Histogram Clustering Techniques

• Histogram clustering is a family of unsupervised techniques that use statistical histograms to represent data and guide clustering via dissimilarity measures and thresholding.
• It leverages diverse representations—from GLCM-based descriptors to probability maps and Wasserstein barycenters—to extract meaningful clusters in various applications.
• Applications include periodic texture inspection, conformal prediction calibration, and density level set estimation, with performance validated by metrics like DSC and coverage differences.

Histogram clustering denotes a family of unsupervised procedures in which the histogram is the primary analytical object. In the cited literature, the histogram may be a compact representation of a periodic image block, a clipped probability map from a segmentation network, a smoothed gray-level frequency curve, a histogram-valued descriptor of an individual, or a histogram density estimator whose thresholded level sets define connected components (Asha et al., 2012, Jaubert et al., 4 Jun 2025, Jassim, 2013, Irpino et al., 2011, Steinwart, 2014). A comparative tutorial on histogram-based image segmentation further emphasizes that histograms have been widely adopted to calculate the clustering means and partitioning thresholds for image segmentation, and organizes classical methods into means-based, Gaussian-mixture-model-based, entropy-based, and feature-points-based categories (Wang, 23 Feb 2025).

1. Problem formulations and scope

The literature does not use histogram clustering in a single sense. In one line of work, histograms are the objects being clustered: examples include flattened gray-level co-occurrence matrices (GLCMs) of periodic image blocks, clipped probability-map histograms used for conditional conformal calibration, and histogram-valued variables compared through Wasserstein geometry (Asha et al., 2012, Jaubert et al., 4 Jun 2025, Irpino et al., 2011). In another line, the histogram itself is partitioned or thresholded: Jassim’s THFCM method applies fuzzy c-means to the frequencies of a smoothed gray-level histogram and identifies a “discerner cluster” for automatic thresholding (Jassim, 2013). A third line uses histogram estimators of density and defines clusters as connected components of empirical level sets (Steinwart, 2014). A related extension constructs possibly-gapped histograms along a hierarchical clustering tree and then uses those bins for ANOHT, a nonparametric alternative to ANOVA-like analysis (Hsieh et al., 2017).

Formulation	Histogram object	Clustering output
Periodic texture analysis	Flattened 64×64 GLCMs	Defective vs. defect-free blocks
Conditional conformal calibration	Clipped probability-map histograms	Equal-size clusters for per-cluster calibration
Histogram-domain thresholding	Smoothed gray-level frequencies	Discerner cluster and binary segmentation
Histogram-valued data analysis	Empirical distributions / quantile functions	$K$ classes with prototypes
Density level-set estimation	Piecewise-constant histogram density	Connected components of level sets

This range of formulations indicates that histogram clustering is better understood as a methodological family defined by representation and objective, rather than as a single algorithmic template.

2. Histogram representations and dissimilarity structures

A central design choice is the representation of the histogram and the metric used to compare two such objects. In the GLCM-based defect-detection method, each periodic block is converted into a 64×64 count matrix after linearly re-quantizing gray levels from $0\ldots255$ to $0\ldots63$ by

$q(x)=\lfloor x\cdot 63/255 \rfloor .$

For each block, GLCMs are computed for $8$ directions $\theta\in\{0,\pi/4,\pi/2,\ldots,7\pi/4\}$ at distance $1$ pixel, summed into a single rotation-invariant GLCM, flattened into a vector of length $K=64\times 64$, and compared by the symmetric chi-square distance

$$\chi^2(H_1,H_2)=\frac12\sum_{k=1}^{K}\frac{(H_1(k)-H_2(k))^2}{H_1(k)+H_2(k)},$$

with bins for which $H_1(k)+H_2(k)=0$ skipped in the sum (Asha et al., 2012).

In Jaubert et al., the histogram is not a texture descriptor but a summary of predictive uncertainty. Let $0\ldots255$0 be the voxel-wise probability output of the segmentation model for image $0\ldots255$1. To remove the overwhelming mass of background voxels, only voxels with $0\ldots255$2 are retained, and the clipped probability map is encoded as a count vector $0\ldots255$3, where $0\ldots255$4–$0\ldots255$5 bins according to $0\ldots255$6 or $0\ldots255$7. No further normalization is applied. Similarity is measured by the Euclidean distance

$0\ldots255$8

The choice of raw counts rather than normalized frequencies makes the histogram encode both uncertainty shape and the amount of suprathreshold probability mass (Jaubert et al., 4 Jun 2025).

