Supervised texture classification by integration of multiple texture methods and evaluation windows
Introduction
The problem of texture classification consists of identifying the texture patterns present in a digital image given a set of known texture patterns (models) of interest. In the most general form of the problem, in which the input image may contain several regions of different texture, the aim is to identify the texture pattern to which every image pixel belongs. This will be referred to as pixel-based texture classification.
Two examples of pixel-based texture classification would be the identification of tissue with a specific granularity in an X-ray image or the location of specific types of soil in a satellite or aerial image. These problems do not just require the segmentation of the given image into distinctive regions – this is the goal of unsupervised texture segmentation algorithms (e.g. [13], [17], [19], [27], [31], [34]) – but the identification of specific regions of interest within the image. Obviously, such an identification will also lead to the segmentation of those regions – in this respect, classification subsumes segmentation (e.g. [16]) – but it will not necessarily lead to the segmentation of the whole image. In this scope, supervised classification algorithms (e.g. [2], [23]) are generally considered to be more suitable than unsupervised segmentation algorithms.
In order to perform pixel-based texture classification, a number of texture measures are computed for every image pixel by applying a predefined set of texture feature extraction methods (texture methods in short) to its neighboring pixels [28]. Usually, these neighborhoods are square windows centered at the considered pixel.
A wide variety of texture feature extraction methods have been proposed in the computer vision literature (e.g. [21], [24], [28], [30], [33]). Their performance basically depends on the type of processing they apply, the size of the neighborhood of pixels over which they are evaluated (window size) and the texture content. When dealing with those methods, two different issues must be addressed: (1) the determination of proper window sizes for evaluating each method, and (2) how the various methods are combined in order to improve the result. These issues are further discussed below.
Although many studies regarding the performance of the different families of texture feature extraction methods have been carried out (e.g. [6], [22], [28]), few have dealt with the issue of determining optimal window sizes, and they do it for specific texture methods (e.g., [3], [8], [21]). Hence, window sizes are commonly defined on an experimental basis, with each method being evaluated over windows of a single size. The role played by both the shape and size of those windows was studied in [11], showing that texture characterization is much more influenced by the window size than by its shape, although no hints on optimal sizes are provided.
The main problem regarding the definition of proper window sizes for texture feature extraction methods is that two contradictory requirements appear when these methods are utilized for image segmentation and classification. Thus, while large windows are appropriate for texture discrimination in wide areas of uniform texture, they perform poorly in contact areas among different textures. In the latter case, small windows are more adequate in order to identify the correct texture near the boundaries. Since the performance of texture methods is greatly influenced by the particular textures to which they are applied, it is not feasible to devise general strategies for determining window sizes that allow optimal discrimination of arbitrary textures.
On top of the problem of selecting window sizes, no single texture method is good enough to completely characterize and, therefore, distinguish the different textures that may appear in nature. Thus, several methods must be combined in order to obtain good classification results. The different proposals typically combine methods that belong to a specific family (e.g. [2], [5], [22], [24]). This combination is commonly tailored to the particular methods that are chosen. Unfortunately, the results they obtain are likely to depend on the type of texture models to which the different proposals are targeted.
Notwithstanding, every texture feature extraction method is potentially useful for texture discrimination to a larger or lesser extent. Hence, instead of trying to find out the best set of texture feature extraction methods – this being a challenge that, in general, appears to be of doubtful feasibility in considering that there is no consensus in the literature about which family of methods or combination of them is the best— [10], [26] show that classification rates can be improved by integrating different families of texture methods. The reason is that no single method is excluded a priori. Instead, all of them are considered in order to improve the result.
This paper presents a new pixel-based texture classifier that combines different families of texture feature extraction methods, with each method being evaluated over multiple windows of different size. A Bayesian scheme based on the application of the Kullback-J divergence is proposed. Experimental results show that this technique yields significantly lower classification errors and better quality segmentations of the sought regions of interest than well-known classifiers and segmenters based on single families of texture methods evaluated over single-sized windows.
The organization of this paper is as follows. Section 2 describes the evaluation of texture methods over multisized windows. Section 3 presents the proposed pixel-based texture classifier. Section 4 shows experimental results of the integration of widely used texture feature extraction methods with the proposed technique, as well as a comparison with a public texture classification framework (MeasTex [32]), the recent LBP texture classifiers [24] and two publicly available unsupervised texture segmenters (JSEG [9] and Edge Flow [18]). A practical application of the proposed technique to fabric defect detection is also described. Finally, conclusions and further improvements are given in Section 5.
