FOLLOW
1 Reviews   |   226 Views   |   2 Followers

Video-Genre Classification

Patent

US8666918

Owner

TU Berlin

Filing Date

August 04, 2009

Priority Date

August 06, 2008

Portfolio

Video - TU Berlin

Intro

This technology enables the automatic classification of audio-visual data for search engines. The main challenge in the field of multimedia content analysis is the transformation of human interpretations of audio-visual data to correlating machine-processable representations. This invention analyses such contents with the help of high-level audio-visual descriptors and classification methods.

Claims

1. Method for classifying a video sequence (VS), characterized by analyzing the video sequence using a plurality of genre-specific detector modules (M...
  1. Method for classifying a video sequence (VS), characterized by analyzing the video sequence using a plurality of genre-specific detector modules (M1-M5), each genre-specific detector module providing a probability value (P1-P5) indicating the probability that the video sequence belongs to the genre assigned to the genre-specific detector module; and analyzing the probability values of the plurality of genre-specific detector modules using a combiner (CM) which analyzes said probability values and generates a classification signal (SC) classifying the video sequence as belonging to a specific genre (g); wherein the probability values of the plurality of genre-specific detector modules are analyzed by a first evaluating unit (EU1) of said combiner, said first evaluating unit providing a first probability vector (V1) comprising for each genre a first probability value indicating the probability that the video sequence belongs to the respective genre, and a first preliminary decision (PD1) indicating which genre the video sequence presumably belongs to, the probability values of the plurality of genre-specific detector modules are further analyzed by a second evaluating unit (EU2) of said combiner, said second evaluating unit providing a second probability vector (V2) indicating for each genre a second probability value indicating the probability that the video sequence belongs to the respective genre, and a second preliminary decision (PD2) indicating which genre the video sequence presumably belongs to, wherein the first evaluating unit and the second evaluating unit differ in their analyzing algorithm; and said classification signal is generated based on an evaluation of said first and second preliminary decisions, wherein if the first and second preliminary decision indicate the same genre, a classification signal is generated which classifies the video sequence as belonging to said same genre, wherein if the first and second preliminary decision indicate different genres, the first and second probability vectors are further analyzed and the classification signal is generated based on the result of said analysis, wherein the step of generating said classification signal includes: adding the first and second probability vectors and generating a sum vector (Vsum), each coordinate of said sum vector being assigned to a specific genre; determining the highest coordinate value of the sum vector; determining the genre which is assigned to the coordinate with the highest coordinate value; and generating a classification signal which classifies the video sequence as belonging to the genre associated with the coordinate having the highest coordinate value.
  2. The method according to claim 1, wherein the step of generating said classification signal further includes: normalizing said vector sum; comparing the highest coordinate value of the normalized sum vector (Vnorm) to a reference value (Th); and generating a classification signal that indicates an unreliable classification result if the highest coordinate value is smaller than the reference value.
  3. Method for classifying a video sequence (VS), characterized by analyzing the video sequence using a plurality of genre-specific detector modules (M1-M5), each genre-specific detector module providing a probability value (P1-P5) indicating the probability that the video sequence belongs to the genre assigned to the genre-specific detector module; and analyzing the probability values of the plurality of genre-specific detector modules using a combiner (CM) which analyzes said probability values and generates a classification signal (SC) classifying the video sequence as belonging to a specific genre (g); wherein the probability values of the plurality of genre-specific detector modules are analyzed by a first evaluating unit (EU1) of said combiner, said first evaluating unit providing a first probability vector (V1) comprising for each genre a first probability value indicating the probability that the video sequence belongs to the respective genre, and a first preliminary decision (PD1) indicating which genre the video sequence presumably belongs to, the probability values of the plurality of genre-specific detector modules are further analyzed by a second evaluating unit (EU2) of said combiner, said second evaluating unit providing a second probability vector (V2) indicating for each genre a second probability value indicating the probability that the video sequence belongs to the respective genre, and a second preliminary decision (PD2) indicating which genre the video sequence presumably belongs to, wherein the first evaluating unit and the second evaluating unit differ in their analyzing algorithm; and said classification signal is generated based on an evaluation of said first and second preliminary decisions, wherein if the first and second preliminary decision indicate the same genre, a classification signal is generated which classifies the video sequence as belonging to said same genre, wherein if the first and second preliminary decision indicate different genres, the first and second probability vectors are further analyzed and the classification signal is generated based on the result of said analysis, wherein the step of generating said classification signal further includes: adding the first and second probability vectors and generating a sum vector, each coordinate of said sum vector being assigned to a specific genre; normalizing said sum vector; determining the highest coordinate value of the normalized sum vector; comparing said highest coordinate value of the normalized sum vector to a reference value; and generating a classification signal indicating an unreliable classification result if the highest coordinate value is smaller than the reference value.
  4. Method for classifying a video sequence (VS), characterized by analyzing the video sequence using a plurality of genre-specific detector modules (M1- M5), each genre-specific detector module providing a probability value (P1-P5) indicating the probability that the video sequence belongs to the genre assigned to the genre-specific detector module; and analyzing the probability values of the plurality of genre-specific detector modules using a combiner (CM) which analyzes said probability values and generates a classification signal (SC) classifying the video sequence as belonging to a specific genre (g); wherein the probability values of the plurality of genre-specific detector modules are analyzed by a first evaluating unit (EU1) of said combiner, said first evaluating unit providing a first probability vector (V1) comprising for each genre a first probability value indicating the probability that the video sequence belongs to the respective genre, and a first preliminary decision (PD1) indicating which genre the video sequence presumably belongs to, the probability values of the plurality of genre-specific detector modules are further analyzed by a second evaluating unit (EU2) of said combiner, said second evaluating unit providing a second probability vector (V2) indicating for each genre a second probability value indicating the probability that the video sequence belongs to the respective genre, and a second preliminary decision (PD2) indicating which genre the video sequence presumably belongs to, wherein the first evaluating unit and the second evaluating unit differ in their analyzing algorithm; and said classification signal is generated based on an evaluation of said first and second preliminary decisions, wherein the first evaluating unit of said combiner calculates said first probability vector based on a given product rule, wherein the first evaluating unit of said combiner calculates the first probability vector (V1) according to the following equation: V⁢⁢1=(P⁢⁢1*(1-P⁢⁢2)*…*(1-Pi)*…*(1-Pn)(1-P⁢⁢1)*P⁢⁢2*…*(1-Pi)*…*(1-Pn)(1-P⁢⁢1)*(1-P⁢⁢2)*…*Pi*…*(1-Pn)…(1-P⁢⁢1)*(1-P⁢⁢2)*…*(1-Pi)*…*Pn) wherein Pi (1≦i≦n) defines the probability value provided by the ith genre-specific detector module associated with the ith genre, and n defines the number of genres and genre-specific detector modules.
  5. Method for classifying a video sequence (VS), characterized by analyzing the video sequence using a plurality of genre-specific detector modules (M1-M5), each genre-specific detector module providing a probability value (P1-P5) indicating the probability that the video sequence belongs to the genre assigned to the genre-specific detector module; and analyzing the probability values of the plurality of genre-specific detector modules using a combiner (CM) which analyzes said probability values and generates a classification signal (SC) classifying the video sequence as belonging to a specific genre (g); wherein the probability values of the plurality of genre-specific detector modules are analyzed by a first evaluating unit (EU1) of said combiner, said first evaluating unit providing a first probability vector (V1) comprising for each genre a first probability value indicating the probability that the video sequence belongs to the respective genre, and a first preliminary decision (PD1) indicating which genre the video sequence presumably belongs to, the probability values of the plurality of genre-specific detector modules are further analyzed by a second evaluating unit (EU2) of said combiner, said second evaluating unit providing a second probability vector (V2) indicating for each genre a second probability value indicating the probability that the video sequence belongs to the respective genre, and a second preliminary decision (PD2) indicating which genre the video sequence presumably belongs to, wherein the first evaluating unit and the second evaluating unit differ in their analyzing algorithm; and said classification signal is generated based on an evaluation of said first and second preliminary decisions, wherein the second evaluating unit of said combiner calculates said second probability vector using a support vector machine, wherein the second evaluating unit of said combiner uses a support vector machine having a Radial Basis Function, RBF, as kernel function and/or a cost parameter between 30000 and 35000 and/or a γ-value of 8. 
read more

