Face Detection Using Candidate Grouping to Reduce False Positives
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing face detection methods often result in false detection of non-face images due to non-face patterns raising indicator values, leading to incorrect identification of face images.
Innovation Solution
A face detection method that cuts out partial images at different positions in an input image, calculates indicator values, groups nearby candidates with high values, and uses integrated indicator values to accurately identify face images within a predetermined area, reducing false positives by considering spatial and temporal proximity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If indicator values are calculated for partial images and a threshold value is used for determination, then detection speed is improved, but false detection of non-face images increases
Solution Approach 1:
The patent segments the detection process into multiple stages: first extracting partial images with indicator values above a first threshold, then grouping nearby candidates, and finally filtering using integrated indicator values. This segmentation allows the system to maintain speed through efficient first-stage filtering while improving accuracy through subsequent refinement stages.
Solution Approach 2:
The patent applies different threshold criteria at different stages of the detection process. The first threshold (Th1) is used for initial candidate extraction, while the second threshold (Th2) applied to integrated indicator values provides a stricter filtering criterion. This local quality approach allows the system to balance speed and accuracy by using appropriate threshold strictness at each processing stage.
2Device complexity
If a single threshold value is used for face image determination, then processing complexity is reduced, but false positive rate increases
Solution Approach 1:
The patent divides the threshold application into two segments: a first threshold (Th1) for extracting initial candidates from partial images, and a second threshold (Th2) for validating candidate groups through integrated indicator values. This segmentation reduces false positives by requiring candidates to satisfy both threshold criteria, while the modular structure keeps processing complexity manageable.
Solution Approach 2:
The patent performs preliminary action by first extracting partial images with indicator values above Th1 before applying the stricter Th2 criterion. This preliminary filtering reduces the number of images requiring full verification, thereby reducing overall processing complexity while maintaining high reliability through the two-stage threshold approach.
3Adaptability or versatility
If all partial images with high indicator values are extracted as candidates, then detection coverage is improved, but false detection increases
Solution Approach 1:
The patent merges nearby candidate images into candidate groups based on their spatial proximity and coordinate distance. By combining multiple high-indicator partial images that are close together into unified candidate groups and evaluating them through integrated indicator values, the system maintains detection coverage while reducing false detections through collective validation.
Solution Approach 2:
The patent changes the evaluation parameter from individual partial image indicator values to integrated indicator values of candidate groups. This parameter change allows the system to maintain broad detection coverage by accepting multiple high-indicator images while filtering false positives through the integrated value calculation that considers multiple candidates together.
Data Source
AI summary
To detect a face image in an inputted image, predetermined-size partial images are cut out at different positions in the inputted image. An indicator value indicating a probability of each partial image being the face image is calculated. The partial images having the indicator values not less than a first threshold are extracted as candidate face images. Each candidate is set as a candidate of interest. If any nearby candidate is present within a predetermined coordinate distance from the candidate of interest, the candidate of interest and the nearby candidate are set in one candidate group. For each candidate group, an integrated indicator value reflecting the indicator values calculated for the candidates forming the candidate group is calculated. Then, an image within a predetermined area in the inputted image containing the candidate group having the integrated indicator value not less than a second threshold is extracted as the face image.