Orientation-Adaptive Face Feature Detection Using a Shared Neural Network
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing face recognition and expression analysis systems struggle with accurately detecting feature points when dealing with laterally-facing individuals, as the eye and mouth corners on the far side from the screen become invisible, necessitating a switch in feature points between front-facing and lateral-facing captures, and require adaptable detection for varying attributes in non-human targets.
Innovation Solution
A detection device and method that includes an acquisition means, common feature quantity extraction, feature information detection, attribute identification, and output means, utilizing a neural network to extract common features and identify attributes across multiple orientations and targets, reducing memory size and improving accuracy through multi-task learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If separate detection models are used for different face orientations (front-facing, lateral-facing), then detection accuracy for each orientation is improved, but device complexity and memory requirements increase
Solution Approach 1:
The patent implements a single detection model that can handle multiple face orientations (front-facing, lateral-facing, and intermediate orientations) by introducing an orientation identification module that determines the face orientation and adjusts the detection process accordingly. This universal model replaces the need for separate specialized models for each orientation, reducing device complexity and memory requirements while maintaining detection accuracy across different orientations.
2Measurement precision
If multiple separate detection models are maintained for different attributes, then detection accuracy for specific attributes is improved, but memory size and processing time increase
Solution Approach 1:
The patent creates a unified detection model that processes multiple attributes (different face orientations, feature points) within a single framework. The model includes an orientation identification module that determines the face orientation and selectively applies appropriate detection parameters, eliminating the need to store multiple separate models in memory. This approach maintains high detection accuracy for specific attributes while significantly reducing memory size.
3Adaptability or versatility
If feature points are switched between front-facing and lateral-facing detection, then adaptability to different orientations is improved, but detection reliability decreases due to visibility issues
Solution Approach 1:
The patent implements a dynamic detection system that identifies the face orientation first and then adaptively selects which feature points to detect based on the identified orientation. For lateral-facing faces, the system dynamically adjusts to detect only the visible feature points (e.g., left eye and left mouth corner for right-facing faces). This dynamic adaptation maintains detection reliability by avoiding attempts to detect invisible features while preserving orientation adaptability.
4Measurement precision
If separate detection processes are used for different attributes, then detection accuracy for each attribute is improved, but processing speed decreases
Solution Approach 1:
The patent segments the detection process into two main stages: (1) orientation identification using a dedicated module that quickly determines face orientation, and (2) feature point detection using a single unified model with orientation-specific parameters. This segmentation allows the system to maintain attribute-specific detection accuracy while improving processing speed by avoiding the need to run multiple separate detection processes sequentially.
Data Source
AI summary
The detection device 1X mainly includes an acquisition means 12X, a common feature quantity extraction means 13X, a feature information detection means 14X, an attribute identification means 15X, and an output means 16X. The acquisition means 12X acquires data relating to a detection target. The common feature quantity extraction means 13X extracts, from the data, common feature quantity common to plural candidates of an attribute of the detection target. The feature information detection means 14X detects the feature information for each of the plural candidates based on the common feature quantity. The attribute identification means 15X identifies the attribute of the detection target based on the data. The output means 16X outputs the feature information corresponding to the identified attribute.