Robust Orthogonal NMF with Label Propagation for Image Clustering

Jingjing Liu, Nian Wu, Xianchao Xiu, Jianhua Zhang · 2025

Read Robust Orthogonal NMF with Label Propagation for Image Clustering by Jingjing Liu, Nian Wu, Xianchao Xiu, Jianhua Zhang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-21472.txt.

Introduction

Jingjing Liu, Nian Wu, Xianchao Xiu, and Jianhua Zhang present a tougher way to sort images into groups without labels, in a 2025 paper. Their base method splits an image collection into two simple parts whose product rebuilds the data. Noise in real photo sets can wreck that split, and a few known labels often go unused. Their version hardens both weak spots at once.

The design adds three guards. A graph term keeps near images near in the split. A label-spreading term lets the few known labels teach the rest. A stricter error measure and a neatness rule on one part blunt the effect of noise. A standard splitting algorithm solves the whole thing, with each sub-step in closed form. On eight public image sets, the method beat top rivals and held up under dirtied data. Code is promised in public.

Catalog record

CREATORJingjing Liu, Nian Wu, Xianchao Xiu, Jianhua Zhang
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE54 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-21472

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Robust Orthogonal NMF with Label Propagation for Image Clustering Jingjing Liu, Nian Wu, Xianchao Xiu, Jianhua Zhang Abstract Non-negative matrix factorization (NMF) is a popular unsupervised learning approach widely used in image clustering. However, in real-world clustering scenarios, most existing NMF methods are highly sensitive to noise corruption and are unable to effectively leverage limited supervised information. To overcome these drawbacks, we propose a unified non-convex framework with label propagation called robust orthogonal nonnegative matrix factorization (RONMF). This method not only considers the graph Laplacian and label propagation as regularization terms but also introduces a more effective non-convex structure to measure the reconstruction error and imposes orthogonal constraints on the basis matrix to reduce the noise corruption, thereby achieving higher robustness. To solve RONMF, we develop an alternating direction method of multipliers (ADMM)-based optimization algorithm. In particular, all subproblems have closed-form solutions, which ensures its efficiency. Experimental evaluations on eight public image datasets demonstrate that the proposed RONMF outperforms state-of-the-art NMF methods across various standard metrics and shows excellent robustness. The code will be available at https://github.com/slinda-liu . keywords Image clustering Non-negative matrix factorization (NMF) Label propagation Non-convex optimization Orthogonal constraint † † corresponding: Corresponding author mode = title¡Robust Orthogonal NMF with Label Propagation for Image Clustering¿ 1 Introduction Image clustering plays an important role in many areas of image and video processing. However, high-dimensional data often contains a large number of redundant features and noise, which makes the clustering performance unreliable. Therefore, image dimensionality reduction has become a key step in clustering. To this end, various dimensionality reduction techniques have been developed, such as principal component analysis (PCA) [ 1 , 2 ] , canonical correlation analysis (CCA) [ 3 , 4 ] , and non-negative matrix factorization (NMF) [ 5 , 6 ] . It is worth noting that compared with PCA and CCA, NMF not only retains the inter-class structure of the original data but also considers the non-negativity of the actual problem [ 7 ] . At present, NMF has been extensively applied in multi-view clustering [ 8 ] , biomedical science [ 9 ] , hyperspectral unmixing [ 10 ] , community detection [ 11 ] . Interested readers can refer to [ 12 ] . During the past few decades, numerous NMF variants have been proposed. For example, to deal with data distribution, Wang et al. [ 13 ] introduced a novel graph-based transductive classification technique [ 14 ] to NMF termed linear neighborhood propagation (LNP). Subsequently, Cai et al. [ 15 ] proposed graph Laplacian regularized NMF (GNMF) by representing geometric data through the nearest neighbor graph. In addition, Guan et al. [ 16 ] proposed a truncated Cauchy NMF framework and solved the non-quadratic objective through semi-quadratic programming. Leng et al. [ 17 ] developed the graph Laplacian regularized smooth NMF (GSNMF), which combined \ell_{p} -norm smoothing constraints with graph regularization. Note that here 1<p<2 , and it is