AU Feature Recognition via Facial Mean Shape and Local SVM Classifiers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current AU feature recognition methods in facial emotion recognition, primarily relying on convolutional neural networks, are not sufficiently accurate due to the complexity of facial expressions and the need for extensive sample collection.
Innovation Solution
An AU feature recognition method and device that utilizes a combination of facial mean shape recognition, local feature extraction, and support vector machine classifiers to identify AU features in real-time facial images, improving recognition efficiency by cropping feature regions and using pretrained classifiers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If convolutional neural network is used for AU feature recognition, then recognition capability is provided, but recognition accuracy is insufficient
Solution Approach 1:
The patent divides the facial image into multiple feature regions corresponding to different AU types (e.g., eyebrow regions for AUs 1-3, eye regions for AUs 4-7, nose regions for AUs 9-17, mouth regions for AUs 20-26). Each region is processed by a dedicated classifier, enabling localized feature recognition that improves accuracy while maintaining manageable complexity through functional decomposition.
Solution Approach 2:
The system performs preliminary actions by pre-training multiple specialized classifiers (SVM, GMM, or CNN models) for different AU types before actual recognition. Facial feature points are detected and feature regions are segmented in advance, so that when a real-time facial image is input, the pre-prepared classifiers can directly process the segmented regions without requiring complex real-time model training.
2Measurement precision
If extensive AU samples are collected and model training is performed, then recognition model is established, but recognition accuracy remains insufficient
Solution Approach 1:
Instead of using a single general model trained on all AU samples, the patent creates multiple specialized classifiers, each optimized for specific AU types and their corresponding facial regions. Each classifier is trained on local feature data from its specific region (e.g., eyebrow curvature features for eyebrow AUs, lip contour features for mouth AUs), enabling more accurate local feature recognition without requiring exhaustive global sampling.
3Productivity
If real-time facial image processing is performed, then recognition speed is improved, but recognition accuracy may be compromised
Solution Approach 1:
The patent segments the facial image into multiple independent feature regions corresponding to different AU types. Each segmented region is processed by a dedicated pre-trained classifier, enabling parallel processing that maintains real-time speed while improving accuracy through specialized local analysis. This segmentation allows the system to focus computational resources on specific facial regions rather than processing the entire image uniformly.
Solution Approach 2:
Multiple classifiers for different AU types are pre-trained in advance on training sets. When a real-time facial image is input, the system directly applies these pre-trained classifiers to the segmented feature regions without requiring real-time model training or complex iterative optimization, thus achieving both real-time processing speed and accurate recognition results.
Data Source
AI summary
The disclosure discloses an action unit (AU) feature recognition method, which includes: acquiring a real-time image shot by a photographic device and extracting a real-time facial image from the real-time image by use of a face recognition algorithm; inputting the real-time facial image into a pretrained facial mean shape and recognizing t facial feature points from the real-time facial image by use of the facial mean shape; determining feature regions matched with each AU in the real-time facial image according to positions of the t facial feature points, extracting local features from the feature regions and generating multiple feature vectors; and inputting the multiple feature vectors into pretrained AU classifiers matched with the feature regions respectively to obtain a prediction result of recognition of the corresponding AU features from the feature regions. The disclosure also discloses an electronic device and a computer-readable storage medium.