Multi-Output CNN for Simultaneous Face Identity and Attribute Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for organizing and recognizing faces in images require multiple convolutional neural networks (CNNs), increasing processing time and memory complexity, especially on mobile platforms, and are limited by small training sets and noise in data, particularly for age prediction.
Innovation Solution
A multi-output extension of a CNN with low inference and memory complexity, pre-trained on the VGGFace2 dataset for face recognition, fine-tuned for facial attribute recognition using the Audience and IMDB-Wiki datasets, employing a two-stage approach for simultaneous face identification and attribute prediction, and hierarchical agglomerative clustering for efficient grouping.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple CNNs are used for face recognition and attribute prediction, then recognition accuracy is improved, but processing time and memory complexity increase
Solution Approach 1:
The patent combines multiple separate CNNs (face recognition network and attribute prediction networks) into a single unified multi-output CNN. This single network performs both face identification and multiple attribute predictions (age, gender, race, emotions) simultaneously, reducing the number of sequential processing steps and thereby decreasing processing time while maintaining recognition accuracy through shared feature extraction layers.
Solution Approach 2:
The unified CNN is designed with multi-functionality to perform diverse facial analysis tasks through a single model. The network includes independent output branches for face recognition and multiple attribute predictions, allowing one system to replace multiple specialized systems, thus reducing memory complexity and processing overhead while maintaining the accuracy benefits of specialized networks.
2Measurement precision
If multiple CNNs are used for face recognition and attribute prediction, then recognition accuracy is improved, but device complexity increases
Solution Approach 1:
By merging multiple CNNs into a single multi-output network, the patent eliminates redundant computational resources. The shared feature extraction layers are stored only once in memory rather than being duplicated across multiple networks, significantly reducing memory complexity while preserving the specialized processing capabilities needed for accurate face recognition and attribute prediction.
3Reliability
If separate CNNs are used for face recognition and attribute prediction, then each task can be optimized independently, but processing efficiency decreases
Solution Approach 1:
The unified CNN enables continuous and simultaneous processing of face recognition and attribute prediction tasks in a single forward pass through the network. This eliminates the need for sequential processing where one task must complete before another begins, thereby maintaining independent task optimization while dramatically improving processing efficiency through parallel computation within the unified architecture.
Data Source
AI summary
A method is provided for simultaneously recognizing facial attributes and identity to organize photo and/or video albums, based on modifying an efficient convolutional neural network (CNN) which extracts facial representations suitable for face identification and attribute (age, gender, ethnicity, emotion, etc.) recognition tasks. The method enables to process all the tasks simultaneously, without a need for additional CNNs. As a result, a very fast facial analytic system is provided, and the system can be installed onto mobile devices.