Facial Feature Detection Using Segmented Neural Learners for Masked Faces
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing technologies fail to accurately detect the positions of facial components from images of individuals wearing masks or sunglasses, as the face is partially covered, leading to insufficient node detection and estimation challenges.
Innovation Solution
An image processing apparatus comprising multiple learners trained using neural networks, where a first learner detects the orientation of a face, and corresponding second and third learners detect the position and orientation of facial components, with a cover determination unit controlling the learners to adapt to partially covered faces, enhancing detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional image processing techniques are used to detect facial features, then the system is simple and fast, but the detection accuracy deteriorates when faces are partially covered by masks or sunglasses
Solution Approach 1:
The patent divides the face detection task into multiple specialized learners: a first learner for detecting face orientation and second learners for detecting positions of specific facial components (eyes, nose, mouth). Each learner is trained for specific conditions (covered or uncovered faces), allowing accurate detection even when parts of the face are obscured by masks or sunglasses.
Solution Approach 2:
The system dynamically selects which second learner to use based on the output from the first learner. When a face is detected as covered, the system selects learners trained for covered faces; when uncovered, it selects learners trained for uncovered faces. This dynamic adaptation maintains high detection accuracy across varying conditions without requiring a completely different system for each scenario.
2Measurement precision
If multiple learners trained for different orientations are used, then detection accuracy for covered faces improves, but processing time increases
Solution Approach 1:
The first learner performs preliminary detection of face orientation and determines whether the face is covered before the second learners perform detailed facial component detection. This preliminary action allows the system to quickly assess the situation and select the appropriate pre-trained learner, reducing the time required for accurate detection compared to using a single general-purpose learner that would need to process all possibilities.
Solution Approach 2:
Different learners are trained with different qualities and characteristics suited to specific conditions. The first learner focuses on orientation detection, while second learners are specialized for detecting specific facial components under specific conditions (covered or uncovered). This local specialization allows each learner to be highly efficient at its specific task, reducing overall processing time while maintaining high accuracy.
3Device complexity
If a single learner is used for all face orientations, then the device complexity is low, but the detection accuracy deteriorates for specific orientations and covered faces
Solution Approach 1:
The first learner serves multiple functions: it detects face orientation and determines whether the face is covered, providing input for selecting the appropriate second learner. This multi-functionality reduces the need for separate specialized learners for every possible condition, maintaining reasonable system complexity while improving accuracy through conditional specialization.
Data Source
AI summary
An image processing apparatus includes a first learner trained to detect an orientation of a face in an image including a human face for receiving an input of a target image including a human face and outputting first information associated with an orientation of a face included in the target image, at least one second learner trained, for the first information, to detect a position of a facial component in an image including a human face for receiving an input of the target image and outputting second information associated with a position of a facial component included in the target image, and a control unit that controls the first learner and the second learner. The control unit uses the first learner to obtain the first information from the target image, and uses the second learner corresponding to the first information to obtain the second information from the target image.