321.

[Cover] [Contents] [Index]

Page 88

2.5.1 Use of texture and context

Texture and context both measure aspects of the spatial structure of an image. They are frequently used in image interpretation and are generally extracted directly from an image. Texture refers to a description of the spatial variability of tones found within part of a scene. In visual terms, texture is expressed as the impression of roughness or smoothness created by the variation of tone or repetition of patterns across a surface. The application of tone and texture in digital image analysis required a quantitative characterisation of these visual concepts.

Two texture descriptors for a defined area can be simply the mean and variance of the region. These are the simplest features to characterise textures, but our experience suggests that they are not sufficient to characterise texture in the best way possible. For example, it is possible for different texture patterns to have the same mean and variance. More comprehensive approaches are required. Different methods used to quantise texture patterns also have a profound effect on classification accuracy. Chapter 5 presents a review and comparative study of texture extraction techniques.

The context of a pixel or a group of pixels refers to its probability of occurrence based on the nature of the pixels in the remainder of the scene. Human photo-interpreters have long exploited spatial information like texture and context, and attempts have been made to incorporate these attributes of the scene to computer-assisted image classification. There is clear evidence that data from future sensors, with their finer spatial resolving power, may not necessarily generate improved classifications when per pixel classifiers are used (Townshend, 1981) because of the higher spectral variability of local areas of the image, which becomes apparent as the resolution becomes finer.

Contextual information can either be included in a statistical classifier or in procedures that amend some preliminary classifier output. A simple way to use context is in the form of a majority filter window (Gurney, 1980, 1981). For instance, let a window be centred on a pixel labelled i. If the majority of the pixels within the window belong to class j, the label of the central pixel is altered from i to j. In most cases, the majority filter is used for classification refinement. Although there is some improvement, the increase in classification accuracy is not impressive. More complex statistical models of spatial context are discussed by Swain et al. (1980), Richards et al. (1982), Haslett (1985), and Kim and Swain (1995).

Chapter 6 also provides details of methods of modelling context as prior probability using a Markov random field (MRF) model, and this prior probability is further combined with class-conditional probability to perform a MAP classification. The increase in classification accuracy compared with traditional methods demonstrates the practical value of this approach (Tso and Mather, 1999).

[Cover] [Contents] [Index]