Face Detection Using Orientation-Specific Threshold Adjustment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing face detection methods in digital images suffer from high false detection rates, particularly for side face images due to their smaller face area and complex contours, leading to inaccurate identification.
Innovation Solution
A face detection method that extracts face candidates by varying position, inclination, and orientation in an input image, calculating indicator values for each partial image to determine the probability of a face, and uses a discriminating step to verify true face images by adjusting threshold values based on the reliability of face candidates with similar inclinations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If separate classifiers are used for frontal faces and side faces, then detection capability for different face orientations is improved, but false detection rate increases particularly for side faces
Solution Approach 1:
The invention segments the face detection task by dividing faces into frontal faces and side faces based on their orientation characteristics. Separate classifiers are trained for each segment (frontal and side faces) to capture their distinct features. This segmentation allows the system to handle different face orientations with specialized detectors while managing false detections through orientation-aware processing.
Solution Approach 2:
Instead of treating all face detections equally, the invention inverts the conventional approach by giving special attention to side faces that tend to have higher false detection rates. When a side face candidate is detected, the system applies additional verification by checking for the presence of corresponding frontal face candidates, effectively using the minority class (frontal faces) to verify the majority class (side faces).
2Difficulty of detecting and measuring
If face detection is performed based on face features in images, then face detection capability is achieved, but false detections occur due to non-face patterns with similar features
Solution Approach 1:
The invention introduces face orientation (frontal vs. side) as an intermediary classification layer between raw image features and final face detection. By first determining the orientation type of a detected face candidate, the system can apply orientation-specific verification rules. This intermediary step helps distinguish true face features from non-face patterns that may resemble faces, reducing false detections while maintaining detection capability.
3Device complexity
If side faces are detected using the same method as frontal faces, then detection process is simplified, but detection accuracy decreases due to smaller face area and complex contours
Solution Approach 1:
The invention applies local quality by treating frontal faces and side faces differently according to their specific characteristics. Frontal faces with larger areas and simpler contours are processed with one set of parameters, while side faces with smaller areas and complex contours are processed with adjusted parameters and additional verification steps. This localized approach optimizes detection accuracy for each face type without unnecessarily complicating the overall process.
Data Source
AI summary
To detect faces contained in an image with changing position, inclination and orientation of a face to be detected, an indicator value for each position, inclination and orientation is calculated, which indicates probability of a partial image at the position being a face image having the inclination and the orientation. Then, the partial images having the indicator values not less than a first threshold are extracted as candidates. When a predetermined evaluation value indicating reliability of a certain candidate having a predetermined inclination and a first orientation is not less than a predetermined threshold value, another candidate having the predetermined inclination and a second orientation is discriminated as being a true face. When the candidates having the indicator values not less than a second threshold are discriminated as being true faces, the second threshold is lowered for another candidate having the predetermined inclination and a second orientation for discrimination.