Face Image Clustering Using GCN Noise Filtering and Graph Merging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing face image clustering algorithms struggle to handle complex clustering scenarios and often generate noisy clusters, reducing the effectiveness of face image clustering.
Innovation Solution
A face image clustering method utilizing a graph convolutional neural network (GCN) to construct a connected graph, detect low-level subgraphs, calculate quality and intersection scores, remove noise points, and merge clusters based on shared vertices, effectively handling complex distributions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If existing clustering methods (K-means, spectral clustering, hierarchical clustering) are used, then the clustering process is simple and fast, but they cannot handle complex distribution scenarios and generate noisy clusters
Solution Approach 1:
The patent segments the clustering process into multiple stages: (1) constructing a connected graph from face images, (2) extracting low-level subgraphs representing potential clusters, (3) aggregating subgraphs into candidate clusters, (4) filtering and refining clusters through GCN-based quality scoring, and (5) removing noise points and merging overlapping clusters. This multi-stage segmentation allows the system to handle complex distributions effectively while maintaining computational efficiency.
Solution Approach 2:
The patent introduces a graph convolutional neural network (GCN) as an intermediary component that bridges the gap between simple clustering algorithms and complex data distributions. The GCN processes the graph structure of face images, extracts meaningful cluster representations, and provides quality scores for filtering, enabling the system to handle complex manifolds without requiring explicit manifold assumptions.
2Device complexity
If existing clustering methods are applied to complex distribution data, then the clustering process remains straightforward, but the results contain much noise and reduce face image clustering effect
Solution Approach 1:
The patent extracts and removes noise points from candidate clusters using a GCN-based filtering mechanism. The GCN computes quality scores for each candidate cluster and identifies noise points that do not conform to the underlying manifold structure. By explicitly extracting and removing these noise points, the system produces cleaner cluster representations while maintaining the overall simplicity of the clustering framework.
Solution Approach 2:
The patent implements feedback mechanisms through quality scoring and iterative refinement. The GCN evaluates each candidate cluster's quality and provides feedback for filtering, while overlapping clusters are identified and merged based on shared vertices. This feedback loop continuously refines the cluster assignments, reducing noise and improving accuracy without significantly increasing algorithmic complexity.
3Ease of manufacture
If label-free data clustering is used to avoid annotation, then annotation cost is reduced, but clustering accuracy in complex scenarios deteriorates
Solution Approach 1:
The patent enables the clustering system to be self-service by automatically discovering cluster structures in label-free face image data. The connected graph construction and subgraph extraction processes automatically identify manifold structures without requiring external labels or annotations. The GCN-based quality scoring and noise removal further refine these automatic discoveries, achieving high accuracy while avoiding the need for manual annotation.
Solution Approach 2:
The patent changes the parameters and representation methods used in clustering. Instead of using traditional distance-based clustering parameters, the system transforms face images into graph-structured data with GCN-based features. This parameter transformation enables the system to capture complex manifold structures in label-free data, achieving accurate clustering without annotation while maintaining the ease of label-free processing.
Data Source
AI summary
A face image clustering method and apparatus, an electronic device, and a storage medium are provided. The method includes: performing feature extraction on samples in a face data set by using a face recognition model to obtain a feature; calculating a cosine distance between each two features, and constructing a connected graph covering all the samples; on the basis of a connected component, searching the connected graph to obtain low-level subgraphs, and aggregating the low-level subgraphs, to obtain first candidate clusters; calculating quality scores and intersection scores, and screening the first candidate clusters to obtain second candidate clusters; outputting probability values of vertexes in the second candidate clusters by using a graph convolutional neural network, and removing noise points to obtain third candidate clusters; and searching for a shared vertex between each of the other third candidate clusters and a reference cluster, and removing the shared vertex.