Next: Measurement of autocorrelation Up: Two-dimensional Imaging of Previous: Scanning electron microscopy

Image segmentation and median filtering

The key to combining information contained in the X-ray and backscattered electron images is that they all have been obtained in viewing the same area of the specimen. Thus, at any pixel (location) in the image, one has six signals available, the backscattered electron intensity and the five acquired X-ray signals. These signals can be interpreted as a set to determine the underlying phase present at each pixel comprising the image. The general algorithm for this is illustrated in Fig. 2. Here, the ratio of calcium to silicon (Ca/Si) is employed to separate the C₃S from the C₂S. Alternatively, the intensity of the backscattered electron image could be employed for this purpose, as discussed in the previous section. In addition, the alkali sulfates are identified as gypsum using the algorithm in Fig. 2. However, if X-ray images are also collected for K and Na, the alkali sulfates can be readily separated from the gypsum by comparing the K and Na intensities to user specified threshold values. The threshold values, X^*, to use for each X-ray image can generally be determined by viewing the image's greylevel (intensity) histogram [7]. This histogram is simply a plot of the number of pixels in an image corresponding to each possible greylevel value (with greylevels typically ranging from 0 to 255). The intensities corresponding to pixels where the element of interest is present will generally comprise either a separate peak or a peak shoulder of this histogram.

Once this initial segmentation is performed, the resulting image still contains a considerable amount of noise which must be removed. To do this, a type of median filtering algorithm is employed. Here, a small square, typically 5 pixels by 5 pixels, is centered at each pixel in the processed image. For all non-void pixels, the number (proportion) of pixels of each phase in this square is tabulated and the current pixel value replaced by the majority phase present in the square, if this majority proportion exceeds 0.5, to create a new ``final'' processed image. Figure 3 shows such a final processed image for a type I ordinary portland cement footnote 2. Using simple point counting procedures (program statsimp.c given in Appendix A), the area percentage of a phase(s) and the percentage of a phase(s) comprising the particle surfaces (perimeter percentage) can be readily determined for these images.