Face Recognition Model Training for Occlusion-Aware Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing models struggle with accurately recognizing faces that are partially occluded, leading to reduced accuracy and increased computational requirements due to reliance on external networks for occlusion detection.
Innovation Solution
An image processing model is trained to determine occlusion patterns directly from input images, using occlusion indication information to update its parameters, enabling end-to-end face recognition without external network assistance, thus improving accuracy and reducing computational load.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If external networks are used for occlusion detection, then occlusion detection capability is improved, but device complexity and computational requirements increase
Solution Approach 1:
The patent merges the occlusion detection function with the face recognition model by integrating an occlusion detection layer into the neural network architecture. This allows the model to simultaneously perform feature extraction and occlusion detection without requiring separate external networks, thereby reducing device complexity while maintaining reliable occlusion detection capability.
Solution Approach 2:
The face recognition model is designed to serve multiple functions: it extracts face features for recognition and simultaneously detects occlusion patterns. This multi-functional design eliminates the need for dedicated external occlusion detection networks, reducing overall system complexity while maintaining both recognition and occlusion detection capabilities.
2Speed
If traditional face recognition methods are used, then processing speed is maintained, but recognition accuracy on occluded faces deteriorates
Solution Approach 1:
The model performs preliminary occlusion detection and feature filtering during the feature extraction phase, before final recognition decisions are made. By identifying and appropriately handling occluded regions early in the processing pipeline, the model maintains fast processing speed while improving recognition accuracy on occluded faces.
Solution Approach 2:
The patent modifies the neural network's parameter space by adding occlusion detection capabilities to the feature extraction process. This allows the model to adaptively adjust feature weights based on occlusion patterns, improving recognition accuracy without significantly increasing processing time due to the integrated design.
3Reliability
If occlusion patterns are learned dynamically, then robustness to occluded faces is improved, but training complexity increases
Solution Approach 1:
The training process merges occlusion pattern learning with standard face recognition training by using a unified loss function that combines recognition accuracy and occlusion detection accuracy. This integrated approach allows the model to learn both capabilities simultaneously without requiring separate training phases or complex multi-stage training procedures.
Solution Approach 2:
The model incorporates feedback mechanisms during training where occlusion detection results are used to adjust feature extraction parameters and recognition decisions. This feedback loop enables the model to dynamically adapt to occlusion patterns during both training and inference, improving robustness while maintaining manageable training complexity through gradient-based optimization.
Data Source
AI summary
An image processing model can more accurately process a face image with an occluded face while reducing calculations and improving operation speed of a processing device, reducing training time and costs. A predicted recognition result of a sample face image and occlusion indication information based on an image processing model is obtained. The occlusion indication information indicates an image feature of a face occlusion area of the sample face image. A recognition error based on the predicted recognition result and a target recognition result is also obtained. A classification error is obtained based on the occlusion indication information and a target occlusion pattern corresponding to the sample face image. An occlusion pattern of the sample face image indicates a position and a size of the face occlusion area. A model parameter of the image processing model is updated based on the recognition error and the classification error.