Convolutional Neural Network Object Attribute Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video surveillance and marketing tracking systems face challenges in efficiently detecting and recognizing humans in video streams and image banks, particularly due to reliance on facial analysis in low-resolution images, leading to insufficient performance and inability to search for individuals or attribute-matched people in stored image collections.
Innovation Solution
A method utilizing a convolutional neural network trained on a diverse learning suite with both attribute-associated and unassociated objects, generating signatures for object identification and attribute determination, and storing records in a database for efficient retrieval, employing a 'centre loss' mechanism for enhanced discrimination and incorporating a Restnet50 network architecture.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If facial analysis is used for tracking humans in video streams, then identification can be performed, but performance is insufficient especially in low-resolution images
Solution Approach 1:
The patent changes the parameters used for identification from facial features alone to a comprehensive set of attributes including clothing color, body shape, and accessories. This parameter expansion allows accurate identification even when facial resolution is low, directly resolving the contradiction between identification capability and performance reliability in low-resolution scenarios.
Solution Approach 2:
The patent segments the identification process into multiple independent attribute detection components (clothing color detection, body shape analysis, accessory identification) rather than relying on a single facial recognition system. This segmentation allows each component to contribute to overall identification accuracy, improving reliability without requiring high-resolution facial images.
2Loss of information
If manual tracking is performed on video streams, then detailed analysis is possible, but it becomes difficult and expensive when video quantities are significant
Solution Approach 1:
The patent implements an automated system where the computer itself performs the complete tracking and attribute extraction process without human intervention. The system automatically detects objects, extracts attributes, stores data in databases, and retrieves information, enabling high-productivity processing of large video quantities while maintaining detailed tracking information through automated attribute extraction.
Solution Approach 2:
The patent replaces the mechanical manual tracking process with an automated computer-based system using convolutional neural networks and database technologies. This substitution eliminates the need for human operators to manually analyze each frame, dramatically increasing processing efficiency while preserving detailed tracking information through automated attribute detection and storage.
3Adaptability or versatility
If existing search tools based on facial analysis are used, then human tracking can be performed, but they cannot search in banks of previously gathered images or find people by attribute sets
Solution Approach 1:
The patent creates a universal system that performs multiple functions: real-time video stream tracking, search in image banks, and attribute-based querying. The same attribute extraction infrastructure serves all three purposes, allowing the system to adapt to different search scenarios (facial recognition, clothing color, body shape, accessories) and different data sources (live video, stored images) without requiring separate systems, thus achieving both versatility and information preservation.
Data Source
AI summary
The presently disclosed subject matter relates to a method for recognizing objects of a predefined type from among a set of types, within a set of digital images, includingdetecting an object of this predefined type within a digital image of the set, and determining a zone of the image encompassing the detected object,generating a signature by a convolutional neural network on the basis of this zone, allowing identification of the object in a one-to-one manner,determining on the basis of the signature of a set of attributes,storing in a database a record relating to the object associating the signature with the set of attributes,wherein the neural network is trained on a learning suite composed of a first set formed of objects associated with a set of attributes and of a second set formed of objects not associated with a set of attributes.