Face Authentication Glasses Occlusion Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing human face authentication technologies are ineffective in accurately verifying identities when comparing photographs with and without glasses, due to poor feature extraction and verification results caused by occlusion and diversity in glasses styles.
Innovation Solution
A human face authentication method that integrates a glasses segmentation algorithm and a glasses photo feature regression network, allowing for accurate feature extraction and comparison by modifying pixel values in images with glasses, using a glasses segmentation model to identify occluded regions and adjusting the authentication model accordingly.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If glasses removal technology is used to restore images without glasses, then verification accuracy improves for thin-frame glasses, but it fails to truly recover face area for thick-rimmed glasses and destroys identity information when relying on others' non-glasses photos
Solution Approach 1:
The patent segments the face image into multiple regions: occluded regions (behind glasses), partially occluded regions, and non-occluded regions. By separately processing and weighting these regions, the system avoids the need for complete glasses removal while preserving identity information from non-occluded areas and appropriately handling occluded areas through regression prediction.
Solution Approach 2:
The patent applies different processing strategies to different regions of the face image. Non-occluded regions are used directly for verification, partially occluded regions undergo regression prediction, and heavily occluded regions are weighted differently. This local differentiation resolves the contradiction by preserving identity information where available while making reasonable predictions where occlusion occurs.
2Adaptability or versatility
If dedicated face verification algorithm is trained using pairs of ID card photos and self-portrait photos with glasses, then verification for glasses wearers improves, but insufficient training samples due to diverse glasses styles result in poor verification results
Solution Approach 1:
The patent changes the approach from collecting diverse real glasses photos to generating synthetic training data by applying occlusion masks to existing non-glasses photos. This parameter change in data generation methodology allows creating unlimited training samples with various occlusion patterns, resolving the contradiction between sample quantity and glasses style diversity.
Solution Approach 2:
The patent performs preliminary occlusion processing on training photos to create synthetic glasses-wearing samples before actual verification. By pre-processing training data with various occlusion patterns, the system prepares comprehensive training examples without needing to collect diverse real-world glasses photos, thus solving the sample quantity problem.
3Quantity of substance
If face verification algorithm directly processes photos with and without glasses mixed together, then large amount of training samples can be provided, but learning accuracy decreases and verification results become unreliable
Solution Approach 1:
The patent segments the training process by creating separate processing paths for occluded and non-occluded regions. During training, the system learns to handle occluded regions through regression prediction while maintaining accurate feature extraction from non-occluded regions. This segmentation allows using large mixed datasets without degrading overall learning accuracy.
Solution Approach 2:
The patent introduces an intermediary regression prediction module that handles occluded regions. This intermediary component predicts the appearance of occluded areas based on visible regions, allowing the system to process mixed photos with and without glasses while maintaining feature extraction accuracy through the mediation of the regression model.
Data Source
AI summary
A device receives an image-based authentication request from a specified object and performs human face authentication in a manner depending on whether the object wears glasses. Specifically, the device designates a glasses region on a daily photograph of the specified object using a glasses segmentation model. If the regions of the human face in the daily photograph labeled as glasses exceed a first threshold amount, the device modifies the daily photograph by changing pixel values of the regions that are labeled as being obscured by glasses. The device extracts features of a daily human face from the daily photograph and features of an identification human face from the identification photograph. The device approves the authentication request if a matching degree between the features of the daily human face and the features of the identification human face is greater than a second threshold amount.