Section snippets
Evaluation of texture methods over multisized windows
Let I be a two-dimensional image of N = R × C pixels containing several regions of uniform texture. Let {τ1, … , τT} be a set of T texture models of interest. Each model τk is described by a sample image Ik that contains a pattern of that texture (the algorithm can be trivially adapted to deal with multiple sample images per texture pattern). The goal of a pixel-based texture classifier is to determine if a pixel I (x, y) belongs to any of the T aforementioned texture models. If the classifier is
Pixel-based texture classification
Given an image I and M texture feature extraction methods, {μ1, … , μM}, which generate corresponding feature vectors, {F1, … , FM}, when FM are evaluated in the neighborhood of a certain pixel I (x, y) by using a set of W windows of different size, {s1 × s1, … , sW × sW}, this section presents a technique for integrating the M feature vectors in order to determine whether pixel I (x, y) can be classified into one of T given texture models {τ1, … , τT}. The proposed technique has five stages.
The first stage
Experimental results
The proposed technique has been extensively evaluated on a set of composite Brodatz images [4] [e.g., Fig. 1(top)] and real outdoor images [e.g. Fig. 1(bottom)]. Fig. 2(top) shows eight Brodatz texture patterns utilized as models for the proposed supervised classifier. Each pattern belongs to one of the eight texture categories proposed by Rao and Lohse [29] as representatives of the variability of natural textures according to human perception. Fig. 2(bottom) shows five outdoor texture
Conclusions and further work
This paper shows that pixel-based texture classification can be significantly improved by integrating texture methods from multiple families, each evaluated over multisized windows. As an example, 14 methods from seven different families have been integrated, with each method evaluated over six window sizes. A practical application to fabric defect detection has also been presented.
The proposed technique consists of an initial training stage that evaluates the behavior of each considered
Acknowledgement
This work has been partially supported by the Spanish Ministry of Education and Science under project DP12004-07993-C03-03.
References (34)
- et al.
Unsupervised texture segmentation via wavelet transform
Pattern Recogn.
(1997) - et al.
A comparative study of texture measures with classification based on feature distributions
Pattern Recogn.
(1996) - et al.
A comparison of texture feature extraction using adaptive Gabor filtering, pyramidal and tree structured wavelet transforms
Pattern Recogn.
(1996) - et al.
A review of recent texture segmentation and feature extraction techniques
CVGIP: Image Understand.
(1993) - et al.
Measuring texture classification algorithms
Pattern Recogn. Lett.
(1997) Statistical Decision Theory and Bayesian Analysis
(1985)- et al.
Bayesian image classification using Markov random fields
Image Vision Comput.
(1996) - et al.
Shape from texture: integrating texture-element surface estimation
IEEE Trans. Pattern Anal. Mach. Intell.
(1989) Textures: A Photographic Album for Artists and Designers
(1999)- et al.
On the classification of image regions by colour, texture and shape
Pattern Recogn.
(1993)
Texture segmentation using fractal dimension
IEEE Trans. Pattern Anal. Mach. Intell.
Unsupervised segmentation of color-texture regions in images and video
IEEE Trans. PAMI
Textural features for image classification
IEEE Trans. Systems Man Cybernet.
Cited by (33)
Two novel local binary pattern descriptors for texture analysis
2015, Applied Soft Computing JournalScale selection for supervised image segmentation
2012, Image and Vision ComputingCitation Excerpt :Pixel classification is one of the common ways to perform supervised image segmentation. It can adapt well to different segmentation tasks with different training data and is able to provide state-of-the-art performance in various settings [53–55,7,42,41,43,5,56]. Its basic idea is to extract a feature vector F(x,y) for each pixel, and to train a classifier h for obtaining labels of pixels.
Statistical methods for segmentation and quantification of minerals in ore microscopy
2012, Minerals EngineeringA new method to quantify parameters of membrane morphology from electron microscopy micrographs by texture recognition
2011, Chemical Engineering ScienceMulti-level pixel-based texture classification through efficient prototype selection via normalized cut
2010, Pattern RecognitionCitation Excerpt :Pixel-based texture classification [3] is a combination of the two aforementioned application domains in the sense that it determines the class to which every pixel of an input image belongs, which obviously leads to the segmentation of the image as a collateral effect. Hence, pixel-based classification is a possible way to perform supervised segmentation [4,19–25]. In pixel-based texture classification, several measures are computed for each image pixel by applying a set of texture feature extraction methods to its neighboring pixels.