Face Recognition Projection Matrix Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing face recognition technologies face low recognition efficiency and accuracy due to the need to determine category information and significant differences between intra-class samples from the same user, especially in uncontrolled environments.
Innovation Solution
A method involving the use of a preset projection matrix formed by training with a positive and negative template image feature pair, which projects client and template image features to generate a projection feature pair, followed by similarity calculation using a preset similarity calculation rule to improve recognition efficiency and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If face image recognition is performed by calculating similarity between all samples to determine user category, then comprehensive recognition can be achieved, but recognition efficiency is low
Solution Approach 1:
The patent segments the recognition process into two distinct modes: verification mode (one-to-one comparison) and recognition mode (one-to-many comparison). By defaulting to verification mode which only compares test images against a single template, the system achieves high efficiency while maintaining accuracy when needed. This segmentation eliminates the need to calculate similarity between all samples in every case.
Solution Approach 2:
The patent implements partial action by performing only necessary similarity calculations. In verification mode, only one template comparison is performed instead of comparing against all samples. The system performs minimal necessary computations (partial action) to achieve the recognition goal, avoiding excessive calculations that would reduce efficiency.
2Measurement precision
If PCLDA is used to obtain identifiable features by maximizing inter-class covariance and minimizing intra-class covariance, then feature discrimination is improved, but recognition accuracy is low under uncontrolled environments due to big differences between intra-class samples
Solution Approach 1:
The patent changes the fundamental parameter of feature extraction from PCLDA (which optimizes for category discrimination) to a verification-based feature extraction method that optimizes for identity matching. The system uses verification loss functions and similarity metrics that are specifically designed to handle variations within the same class, thereby improving recognition accuracy under uncontrolled environments while maintaining feature discrimination capability.
3Adaptability or versatility
If category information is determined for training samples to perform LDA, then classification capability is enhanced, but training complexity and time consumption increase
Solution Approach 1:
The patent performs preliminary action by pre-computing verification-specific features and storing template images during registration. The system prepares verification templates in advance with features extracted using verification-oriented methods, so that during operation, only simple similarity comparisons are needed. This preliminary preparation eliminates the need for complex category-based LDA training during deployment.
Solution Approach 2:
Instead of using category information to train LDA classifiers (traditional approach), the patent inverts the approach by using verification pairs (same-person positives and different-person negatives) to train verification models. This inversion eliminates the need for category labels during training while achieving better performance for verification tasks, reducing training complexity and time.
Data Source
AI summary
In the present disclosure, a client feature and a pre-stored template image feature are obtained; the obtained client feature and template image feature are projected according to a preset projection matrix, to generate a projection feature pair, where the projection matrix is formed by training of a first template image feature of a same object and a second template image feature of a different object; and similarity calculation is performed on the projection feature pair according to a preset similarity calculation rule, to generate a similarity result and prompt the similarity result to a client.