For histogram-valued data in the sense of symbolic or distributional data analysis, the histogram is itself a descriptor of an individual. If $0\ldots255$9 is the quantile function of histogram $0\ldots63$0, the $0\ldots63$1-Wasserstein distance is

$0\ldots63$2

Its squared form decomposes into a location term and a dispersion/shape term: $0\ldots63$3 The same paper further writes

$0\ldots63$4

which makes explicit the contributions of means, standard deviations, and quantile correlation (Irpino et al., 2011).

A different representation appears in histogram-based density clustering. Given a partition $0\ldots63$5 of $0\ldots63$6, the empirical histogram density estimator is

$0\ldots63$7

where

$0\ldots63$8

Here the histogram is not compared to another histogram by a pairwise metric; rather, it is thresholded at level $0\ldots63$9 to form the empirical level set

$q(x)=\lfloor x\cdot 63/255 \rfloor .$0

and clusters are the connected components of that set (Steinwart, 2014).

3. Algorithmic families

The segmentation tutorial identifies four major categories of histogram-based image segmentation methods: means-based, Gaussian-mixture-model-based, entropy-based, and feature-points-based methods (Wang, 23 Feb 2025). Means-based methods include K-means, Otsu’s method, and fuzzy c-means (FCM), all of which iteratively seek class-means that best explain the $q(x)=\lfloor x\cdot 63/255 \rfloor .$1D gray-level histogram. For histogram bins indexed by $q(x)=\lfloor x\cdot 63/255 \rfloor .$2 with probability $q(x)=\lfloor x\cdot 63/255 \rfloor .$3, the K-means objective is

$q(x)=\lfloor x\cdot 63/255 \rfloor .$4

whereas FCM introduces memberships $q(x)=\lfloor x\cdot 63/255 \rfloor .$5 and exponent $q(x)=\lfloor x\cdot 63/255 \rfloor .$6: $q(x)=\lfloor x\cdot 63/255 \rfloor .$7 These are histogram-domain clustering objectives even when the final output is an intensity threshold rather than a partition of histogram-valued objects (Wang, 23 Feb 2025).

THFCM is a more specific histogram-domain construction. It computes the gray-level histogram $q(x)=\lfloor x\cdot 63/255 \rfloor .$8, smooths it to obtain $q(x)=\lfloor x\cdot 63/255 \rfloor .$9, treats each bin $8$0 as a data point with coordinate $8$1, and applies standard FCM to the one-dimensional data set $8$2 with $8$3. The “discerner cluster” $8$4 is the cluster whose center $8$5 has the largest value, and all gray levels $8$6 whose frequency $8$7 has its highest membership in that cluster are labeled “object,” while the other clusters represent “background” (Jassim, 2013).

In periodic texture inspection, the workflow is pairwise dissimilarity construction followed by agglomerative clustering. After blockwise GLCM computation and histogram flattening, one builds an $8$8 symmetric matrix $8$9 with

$\theta\in\{0,\pi/4,\pi/2,\ldots,7\pi/4\}$0

noting that $\theta\in\{0,\pi/4,\pi/2,\ldots,7\pi/4\}$1 and $\theta\in\{0,\pi/4,\pi/2,\ldots,7\pi/4\}$2. The paper uses agglomerative hierarchical clustering with Ward’s linkage criterion and then cuts the dendrogram at the top level into two clusters. Since it is known a priori that defective blocks are fewer than defect-free blocks, the smaller cluster is labeled “defective” and the larger cluster is labeled “defect-free.” No explicit distance threshold is set (Asha et al., 2012).

