Texture classification via morphological scale-space: Tex-Mex features Paul Southam Richard Harvey University of East Anglia School of Computing Sciences Norwich, NR4 7TJ, United Kingdom Abstract. We consider the problem of classifying textures. First, we consider images where the orientation of the texture is known. Then, we consider the classification of textures where the orientation is unknown. Last, classification in real scenes is considered. A wide variety of techniques are tested using the Outex framework. We introduce a new grayscale multiscale texture classification method based on a class of morphological filters called sieves. The method, denoted Tex-Mex because it extracts TEXture features using Mor- phological EXtrema filters, is shown to be among the best perform- ing texture feature extraction methods. Tex-Mex features can be computed rapidly and are shown to be more robust and compact than the alternatives. Furthermore, they may be applied over win- dows of arbitrary size and orientation, a useful attribute when clas- sifying texture in real scenes. © 2009 SPIE and IS&T. DOI: 10.1117/1.3258441 1 Introduction and Related Work Texture analysis is a problem that has attracted consider- able interest. As a result, texture analysis has developed into two distinct subproblems. First, there are methods that seek to model texture at a physics level—the study of sur- face roughness and so on. Second, there are analysis tech- niques that operate on “image texture,” 1 in which texture is treated as a gray-level or color variation without knowledge of the physical properties. Image texture methods are usu- ally evaluated on the standard tasks of texture detection and classification. Classification is the most studied, since its solution would appear to be a prerequisite for the solution of further problems including texture segmentation which is closely related to classification. Thus, this paper studies image texture classification. The first stage of image texture classification is the cap- ture of suitable images. In this paper, we work with stan- dard image databases since we are interested in comparing the performance of several systems; however, as you will see later, the choice of database is an important issue that can affect the measured performance of algorithms. The next stage of image texture classification is the extraction of features that are invariant to common nuisances such as unknown photometric calibration, yet vary with texture class. The final stage is classification. This paper focuses on the middle stage: feature extraction. There are many choices for texture features, so it is difficult to be confident that our review is comprehensive, but we hope to have selected features that have either been demonstrated to have good classification performance or are representative of a general class of features. Reviewing these features leads us to introduce a set of desiderata for texture features. Section 2 introduces a new set of features, the Tex-Mex features, that satisfy these desiderata and in Secs. 3 and 4 these are tested along with a selection of conventional features. Sec- tion 5 concludes that the new features have improved per- formance, gives reasons for their superiority, and discusses how they might be extended. Reviews of various texture features can be found in Refs. 2–7. Reference 8 summarizes texture analysis meth- ods into three classes, statistical, signal processing, and morphological methods. Statistical methods analyze the distribution of gray-level intensities in the textured image and were some of the ear- liest proposed texture models. 9–11 This class includes meth- ods such as local binary patterns LBP, 12 gray-level cooc- currence matrices GLCM, 13 and those derived from the autocorrelation function and Markov random fields MRFs. The signal processing class contains methods where tex- ture features are extracted using multiple filters. The sim- plest methods involve convolving a textured image with edge detectors, such as Sobel, Canny, or Laplacian of Gaussian. Other methods in this category include wavelets such as the dual tree complex wavelet transform DTCWT 14 and other filter-banks such as MR8 15 and LM. 16 The morphological category includes the use of max and min operators, which have had a long history in the analy- sis of image texture. In Ref. 17, for example, a method based on run-lengths is introduced that was later extended 18 to a system based on openings and closings using a struc- turing element. Later, granulometries for texture classifica- tion were introduced. 19 Granulometries are the application of a varying scale, but fixed shape, structuring element to produce an opening or a closing. Granulometric openings contain information about image structures brighter than their neighborhood, whereas granulometric closings capture image structures darker than their neighborhood. Usually, the mean polarity of the texture is unknown i.e., mostly white structures on a black background or vice versa, so it Paper 09036R received Mar. 30, 2009; revised manuscript received Aug. 28, 2009; accepted for publication Sep. 28, 2009; published online Nov. 20, 2009. 1017-9909/2009/184/043007/16/$25.00 © 2009 SPIE and IS&T. Journal of Electronic Imaging 18(4), 043007 (Oct–Dec 2009) Journal of Electronic Imaging Oct–Dec 2009/Vol. 18(4) 043007-1