Masked Face Recognition via Segmentation-Gated Feature Refinement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current face recognition algorithms face significant challenges in achieving high accuracy when dealing with masked faces due to incomplete visual parts and insufficient identity information, as they treat all input pixels equally, including both facial and occluded regions.
Innovation Solution
The proposed solution involves a masked face recognition apparatus and method that employs guided gating convolution to differentiate between facial and masked regions, using a deep learning network that includes a segmentation model to estimate features of masked regions and refine image features by focusing on un-masked regions, thereby discarding occluded information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If all input pixels including both facial and occluded regions are treated equally, then the face recognition system can process complete face images, but the recognition accuracy dramatically drops when dealing with masked faces
Solution Approach 1:
The patent applies segmentation by dividing the face image into multiple regions: masked regions and un-masked regions. The system generates a mask map that segments the input face image to identify which regions are occluded, allowing different processing strategies to be applied to different segments. This resolves the contradiction by enabling the system to handle masked faces adaptively while maintaining accuracy through region-specific feature extraction.
Solution Approach 2:
The patent implements local quality by applying different processing weights to different regions of the face image. Instead of treating all pixels equally, the system assigns higher importance to un-masked regions and reduces or eliminates the influence of masked regions through the gating mechanism. This allows the system to maintain high recognition accuracy by focusing computational resources on reliable, visible facial features while adapting to various masking scenarios.
2Measurement precision
If a segmentation model is introduced to differentiate masked and un-masked regions, then recognition accuracy improves by focusing on valid facial features, but the device complexity increases
Solution Approach 1:
The patent merges the segmentation functionality with the existing face recognition network by integrating the mask estimation module and gating mechanism directly into the feature extraction pipeline. Rather than adding a separate, independent segmentation system, the patent combines multiple functions (feature extraction, mask estimation, and region weighting) into a unified network architecture. This reduces overall system complexity while maintaining the benefits of region-aware processing for improved recognition accuracy.
Solution Approach 2:
The patent implements universality by designing a multi-functional network that simultaneously performs face feature extraction, mask region estimation, and adaptive feature refinement. The same network architecture handles both masked and non-masked faces, as well as various types of occlusions, making the system universally applicable to different scenarios without requiring separate specialized modules for each case.
3Extent of automation
If guided gating convolution is used to force the network to learn facial regions and discard masked regions, then the system can automatically distinguish valid from occluded areas, but the computational overhead increases
Solution Approach 1:
The patent applies preliminary action by performing mask region estimation and generating the gating mask early in the processing pipeline, before the main feature extraction and classification steps. The gating mask is computed from the input image and used to guide subsequent feature extraction operations. This preliminary segmentation allows the network to automatically identify and focus on valid facial regions from the outset, reducing unnecessary computational waste on masked areas while maintaining automation.
Data Source
AI summary
Provided is a face recognition method comprising acquiring a masked face image including a masked region and an un-masked region; obtaining an image feature from the masked face image; inputting the image feature to a pre-trained segmentation model to automatically estimate a feature of the masked region; and refining the image feature using the estimated feature of the masked region, wherein the refining step comprising focusing on a feature of the un-masked region and discarding the estimated feature of the masked region.