Jaubert et al. use constrained K-means clustering of histograms inside a conformal prediction pipeline. The calibration histograms $\theta\in\{0,\pi/4,\pi/2,\ldots,7\pi/4\}$3 are partitioned into $\theta\in\{0,\pi/4,\pi/2,\ldots,7\pi/4\}$4 clusters $\theta\in\{0,\pi/4,\pi/2,\ldots,7\pi/4\}$5 with equal cardinality $\theta\in\{0,\pi/4,\pi/2,\ldots,7\pi/4\}$6 (rounded), minimizing the within-cluster sum of squared $\theta\in\{0,\pi/4,\pi/2,\ldots,7\pi/4\}$7 distances: $\theta\in\{0,\pi/4,\pi/2,\ldots,7\pi/4\}$8 The centroid is the average histogram,

$\theta\in\{0,\pi/4,\pi/2,\ldots,7\pi/4\}$9

and at test time the histogram $1$0 is assigned by

$1$1

This cluster assignment determines which conformal correction $1$2 is used for the predictive interval $1$3 (Jaubert et al., 4 Jun 2025).

For histogram-valued data, Irpino, Verde, and de Carvalho propose a dynamic clustering algorithm based on adaptive squared Wasserstein distances. The clustering criterion is

$1$4

and optimization alternates among a representation step, a weight-adaptation step, and an assignment step. Prototypes are Wasserstein barycenters defined by averaged quantile functions,

$1$5

and adaptive weights can be global or cluster-dependent, with separate components for means and dispersions (Irpino et al., 2011).

4. Hierarchical and level-set constructions

A distinct branch of histogram clustering is driven by hierarchical clustering and level-set topology rather than by centroid updates. In the possibly-gapped histogram framework, the ensemble of all candidate histograms is encoded by a two-layer Ising model. Each gap between consecutive sorted observations carries a layer-1 spin indicating whether adjacent points lie in the same bin or different bins, and each boundary position may also carry a layer-2 spin indicating whether there is an empty gap. The total number of configurations is

$1$6

This exponential growth makes exhaustive search infeasible, so the paper proposes a nearly-optimal algorithm that follows the branching hierarchy of a hierarchical clustering tree and decides whether to accept a split using a decoding-error sum-of-squares criterion (DESS) based on Uniform-within-bin coding (Hsieh et al., 2017).

For a bin $1$7 containing $1$8 points, the DESS is computed from the discrepancy between observed order statistics and ideal Uniform order statistics; a bin is declared “uniform enough” if

$1$9

up to a tolerance $K=64\times 64$0. Splitting a bin into $K=64\times 64$1 sub-bins is accepted only if the sum of the sub-bin decoding errors plus the boundary cost $K=64\times 64$2 is strictly smaller than the parent criterion. The resulting algorithm runs hierarchical clustering on the $K=64\times 64$3D data, traverses the tree top-down, tests local uniformity, and finally checks for existential gaps between adjacent leaf-bins. The summary states complexity $K=64\times 64$4 for the HC-driven construction (Hsieh et al., 2017).

Steinwart’s histogram-based density clustering addresses a different question: the smallest density level $K=64\times 64$5 at which the population level set splits from one component into two. The algorithm increments a level parameter $K=64\times 64$6, forms the empirical level set $K=64\times 64$7, finds all $K=64\times 64$8-connected components that persist at level $K=64\times 64$9, and stops when the number of surviving components is no longer one. It then returns $$\chi^2(H_1,H_2)=\frac12\sum_{k=1}^{K}\frac{(H_1(k)-H_2(k))^2}{H_1(k)+H_2(k)},$$0 and the two surviving components (Steinwart, 2014).

The theoretical contribution is a finite-sample analysis for histogram-based level set estimates. Under normality, clusterability, thickness of level sets, a separation exponent $$\chi^2(H_1,H_2)=\frac12\sum_{k=1}^{K}\frac{(H_1(k)-H_2(k))^2}{H_1(k)+H_2(k)},$$1, and a flatness exponent $$\chi^2(H_1,H_2)=\frac12\sum_{k=1}^{K}\frac{(H_1(k)-H_2(k))^2}{H_1(k)+H_2(k)},$$2, Steinwart proves high-probability inclusions of empirical and population level sets and derives both a level-estimation bound and a component-recovery bound. The paper also proposes a fully adaptive data-dependent strategy for selecting the width parameter $$\chi^2(H_1,H_2)=\frac12\sum_{k=1}^{K}\frac{(H_1(k)-H_2(k))^2}{H_1(k)+H_2(k)},$$3 by grid search, with rates that, up to logarithms, match the non-adaptive rates (Steinwart, 2014). In this formulation, histogram clustering is a topological estimation problem rather than a dissimilarity-minimization problem.

