of the different types of photo (video frames) monitoring can be further divided, for example, bad language detection can be divided in to:

  • image filter & image text OCR
  • running obtained with OCR words through the dictionary

The main issue here can be false-positive marks on parts of the words or foreign words. The solution to this issue is so-called “Language Guesser” systems based on categorization algorithm presented in Cavnar, W. B. and J. M. Trenkle, “N-Gram-Based Text Categorization''.

Violence detection also has its own specifics.


Products evaluated:

  • PicBlock 3 from CinchWorks
  • Illicit Image Detection from Paraben Corporation
  • SurfRecon
  • PNWatch
  • Snitch from Hyperdyne Software
  • IMAGEmanager from Clearswift

The Benefits

As part of this project, Softjourn also evaluated a number of existing software products which could be used for nudity detection and made recommendations as to which product suited our customer’s online service and what type of customization could be made to address, to the extent possible the other types of automated detection necessary.