Face Detection Using Hierarchical Neighborhood Re-examination
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing methods for face detection in images face challenges in achieving both high precision and short computation time, as they are influenced by changes in object posture, illumination conditions, and noise, leading to impaired detection precision even when combining feature detection and neural network-based approaches.
Innovation Solution
An image processing method that detects feature candidate regions, determines object candidate regions, sets neighboring regions for re-examination, and outputs data on object candidate regions, utilizing a combination of feature detection and neural networks to improve detection precision and speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If feature detection method is used to detect face based on geometric positional relationship among features, then detection speed is improved, but detection precision deteriorates under changes in posture, illumination conditions, and noise
Solution Approach 1:
The patent segments the face detection process into multiple stages: initial feature detection (eyes, nose, mouth) to locate candidate regions, followed by refined detection using neighboring regions to correct positional deviations. This segmentation allows the system to achieve both speed (through initial coarse detection) and precision (through subsequent refinement steps).
Solution Approach 2:
The patent performs preliminary feature detection to identify candidate regions before conducting the final precise detection. By pre-locating features such as eyes and mouth, the system narrows down the search space, enabling faster subsequent processing while maintaining high detection precision through the preliminary positioning information.
2Measurement precision
If neural network-based face detection is used to achieve robust detection against posture and illumination changes, then detection precision is improved, but computation time increases significantly
Solution Approach 1:
The patent divides the detection process into an initial fast feature detection stage and a subsequent refined detection stage. The initial stage quickly identifies candidate regions using simple geometric relationships, while the refined stage applies more computationally intensive methods only to these limited candidate regions, thereby reducing overall computation time while maintaining high precision.
Solution Approach 2:
Instead of applying complex detection algorithms to the entire image, the patent applies refined detection methods only to specific neighboring regions around detected features. This partial application of intensive processing to only the necessary areas reduces computation time while ensuring high detection precision where it matters most.
3Measurement precision
If combination of feature detection and neural networks is used, then detection precision is improved, but detection speed deteriorates
Solution Approach 1:
The patent segments the combined detection approach into hierarchical stages: initial feature detection for rapid candidate identification, followed by refined detection in neighboring regions. This segmentation ensures that the more time-consuming refined detection is applied only to limited candidate regions rather than the entire image, maintaining detection speed while improving precision.
Solution Approach 2:
The patent performs preliminary feature detection to identify and narrow down candidate regions before applying refined detection methods. This preliminary action reduces the search space, allowing the combination of feature detection and neural networks to achieve high precision without sacrificing detection speed, as the intensive processing is confined to pre-identified candidate areas.
Data Source
AI summary
A luminance image which is formed by luminance components of an image is generated, and a plurality of reduced-scale images by recursively reducing the luminance image are generated. Positions of eye candidate regions from each of the reduced-scale images are detected, it is determined that whether or not a face candidate region which is specified by positions of the detected two eye candidate regions is a face pattern. If the face candidate region is determined not as the face pattern, neighboring regions in the neighborhood of the positions of the eye candidate regions in the face region are set. It is determined whether or not a face candidate region which is specified by a first position in one neighboring region and a second position in the other neighboring region as positions of eye candidate regions, is the face pattern.