Facial Feature Point Estimation via Nearest Neighbor Template Gradient
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing feature point location detection methods, such as Active Appearance Model (AAM), are prone to local solutions and struggle with high accuracy when detecting facial feature points in images with varying illuminations and postures, especially when dealing with multiple individuals.
Innovation Solution
A feature point location estimation device and method that uses nearest neighbor templates from positive and negative examples to compute a gradient direction for updating the search parameter, reducing the likelihood of local solutions and improving precision by iteratively refining the facial feature point location.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If Active Appearance Model (AAM) is used for feature point location detection, then detection accuracy is improved for few persons, but performance deteriorates remarkably when used for many persons under non-control conditions
Solution Approach 1:
The patent changes the fundamental parameters of the detection approach by switching from a statistical model-based method (AAM) to a template-based discriminative method. The system uses a template library with multiple facial images and computes feature point locations by comparing templates against the input image, fundamentally altering how detection is performed to achieve both accuracy and adaptability.
Solution Approach 2:
The patent segments the detection process into distinct components: template extraction, feature point location computation, and classification. By separating these functions and using a library of segmented facial templates, the system can handle diverse conditions more effectively while maintaining detection accuracy across different persons and scenarios.
2Productivity
If Boosting method is used for feature point detection, then processing speed is improved, but the solution becomes local and accuracy deteriorates
Solution Approach 1:
The patent transitions from one-dimensional Boosting feature selection to a multi-dimensional template-based approach. By using a library of facial templates with varying characteristics and computing distances in a multi-dimensional feature space, the system achieves both speed and accuracy, avoiding the local solution trap of traditional Boosting methods.
Solution Approach 2:
The patent uses template copying and comparison instead of iterative Boosting classification. By pre-extracting and storing facial templates as copies, then comparing these copies against the input image to compute feature point locations, the system achieves faster processing without sacrificing accuracy, as the templates provide direct geometric references rather than iterative approximations.
Data Source
AI summary
Provided is a feature point location estimation device, whereby fitting of a model is less prone to local solutions, and it is possible to detect a facial feature point with more practical precision. The feature point location estimation device comprises: an image input unit which receives input of a facial image from an external source; a nearest neighbor template computation unit which obtains a nearest neighbor template from among positive examples and a nearest neighbor template from among negative examples, the nearest neighbor templates being the templates nearest to an image that is computed from the inputted facial image using a facial feature point location search parameter from among a plurality of facial image templates constituted by the positive examples and the negative examples; a parameter gradient direction computation unit which computes, from the facial feature point location search parameter and the nearest neighbor templates, a gradient direction of the facial feature point location search parameter; and a search parameter update unit which updates the facial feature point location search parameter by adding the computed gradient direction.