Feature Extraction & Image Processing, 2nd Edition

Of the other approaches, Korn (1988) developed a unifying operator for symbolic repre-

sentation of grey level change. The Susan operator (Smith and Brady, 1997) derives from

an approach aimed to find more that just edges, since it can also be used to derive corners

(where feature boundaries change direction sharply, as in curvature detection in Section 4.8) and

structure-preserving image noise reduction. Essentially, SUSAN derives from smallest univalue

segment assimilating nucleus, which concerns aggregating the difference between elements in

a (circular) template centred on the nucleus. The USAN is essentially the number of pixels

within the circular mask that have similar brightness to the nucleus. The edge strength is then

derived by subtracting the USAN size from a geometric threshold, which is, say, three-quarters

of the maximum USAN size. The method includes a way of calculating edge direction, which

is essential if non-maximum suppression is to be applied. The advantages are in simplicity (and

hence speed), since it is based on simple operations, and the possibility of extension to find

other feature types.

4.5 Comparison of edge detection operators

The selection of an edge operator for a particular application depends on the application itself.

As has been suggested, it is not usual to require the sophistication of the advanced operators

in many applications. This is reflected in analysis of the performance of the edge operators on

the eye image. To provide a different basis for comparison, we shall consider the difficulty

of low-level feature extraction in ultrasound images. As has been seen earlier (Section 3.5.4),

ultrasound images are very noisy and require filtering before analysis. Figure 4.31(a) is part

of the ultrasound image which could have been filtered using the truncated median operator

(Section 3.5.3). The image contains a feature called the pitus (the ‘splodge’ in the middle) and

we shall see how different edge operators can be used to detect its perimeter, although without

noise filtering. Earlier, in Section 3.5.4, we considered a comparison of statistical operators on

an ultrasound image. The median is perhaps the most popular of these processes for general

(i.e. non-ultrasound) applications. Accordingly, it is of interest that one study (Bovik et al., 1987)

has suggested that the known advantages of median filtering (the removal of noise with the

preservation of edges, especially for salt and pepper noise) are shown to good effect if it is used

as a prefilter to first and second order approaches, although with the cost of the median filter.

However, we will not consider median filtering here: its choice depends more on suitability to

a particular application.

The results for all edge operators have been generated using hysteresis thresholding, where

the thresholds were selected manually for best performance. The basic first order operator

(Figure 4.31b) responds rather nicely to the noise and it is difficult to select a threshold that

reveals a major part of the pitus border. Some is present in the Prewitt and Sobel operators’ results

(Figure 4.31c and d, respectively), but there is still much noise in the processed image, although

there is less in the Sobel. The Laplacian operator (Figure 4.31e) gives very little information

indeed, as to be expected with such noisy imagery. However, the more advanced operators

can be used to good effect. The Marr–Hildreth approach improves matters (Figure 4.31f), but

suggests that it is difficult to choose a LoG operator of appropriate size to detect a feature

of these dimensions in such noisy imagery, illustrating the compromise between the size of

operator needed for noise filtering and the size needed for the target feature. However, the

Canny and Spacek operators can be used to good effect, as shown in Figure 4.31(g) and (h),

respectively. These reveal much of the required information, together with data away from the

Low-level feature extraction (including edge detection) 145