Home Articles New texural extraction method using rolling ball and riping membrane transforms

New texural extraction method using rolling ball and riping membrane transforms

ACRS 1996

Digital Image Processing

New Texural Extraction Method Using Rolling Ball and Riping Membrane Transforms

Mazlan Hashim
Faculty of Geoinformation Science & Engineering
Universiti Teknologi Malaysia
Locked Bag 791, 80990 Johor Bahru, Malaysia
Tel . + (07)-5502969, Fax: + (07)-5566163
e-mail: [email protected]

Abstract
The inherent spatial information within a satellite remotely sensed data has shown a significant contribution to the classification of an image. In this, paper, two new textural approaches, namely the Rolling Ball Transform (RBT) and Ripping membrane (RM), were examined and analysed for image classification. the classification results are presented in the following models: (I) textural information alone, (ii) combination of best textural information and spectral data, and (iii) spectral data alone. Comparison of classification results obtained from the two new textural methods with three commonly textural extraction techniques : grey level co-occurrence, local statistical transform and convolution filtering masks were also carried out. Result of this study shows that the new textural extraction methods as one of the possible methods to increase the classification accuracies of the land use classes. The best results were obtained when textural information was combined with raw data. The new textural approaches- the RBT and RM both showed the most stable textures. RM textures from range 20 was the best overall textural extraction method.

1.Introduction
The use of spectral data alone in the classification has limited success due to the high variability within the spectral data. This variability is attributed to the within -class variances caused by the fragmented nature of the class e.g. urban contains varieties of urban structures like roads of various sizes and surface types, building variety and also variable vegetation cover. Even in the current high spatial resolution satellite data such as the Landsat TM, these small fragments and also other fragments found in most land use categorical classes as previously shown in Hashim (1995) are usually at subpixel resolution. Hence, classification of such classes spectral data alone is difficult.

The above-mentioned spectral variations are often referred to as inherent spatial information. Spatial information of remotely sensed data is assumed to be characterizes by the image texture (Haralick et. Al., 1973) which is an amalgam of geometrical properties (shape, size position, site, distribution) and the sensor imaging characteristics. Incorporating textural information into the image classification process may therefore provide a means of characterizing the classes within the land use classification system.

In this paper the incorporation of texture for improving image classification for land use mapping will be examined. Five textural extraction methods are employed : (1) grey level co-occurrence matrix (GLCM) (Haralick et. Al., 1973); (ii) Local statistical transform (LST) using median filter; (iii) convolution filtering masks (CFM) using Thomas et. Al. (1987), and Ford et. Al. (1983); (iv) rolling ball transformation (Sternberg, 1983; Stenberg, 1986); and (v) ripping membrane transformation (Blake and Zisserman, 1988). Emphasis, however, is place to the last two new textural extraction methods. These two methods were originally proposed as algorithms for use in biomedical image processing and visual reconstruction, respectively. In this study, the adapting of these methods will be examined by varying the parameters that best describe the textures of interest.

2.The Rolling Ball Transformation (RBT)
The RBT is mathematical morphological transformation first proposed as an algorithm to minimise image background noises. The RBT was proposed in this paper as a new method to minimize inhomogeneous effects due to inherent spatial information.

In the RBT transformation process, the image is viewed as a set of boxes or umbra (cubical pixels) in 3-D space (see Figure 1). The pixel and row number forms the umbra basis and ordinate while the pixel intensity is the height of the boxes. The ball being the structure element is moved freely on the umbra. The center of the ball is determined by the finding manima and then minima within the umbra of minima-template. The trajectory of the ball’s center is the smoothed image. The transformation process is then repeated with the ball placed below the umbra surface for picking up the dimples; this process is known as opening. The closing is the former transformation process where the ball glides on the umbra. The opening and closing are the terms in mathematical morphology where the structural elements are used as operator to transform input pixel in an image according to the relationship of each pixel to other pixels in its neighbourhood. The structural element in the RBT is a sphere (ball). The opening of greyscale image X by structuring element Y is given by:

XY = [XQ Y]