Viewed simply, MeasTex is a framework for measuring the performance of a texture classification algorithm on a suite of texture classification problems.
On any single classification, "Is this homogeneous image more like this texture or that texture?", a classification algorithm will be either right or wrong. Sometimes, a better algorithm will give an incorrect answer to a single classification that a poorer algorithm gives a correct answer to. A meaningful, replicable measure of an algorithm's performance requires combining the results of classifying many images. The MeasTex framework specifies a method for combining the results of many classifications.
MeasTex is not just a theoretical framework. MeasTex specifies a format for a suite of texture classification problems, and specifies the format of the input and output of a texture classification algorithm. Given these formats, we have implemented software which applies an algorithm to suites of texture classification problems. The software also processes the output of the algorithms, giving a score of the algorithms performance on each texture problem and a summary score for each suite of texture problems.
Texture classification algorithms all analyse textures: but there is a great diversity of textures, and consequently of texture classification problems. For example, textures are generated by objects as diverse as forests, farmlands, cities and oceans viewed from space, as diverse as different weaves of cloth, as diverse as the pattern of text and photographs in newspaper, as diverse as any homogeneous region that can be viewed as an image. The MeasTex framework is modular: it allows a test suite to be constructed from any set of homogeneous textures.
The MeasTex site contains several test suites. The Brodatz and VisTex test suites are not specific to any single texture domain, but rather are based on the Brodatz and VisTex image sets. These image sets contain images from diverse sources, such as cloth weave, vegetation, rocks and brickwork. There are several test suites based on artificially contructed textures: these measure the performance of an algorithm on microtextures and macrotextures, and on a sequence of similar texture problems of varying difficulty. It is our intention to add further, more specific, test suites.
The primary goal of MeasTex is to measure the performance of texture classification algorithms. To facilitate this, the design of MeasTex is modular. The MeasTex framework (and associated software) will measure any algorithm which accepts input and generates output in the format assumed by the framework. The Meastex site contains implementations of some algorithms (Grey Level Coocurrence Matrix, Gabor Energies, Gauss Markov Random Fields, and Fractal Dimension).
We hope that MeasTex will provide a valuable tool to the texture analysis community, enabling the quantitative comparisons of algorithms. The implementation of any algorithm requires choosing values for parameters (such as the set of frequencies in a Gabor Energy algorithm) and making design decisions (such as which of the textural features should be used in a Grey Level Cooccurence algorithm). MeasTex allows a quantitative evaluation of each of these design decisions.
We hope MeasTex will become a widely used tool. Consequently the framework is publicly available, including the test suites, the supporting software and sample algorithm implementations.