obviously a convex function between \ell_{1} -norm and \ell_{2} -norm. Recently, Wang et al. [ 18 ] considered a robust random graph regularized matrix factorization method to achieve more effective data clustering. In practical scenarios, the high costs associated with sample labeling for high-dimensional data present a significant challenge. To address this issue, one can utilize a small portion of labeled data to improve image clustering performance [ 19 ] . Liu et al. [ 20 ] proposed to enhance discrimination by incorporating label information into NMF, leading to the development of constrained NMF (CNMF). Li et al. [ 21 ] introduced class-driven concept decomposition (CDCF), which incorporates class-driven constraints to link the class labels of data points with their feature representations. Building on this idea, Jia et al. [ 22 ] proposed integrating complementary regularizers into the traditional NMF framework, resulting in a method called SNMFDSR. Later, Lan et al. [ 23 ] combined semi-supervised learning with NMF to create an image clustering method that leverages label propagation. Extending these advancements, Liu et al. [ 24 ] introduced label propagation-constrained NMF (LpCNMF), an efficient variant that incorporates label information from both labeled and unlabeled data. This method also uses the intrinsic geometric structure of the data as a regularization term, effectively constraining the model. Note that the aforementioned methods typically use the Frobenius norm to measure reconstruction error, which can be sensitive to noise. To solve the above problems, Kong et al. [ 25 ] adopted \ell_{2,1} -norm [ 26 ] to measure the reconstruction error to enhance the robustness. Similar ideas are also considered in [ 27 , 28 ] . Furthermore, Li et al. [ 29 ] introduced robust structured NMF (RSNMF) to explore block diagonal structures by replacing \ell_{2,1} -norm with \ell_{2,p} -norm (0<p<1) . In fact, compared with convex functions, non-convex functions can achieve better sparsity, thereby achieving better feature extraction and data dimensionality reduction. Thanks to the rapid development of sparse optimization [ 30 , 31 , 32 ] , one can employ various non-convex functions such as capped- \ell_{1} [ 33 ] , minimax concave penalty (MCP) [ 34 ] , smoothly clipped absolute deviation (SCAD) [ 35 ] , and exponential-type penalty (ETP) [ 36 ] . Unfortunately, we find that the structured non-convex sparsity is rarely considered to measure the reconstruction error in NMF. This observation raises the question: can we leverage the strengths of the graph regularization, label propagation, and these non-convex functions to enhance the robustness and effectiveness of image clustering? Figure 1: ACC performance of our proposed RONMF (including RONMF-M, RONMF-S, and RONMF-E) and competitors on MNIST. Motivated by the above studies, in this paper, we propose a robust orthogonal NMF method, termed RONMF. Note that the orthogonal constraints are imposed on the basis matrix to improve the discriminability of features and further improve the robustness [ 37 , 38 ] . The primary contributions of this paper can be described in the following three aspects. [1)] 1. We propose a unified NMF framework that utilizes non-convex functions as the error loss for better feature selection. To the best of our knowledge, this non-convex NMF method has not been investigated in the literature. 2. We enforce orthogonal constraints on the basic matrix to enhance sparsity, thereby reducing the correlation among basis vectors and improving the generalization performance. Besides, we consider the label information of labeled and unlabeled data with different weights, where graph Laplacian regularization explores the internal geometric structure of high-dimensional data and label propagation obtains the predicted membership values of each category. 3. We propose an optimization algorithm based on the alternating direction method of multipliers (ADMM), whose subproblems admit explicit solutions. Extensive experiments demonstrate the superiority and robustness of the proposed method in various indicators. As shown in Figure 1 , RONME with different non-convex functions (marked by red stars) can achieve high accuracy in an acceptable runtime. The structure of this paper is as follows. Section 2 briefly reviews NMF basics and variants. Section 3 presents the proposed method and optimization algorithm. Section 4 provides the experimental results and discussions. Finally, Section 5 summarizes this paper. 2 Related Work This section first introduces NMF basics, followed by some popular NMF variants. 2.1 NMF Basics NMF is designed

Read the complete text