Robust image clustering via context-aware contrastive graph learning. (June 2023)

• Record Type: Journal Article
• Authors: Fang, Uno; Li, Jianxin; Lu, Xuequan; Mian, Ajmal; Gu, Zhaoquan
• Is Part Of: Pattern recognition. Volume 138(2023)
• Publication Date: 2023-06
• Subjects: Supervised clustering -- Graph convolution network -- Contrastive graph learning -- Graph view generation; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2023.109340
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We introduce a novel context-aware clustering framework via contrastive graph learning, which reasons intra-class relationships and intra-class boundaries. We devise an effective algorithm to construct influential graph views (IGVs) and topological graph views (TGVs) to form the inputs to contrastive learning. We design a loss function that modifies terms for inter-view and intraview while considering nodes residing on cluster boundaries, to facilitate the node-level and the graph-level contrasting. Graphical abstract: We proposed a context-aware contrastive graph learning framework (CA-CGL). In CA-CGL, firstly, a pre-trained CNN is used to extract image feature vectors and their dimensionality is reduced to 2 using t-SNE (t-distributed stochastic neighbour embedding). Then, the proposed algorithm finds furthest nodes (FNs) and local densest groups (LDGs) in each class. Based on FNs and LDGs of each class, we find confusers (i.e. nearest nodes from other class) and generate two graph views, IGV and TGV, which are fully connected with edge labels indicating factual statements of linkages. Then, every pair of graph views is contrasted in a shared Graph Convolutional Network (GCN) to learn representations and relationships and is contrasted to maximise mutual information. This makes the linkages reasonable, and the cluster boundaries clear. During testing, the data is established as a fully connected graph. Through the trained GCN, linkage predictions are estimated to consolidate the most possible linkages, and transitively incorporate linked nodes to define the final clusters. Abstract: Graph convolution networks (GCN) have recently become popular for image clustering. However, existing GCN-based image clustering techniques focus on learning image neighbourhoods which leads to poor reasoning on the cluster boundaries. To address this challenge, we propose a supervised image clustering approach based on contrastive graph learning (CGL). Our method generates an influential graph view (IGV) and a topological graph view (TGV) for each class to represent its global context from different viewpoints. These generated graph views are used to reason the inter-cluster relationships and intra-cluster boundaries from the local context of each node in a contrastive manner. Our method considers each class as a fully connected graph to explore its characteristics and strategically generate directional graph views. This enhances the transferability of the proposed approach to handle data with a similar structure. We conduct extensive experiments on open datasets such as LFW, CASIA-WebFace, and CIFAR-10 and show that our method outperforms state-of-the-art including deep GRAph Contrastive rEpresentation learning (GRACE), GraphCL, and Graph Contrastive Clustering (GCC).