5. Empirical domains and reported performance

In patterned-texture inspection, the GLCM-based chi-square method was demonstrated on defective real-fabric images of $$\chi^2(H_1,H_2)=\frac12\sum_{k=1}^{K}\frac{(H_1(k)-H_2(k))^2}{H_1(k)+H_2(k)},$$4 major wallpaper groups, $$\chi^2(H_1,H_2)=\frac12\sum_{k=1}^{K}\frac{(H_1(k)-H_2(k))^2}{H_1(k)+H_2(k)},$$5 and $$\chi^2(H_1,H_2)=\frac12\sum_{k=1}^{K}\frac{(H_1(k)-H_2(k))^2}{H_1(k)+H_2(k)},$$6. The operational pipeline is highly structured: periodic blocks are extracted from four crops, blockwise GLCMs are accumulated across eight directions, pairwise chi-square distances are computed, Ward’s agglomerative clustering produces two groups, and the four-crop results are fused by merging the bounding-boxes of defective blocks, morphological closing/filling, and Canny contour extraction (Asha et al., 2012).

In coronary calcium quantification, histogram clustering appears as the calibration layer of a cluster-based conditional conformal prediction framework. The authors sweep $$\chi^2(H_1,H_2)=\frac12\sum_{k=1}^{K}\frac{(H_1(k)-H_2(k))^2}{H_1(k)+H_2(k)},$$7 and report that, at target coverage $$\chi^2(H_1,H_2)=\frac12\sum_{k=1}^{K}\frac{(H_1(k)-H_2(k))^2}{H_1(k)+H_2(k)},$$8, the highest difference $$\chi^2(H_1,H_2)=\frac12\sum_{k=1}^{K}\frac{(H_1(k)-H_2(k))^2}{H_1(k)+H_2(k)},$$9 occurs at $H_1(k)+H_2(k)=0$0. For $H_1(k)+H_2(k)=0$1, Table 2 gives measured coverage $H_1(k)+H_2(k)=0$2, $H_1(k)+H_2(k)=0$3, $H_1(k)+H_2(k)=0$4, and $H_1(k)+H_2(k)=0$5. In patch-based calibration over $H_1(k)+H_2(k)=0$6 splits of $H_1(k)+H_2(k)=0$7 augmented patches, conventional conformal prediction achieves coverage $H_1(k)+H_2(k)=0$8 with $H_1(k)+H_2(k)=0$9, whereas the $0\ldots255$00-cluster method has coverage $0\ldots255$01 and $0\ldots255$02. On the full-volume test set $0\ldots255$03, ensemble $0\ldots255$04 clustered CP $0\ldots255$05 attains coverage $0\ldots255$06, $0\ldots255$07, $0\ldots255$08, and $0\ldots255$09, compared to conventional CP’s $0\ldots255$10 (Jaubert et al., 4 Jun 2025).

In histogram-based image segmentation, the comparative tutorial reports that on synthesized three-class images with uniform $0\ldots255$11 Gaussian noise, FCM achieved highest $0\ldots255$12, while SDD had the lowest AMD. On cardiac MRI, SDD and max-entropy gave best $0\ldots255$13; on hand-gesture images with uniform background, K-means and Otsu reached $0\ldots255$14; on CT liver from CHAOS, only FCM and SDD gave meaningful contours with $0\ldots255$15. The same tutorial states that histogram methods are extremely cheap, require no training, and can outperform generalist deep nets without retraining on simple or structured scenes (Wang, 23 Feb 2025).

