Face Recognition Model Training with Calibrated Identity Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
General face recognition models exhibit low discrimination in image management technologies, leading to inaccurate image management due to high similarity between face features of individuals wearing glasses, even when the same person is recognized with or without glasses.
Innovation Solution
A target face recognition model is generated by training a general face recognition model using a training sample of face images calibrated with identity information, minimizing the mapping distance between face features to improve discrimination and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a general face recognition model trained on public face datasets is used, then the model can recognize face images, but the discrimination of face features is low leading to inaccurate image management
Solution Approach 1:
The patent changes the training parameters by using calibrated identity information as training targets instead of standard face recognition training. The loss function is modified to minimize mapping distance between face features and calibrated identity information, thereby improving face feature discrimination and image management accuracy simultaneously
Solution Approach 2:
The patent performs preliminary calibration of identity information for face images before training the model. This preliminary action creates a calibrated dataset that enables the model to learn more discriminative face features, resolving the contradiction between measurement precision and reliability
2Ease of operation
If face features are extracted using a general face recognition model, then face recognition can be performed, but high similarity between face features of individuals wearing glasses causes inaccurate recognition
Solution Approach 1:
The patent modifies the training target parameters to use calibrated identity information that explicitly accounts for accessories like glasses. By changing what the model optimizes for (from generic face similarity to calibrated identity matching), the model maintains face recognition capability while improving discrimination of face features even when accessories are present
Data Source
AI summary
A face recognition method includes generating a target face recognition model, performing face detection on an image to obtain a first face image, and performing face recognition on the first face image in the image by using the target face recognition model to obtain a first face feature. Generating the target face recognition model includes determining a training sample, the training sample comprising a training face image calibrated with identity information, and training a general face recognition model by using the training sample, and updating a parameter of the general face recognition model based on a training target to obtain the target face recognition model, the training target being a prediction result of the general face recognition model predicting identity information of a new training face image in the training sample to be the calibrated identity information of the training face image in the training sample.