Face Recognition via Random Path Similarity and Patch Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current face recognition methods struggle with large intra-personal variations such as pose, illumination, and expression, and are sensitive to noise and outliers, failing to effectively capture structural information for robust discrimination.
Innovation Solution
The introduction of the Random Path (RP) measure and the use of two novel face patch networks, the in-face network and the out-face network, which segment faces into overlapping patches to form KNN graphs, allowing for the calculation of patch similarities and capturing both local and global structural information to improve recognition performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If classical measurement approaches use pair wise distances to compute similarities, then the computation is simple, but they cannot capture structural information for high-quality discrimination
Solution Approach 1:
The face image is divided into multiple overlapping patches, and a KNN graph is constructed where each node represents a patch and edges represent similarity relationships. This segmentation allows the method to capture local structural information while maintaining computational feasibility through graph-based measurements.
Solution Approach 2:
The patent introduces the Random Path measure as an intermediary that computes similarity through all paths in the KNN graph rather than direct pairwise distances. This mediator captures global structural information by integrating contributions from all possible paths between patches, significantly improving discrimination accuracy.
2Measurement precision
If some studies apply structural information in measurements, then discrimination quality improves, but the algorithms become sensitive to noise and outliers
Solution Approach 1:
The patent converts the harmful effect of noise and outliers into a benefit by using the Random Path measure that integrates contributions from all paths in the graph. Noisy edges (outliers) naturally receive lower weights in the path integration, allowing the method to robustly capture structural information while automatically down-weighting noisy connections through the probabilistic path sampling process.
Data Source
AI summary
A method and a system for recognizing faces have been disclosed. The method may comprise: retrieving a pair of face images; segmenting each of the retrieved face images into a plurality of image patches, wherein each patch in one image and a corresponding one in the other image form a pair of patches; determining a first similarity of each pair of patches; determining, from all pair of patches, a second similarity of the pair of face images; and fusing the first similarity determined for the each pair of patches and the second similarity determined for the pair of face images.