Video Concept Recognition via Classifier Combination
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automatic concept labeling methods for digital videos are limited by their reliance on predefined taxonomies and user-supplied metadata, which can be incomplete, inaccurate, or intentionally false, failing to capture the richness and diversity of video concepts, especially when new events occur.
Innovation Solution
A computer-implemented method that accesses concepts derived from textual metadata, uses content-based classifiers to score video likelihood for each concept, and trains text-based classifiers based on identified training sets, combining scores from different feature types to improve accuracy and recognize diverse concepts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If predefined taxonomies and user-supplied metadata are used for concept labeling, then the labeling process is simple and fast, but the accuracy and completeness of concept recognition deteriorates due to incomplete, inaccurate, or intentionally false metadata
Solution Approach 1:
The patent introduces an intermediary classification system that bridges user-supplied metadata and final concept labeling. Multiple classifiers (content-based, text-based, and hybrid) act as intermediaries to verify and correct metadata inaccuracies, thereby maintaining labeling speed while improving recognition accuracy through layered verification
Solution Approach 2:
The patent employs composite classification by combining multiple classifier types (content-based classifiers analyzing video/audio features, text-based classifiers analyzing metadata, and hybrid classifiers combining both). This composite approach leverages the strengths of each classifier type to achieve high accuracy while maintaining efficiency through specialized division of labor
2Ease of manufacture
If a manually-specified taxonomy of concepts is used, then the labeling system is easy to implement, but it fails to capture the full richness and diversity of concepts in large video corpuses
Solution Approach 1:
The patent implements dynamic concept discovery where the taxonomy is not static but continuously adapted based on video content analysis. The system dynamically identifies new concepts from video corpora and updates the taxonomy structure, allowing it to evolve with diverse content while maintaining implementation feasibility through automated learning processes
Solution Approach 2:
The patent creates a universal classification framework that can handle both predefined concepts and newly discovered concepts through the same infrastructure. The hybrid classifier system universally processes multiple concept types (predefined taxonomy concepts and dynamically discovered concepts) using unified algorithms, enabling the system to adapt to diverse concepts without requiring separate implementation mechanisms
3Loss of time
If user-supplied metadata is accepted without verification, then the processing is fast and simple, but the labeling quality deteriorates due to unintentional errors and intentional false information
Solution Approach 1:
The patent implements feedback mechanisms where classifier outputs are used to verify and correct input metadata. The system provides feedback loops where initial classification results inform subsequent verification steps, allowing the system to identify and correct metadata errors while maintaining efficient processing through intelligent feedback routing that only processes suspicious cases in detail
Solution Approach 2:
The patent performs preliminary classification using content-based features before final labeling, creating an early verification layer that identifies potential metadata inaccuracies before they affect final results. This preliminary action filters out obviously incorrect metadata early in the process, reducing the need for extensive verification while maintaining high reliability
Data Source
AI summary
A classifier learning module trains video classifiers associated with a stored set of concepts derived from textual metadata of a plurality of videos. Specifically, a first type of classifier (e.g., a content-based classifier) and a second type of classifier (e.g., a text-based classifier) are trained, the classifiers when applied to a video indicating a likelihood that the video represents one or more concepts corresponding to the classifier. The first type of classifier can be used to determine the training set for the second type of classifier. The learning process does not require any concepts to be known a priori, nor does it require a training set of videos having training labels manually applied by human experts. Scores from the first type of classifier are combined with scores from the second type of classifier to obtain video classification of enhanced accuracy.