Abstract

An exemplary embodiment of the invention relates to a method for classifying a video sequence (VS), characterized by the steps of analyzing the video ...
An exemplary embodiment of the invention relates to a method for classifying a video sequence (VS), characterized by the steps of analyzing the video sequence using a plurality of genre-specific detector modules (M1-M5), each genre-specific detector module providing a probability value (P1-P5) indicating the probability that the video sequence belongs to the genre assigned to the genre-specific detector module; and analyzing the probability values of the plurality of genre-specific detector modules using a combiner (CM) which analyzes said probability values and generates a classification signal (SC) classifying the video sequence as belonging to a specific genre (g).
read more
Write Review

Here are the reviews from the crowd!

Rating categories
WOW!
GOTCHA!
DEAL!
Where should the donation go to?
Conroy
October 08, 2018

The challenge today, is that with the growth of AI/ML is the arrival and adoption of rival algorithms. Therefore in such a climate, it is an imperative for the described techniques to demonstrate that it is superior other techniques to merit investment. Even so, it seems likely the large content providers may now have rival video genre classification solutions. In such a case, if the technique is still superior, a campaign or review via one of the many established YouTube AI/ML technical channels could significantly elevate the value of the patent.
Applications: Possible complementary use cases, are categorization of video surveillance security, (e.g. scene categorization) or video based/augment IoT scenarios (e.g. operation success or failure recognition).
Companies: It highly probable, the large video content providers may now have rival solutions using alternate techniques. However, if the patent is still technically superior to current techniques, commercial success would most likely come from a trained model with the model or API licensed or exposed to third parties respectively.

LIKE       COMMENT


Rate the patent & write your own review!

Rating categories
WOW!
GOTCHA!
DEAL!
Where should our donation go to?
Your review
Write about potential applications
Write about potentially interested companies
Choose a name

Convinced about the patent? Contact us!

Your bid
Your e-mail
Your message