Facial Recognition via Region-Based Boosted Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing facial recognition systems face challenges in accurately recognizing faces due to variations in illumination, pose, expression, occlusion, and other factors, leading to inefficient computation and reduced recognition accuracy, especially when dealing with video frames where faces change over time.
Innovation Solution
The system employs a facial tracking and recognition method that normalizes facial images, extracts feature data from selected regions using a boosting machine learning process, and combines individual two-class classifications to determine a similarity score, allowing for efficient and accurate identification of known faces in still images and videos.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If existing facial recognition systems perform comparison over extraneous facial image data, then recognition can be attempted on varied data, but computation efficiency deteriorates and recognition delays occur
Solution Approach 1:
The patent segments the face into multiple regions (e.g., eyes, nose, mouth, cheeks) and performs classification on each region independently using boosted classifiers. This segmentation allows the system to focus computation only on relevant facial features rather than processing the entire face image, thereby improving computation efficiency while maintaining recognition capability on varied data.
Solution Approach 2:
The patent extracts and selects specific discriminative regions from the full facial image data. By taking out only the most informative regions (such as eye regions, nose regions, mouth regions) and performing classification on these extracted regions, the system achieves efficient computation while maintaining adaptability to different facial variations.
2Productivity
If existing recognition techniques use strong dependence on training data and develop identification thresholds, then recognition can be performed on training data, but recognition accuracy deteriorates when applied to test data other than training data
Solution Approach 1:
The patent employs boosted classifiers that dynamically adjust decision thresholds and classification parameters based on the input data characteristics. Instead of using fixed thresholds trained on specific data, the boosted classifiers adapt their parameters to handle variations in illumination, pose, and expression, thereby maintaining high accuracy on test data while preserving efficient processing.
Solution Approach 2:
The system uses dynamic threshold adjustment and adaptive classification where the recognition thresholds are not fixed but can shift based on the confidence scores from multiple boosted classifiers. This dynamic approach allows the system to maintain reliability on unseen test data by adjusting decision criteria based on the actual data characteristics rather than relying on static training-specific thresholds.
3Measurement precision
If facial recognition systems process all facial image data, then complete face information is analyzed, but computational complexity increases especially in video frames where faces change over time
Solution Approach 1:
The patent divides the facial recognition task into multiple independent region-based classification problems. Instead of processing the entire face as one unit, the system segments the face into distinct regions (eye regions, nose regions, mouth regions, cheek regions) and applies boosted classifiers to each region independently. This segmentation dramatically reduces computational complexity while maintaining measurement precision through the combined results of region-level classifications.
Solution Approach 2:
The patent applies partial action by performing classification only on selected discriminative regions rather than processing every pixel and feature of the entire face. By focusing computational resources on the most informative regions (partial action) rather than exhaustive processing (excessive action), the system achieves high recognition accuracy with reduced computational complexity, especially important for video processing where faces change over time.
Data Source
AI summary
Methods, apparatuses, and articles associated with facial tracking and recognition are disclosed. In embodiments, facial images may be detected in video or still images and tracked. After normalization of the facial images, feature data may be extracted from selected regions of the faces to compare to associated feature data in known faces. The selected regions may be determined using a boosting machine learning processes over a set of known images. After extraction, individual two-class comparisons may be performed between corresponding feature data from regions on the tested facial images and from the known facial image. The individual two-class classifications may then be combined to determine a similarity score for the tested face and the known face. If the similarity score exceeds a threshold, an identification of the known face may be output or otherwise used. Additionally, tracking with voting may be performed on faces detected in video. After a threshold of votes is reached, a given tracked face may be associated with a known face.