THFCM itself was evaluated on three standard $0\ldots255$16 gray-level images—F-16 aircraft, Lena, and Baboon. The paper shows the smoothed histogram curve with the three FCM-clusters overlaid, the discerner cluster marked on the highest peak, and the resulting binary segmentation. It does not report explicit numerical values for accuracy, PSNR, or execution time, but remarks that the visual results are “quite acceptable” and that the method is fully automatic and lightweight because FCM runs on $0\ldots255$17 frequency values rather than on $0\ldots255$18 pixels (Jassim, 2013).

For histogram-valued data, the dynamic clustering framework was studied on synthetic data and on $0\ldots255$19 Chinese weather stations with $0\ldots255$20 histogram variables. In the real dataset, CDC-AWD selected $0\ldots255$21 by the Calinski–Harabasz index and obtained $0\ldots255$22, compared with $0\ldots255$23 for STD and $0\ldots255$24 for GC-AWD. The learned weights revealed which variables and which components—mean versus dispersion—drove each cluster, and the cluster prototypes matched known climate zones (Irpino et al., 2011).

The possibly-gapped histogram and ANOHT framework was illustrated on Iris data, MLB pitch data, and divorce data. For Iris, the resulting species-tree showed $0\ldots255$25–$0\ldots255$26 authenticity on the known clustering $0\ldots255$27 versus $0\ldots255$28. For MLB pitch data and divorce data, the paper reports mechanistically sensible or substantively interpretable branching structures at high authenticity levels (Hsieh et al., 2017).

6. Assumptions, limitations, and recurrent misconceptions

A recurrent misconception is that histogram clustering refers to a single, uniform technique. The cited work shows instead that the phrase covers centroid-based clustering, fuzzy partitioning, agglomerative clustering on histogram distances, histogram-domain thresholding, Wasserstein barycentric clustering, and density level-set recovery (Asha et al., 2012, Jassim, 2013, Irpino et al., 2011, Steinwart, 2014). This suggests that methodological differences in representation, metric, and output type are as important as the shared reliance on histograms.

Many methods also depend on strong structural assumptions. In the defect-detection pipeline, the block size must exactly match one full pattern repeat; otherwise, violation of stationarity degrades clustering. The method also requires exactly two groups and relies on the assumption that defective blocks are fewer than defect-free blocks, so if the defect population were very large, the “smaller-cluster = defect” rule would break (Asha et al., 2012). In the conformal-calibration framework, K-means assignments are constrained by equal-size clusters, and performance depends on the choice of $0\ldots255$29, although the authors report best triage results at $0\ldots255$30 in their ablation (Jaubert et al., 4 Jun 2025).

Classical histogram-based segmentation methods have their own sensitivities. The tutorial states that K-means and Otsu are very fast, but K-means is sensitive to initialization and all three means-based methods are sensitive to histogram shape. GMM methods may converge to local optima and are sensitive to the initial guess. Entropy methods can be time-consuming for more than two thresholds because the search cost grows combinatorially as $0\ldots255$31. Feature-point methods such as HPV and SDD are very fast, but are sensitive to the smoothing parameter and to $0\ldots255$32 (Wang, 23 Feb 2025).

The hierarchical and level-set approaches replace some of these sensitivities with different ones. The possibly-gapped histogram formalism begins from an exponentially large candidate ensemble of size $0\ldots255$33, which is precisely why the practical algorithm is only nearly optimal and relies on a particular HC-driven scaffold rather than exhaustive search (Hsieh et al., 2017). Steinwart’s finite-sample guarantees require normality, clusterability, thickness of level sets, a separation exponent, and a flatness exponent. The guarantees are rigorous, but they are attached to a specific level-set notion of cluster, not to arbitrary similarity-based partitions (Steinwart, 2014).

The comparative evidence also argues against a second misconception: that histogram methods are either obsolete or universally superior. The tutorial states that no single hist-based method dominates all scenarios. Histogram methods are extremely cheap and require no training; they excel on uniform or simple backgrounds. Deep learning methods such as Cellpose, MIL, and SAM can segment highly complex scenes, but need extensive annotated data and retraining to match simple cases. The empirical record therefore supports a conditional view: histogram clustering is most effective when the histogram is a sufficient low-dimensional summary of the signal structure, and less compelling when the task depends on spatial semantics that the histogram discards (Wang, 23 Feb 2025).