Is this the same as a gray-level co-occurrence matrix?

It uses a related counting idea, but the paired values come from different scale representations under the stated construction.

Must the matrix sum to one?

A normalized version does; the raw count matrix sums to the number of included pairs.

Are all wavelet co-occurrence methods identical?

No. Methods may pair different bands, scales or positions, so their definitions and feature formulas must be checked separately.

AI Glossary / Mathematics and optimization / Advanced

Scale co-occurrence matrix

Definition

A scale co-occurrence matrix is an image-texture descriptor that counts how values at corresponding locations occur together across two image scales. In the method introduced by Wu Jun and Zhao Zhongming, a wavelet-frame decomposition produces aligned low-frequency images at successive scales. Their paired values are counted in a matrix, from which statistical texture features can be calculated.

What it means

Texture is not described only by the brightness distribution at one resolution. A surface can look similar to another at a fine scale yet change differently when progressively smoothed. A scale co-occurrence matrix records part of that cross-scale relationship. This differs from a conventional gray-level co-occurrence matrix that pairs values at spatially separated positions within one image. The particular scales, transforms, value bins and correspondence rule define the descriptor, so the name alone is insufficient to reproduce a calculation.

How it works

In the original Wu - Zhao construction, use a discrete wavelet frame without downsampling so the low-frequency images retain matching dimensions. For a selected adjacent-scale pair, map values to declared discrete levels and inspect corresponding locations. Increment matrix cell (i, j) whenever the first selected scale has level i and the second has level j. Divide by the total number of valid pairs to obtain a joint-frequency distribution if normalized features are required. Repeat for further scale pairs, then calculate chosen statistics such as energy, entropy or contrast. Fix row and column conventions explicitly. A decimated transform produces different-sized arrays and would require an additional correspondence rule rather than direct position-by-position pairing.

Three practical examples

A remote-sensing experiment can evaluate cross-scale texture features for distinguishing image regions, while testing whether sensor resolution and land-cover variation affect transferability.
A materials-inspection prototype can compare texture patches from surfaces photographed under controlled conditions, using the matrix as one candidate feature representation.
An image-retrieval system can investigate whether cross-scale relationships help retrieve visually related textures. Similarity in descriptor space still needs evaluation against the intended meaning of a useful match.

Think of it this way

Think of comparing the same patterned cloth through two levels of blur and recording which fine-view shades accompany which blurred-view shades at matching places. The record describes how the pattern changes with viewing scale. The analogy captures correspondence across representations, but the actual method uses specified filters and numerical bins. It does not simply compare two photographs taken at arbitrary distances or guarantee invariance to zoom.

Worked example

Consider six illustrative paired locations after a fixed transform and quantization step. Let first-scale levels be [0, 0, 1, 1, 1, 0] and second-scale levels [0, 1, 1, 1, 0, 0]. Use first-scale levels as rows and second-scale levels as columns, both ordered 0 then 1. The count matrix is [[2, 1], [1, 2]], and its entries sum to six. Normalization gives [[1/3, 1/6], [1/6, 1/3]]. Energy, defined as the sum of squared probabilities, is 2(1/3)² + 2(1/6)² = 5/18, approximately 0.278. Contrast, defined here as Σ(i − j)²P(i, j), is 1/6 + 1/6 = 1/3. Both marginal distributions are [1/2, 1/2]. Now replace the paired data by three (0, 0) pairs and three (1, 1) pairs. The marginals remain identical, but the normalized matrix becomes [[1/2, 0], [0, 1/2]], with energy 1/2 and contrast zero. Thus separate histograms miss a relationship that the joint matrix captures. These arrays illustrate counting after preprocessing; they are not claimed to be a complete wavelet decomposition of an invented source image.

Illustrative case study

In a fictional project, Alder Surfaces compares candidate descriptors for fabric photographs. The team locks the camera setup, patch dimensions, filters and quantization boundaries before evaluation. It separates photographs from different physical samples across development and test sets, rather than letting nearby crops of one sample appear in both. When performance changes under new lighting, the team investigates that shift instead of assuming the multiscale descriptor is automatically illumination invariant. No accuracy figure or commercial outcome is asserted.

Common misconceptions and mistakes

Cross-scale does not mean scale invariant. Filtering, image resolution, crop size and quantization can all affect the matrix. It is a joint count, not a covariance matrix, and it need not be symmetric: swapping scale order generally transposes it. The six-pair example happens to be symmetric, which is not a universal property. More bins can retain detail but create sparse, unstable estimates. A high feature value is not automatically better; each statistic summarizes a different aspect of the joint distribution. The matrix also discards much spatial arrangement, so it cannot completely describe every texture.

Business use

Use this descriptor when a concrete image-analysis task may benefit from relationships across scales and interpretable feature calculations are valuable. Record preprocessing, filter family, boundary handling, scale count, quantization, valid-pixel mask and normalization. Compare with simpler single-scale and independent-band features using the same evaluation split. If a classifier consumes the resulting features, report classifier performance separately from the descriptor's mathematical definition. Investigate sensitivity to acquisition changes before using a threshold operationally, and retain enough provenance to recreate the features from the original image.

Check your understanding

Two textures have identical histograms at each of two scales. Must their scale co-occurrence matrices match?

Answer: No. The worked example has identical marginals but different pairings, matrices, energy and contrast.

Related terms

Feature extraction
wavelet frame
multiresolution analysis
gray-level co-occurrence matrix
joint histogram
texture classification

Learn next

reproduce both binary matrices from the example and calculate their row and column sums. Then construct a nonsymmetric pair table and reverse the scale order. Finally, test how changing quantization boundaries alters a descriptor while keeping the underlying images fixed.

Sources

Sources: Wu Jun and Zhao Zhongming, Scale Co-occurrence Matrix for Texture Analysis Using Wavelet Transform, Journal of Remote Sensing 5(2), 100–103, https://doi.org/10.11834/jrs.20010205 ; publisher article record and linked four-page paper, https://www.ygxb.ac.cn/en/article/10638453 . The original paper's wavelet-frame construction and counting equations were reviewed. The binary matrices and fictional inspection study are original.

Previous: Scale analysis (statistics) Next: Scale-invariant prior