Face Detection Using Region Prediction and Adaptive Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current face detection algorithms are inefficient, particularly in grayscale images, due to the need for sequential matching of different size templates across the entire image, which is time-consuming and not suitable for both color and grayscale images.
Innovation Solution
A system and method for face detection using face region location and size predictions, involving image regionalization, region labeling, feature capturing, distance transform with constraint, and adaptive face detection to predict face locations and sizes, reducing unnecessary matching by dynamically determining optimal face locations for each scale.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If sequential matching of different size templates is used across the entire image, then face detection accuracy is maintained, but detection speed deteriorates significantly
Solution Approach 1:
The patent applies preliminary action by performing image regionalization and distance transform before template matching to predict potential face locations. This pre-processing step identifies candidate regions where faces are likely to exist, allowing the subsequent template matching to be performed only in these restricted areas rather than across the entire image, thus maintaining accuracy while improving speed
Solution Approach 2:
The patent segments the image processing task into distinct stages: image regionalization to divide the image into meaningful regions, distance transform to calculate distance maps, candidate location prediction to identify potential face positions, and finally template matching to verify faces. This segmentation allows each stage to contribute to filtering out non-face regions, reducing the computational burden of exhaustive template matching
2Reliability
If different size templates are used to scan the entire image, then comprehensive face detection is achieved, but computational complexity increases
Solution Approach 1:
The patent applies local quality by making the template matching process adaptive to local image characteristics. The distance transform and regionalization identify regions with different properties (potential face regions vs. non-face regions), and the algorithm applies template matching selectively based on these local characteristics. This allows comprehensive detection in relevant regions while avoiding unnecessary computations in irrelevant regions
Solution Approach 2:
The patent performs preliminary image regionalization and distance transform to create a structured representation of the image before template matching. This preliminary action organizes the computational work by identifying candidate regions upfront, allowing the system to maintain reliability through thorough searching in relevant areas while reducing computational complexity by excluding irrelevant areas from exhaustive template matching
3Productivity
If skin filter and sequential search are used for face detection, then detection speed is improved, but applicability to grayscale images deteriorates
Solution Approach 1:
The patent applies universality by designing an image regionalization and distance transform approach that works uniformly across different image types (color and grayscale). Unlike skin filter methods that rely on color information, the patent's method uses structural and spatial features that are independent of color, making it universally applicable to both color and grayscale images while maintaining improved detection speed through candidate region prediction
Data Source
AI summary
In one exemplary embodiment, a face detection method implemented in a computer system estimates location and size of face candidate regions for an input image, which includes image regionalization, region labeling and feature capturing, distance transform, and face location and size estimation, to predict at least a face location and size range for the input image. According to the at least a predicted face location and size range, the method further performs an adaptive face detection until a predicted face region is verified.