136 CONSTANTINE KOTROPOULOS

et al.

The image processing applications of nonlinear mean filters, defined by the s p

mean filter, were first proposed by Kundu, Mitra, and Vaidyanathan [Kun84]. Fur-

ther generalizations of the

L p

mean filter were later advanced by Pitas and Venet-

sanopoulos [Pit86a, Pit86b, Pit86c]. The nonlinear mean filters remove the impulse

noise very effectively, especially when the impulses occur with a high probability.

They have a very simple structure and are suitable for real-time processing ap-

plications. The class of nonlinear mean filters can be treated as a generalization

of linear filters (e.g., the arithmetic mean filter). They are closely related to ho-

momorphic filters, perhaps one of the oldest classes of nonlinear filters. From a

statistical point of view, nonlinear mean filters rely on the nonlinear means that are

well-known location estimators [Ken73]. The latter approach is found to be more

fruitful because it produces a general filter structure that encompasses the filters

that are based on order statistics, the homomorphic filters, and the morphological

filters in addition to the nonlinear mean filters.

The nonlinear mean filters are first defined and their statistical as well as edge

preservation properties reviewed in Sec. 5.2. Their performance in the presence

of impulse noise is outlined as well. In Sec. 5.3,

L p

mean filters are proven to

be a general filter structure able to suppress signal-dependent noise. A class of

edge detectors obtained from the difference of nonlinear mean filters is studied in

Sec. 5.4. Next, soft grayscale morphological filters based on Lp mean filters are de-

scribed in Sec. 5.5. L2 mean filters are shown to be optimal for both multiplicative

Rayleigh speckle and signal-dependent Gaussian speckle noise. These properties

supported the application of the s mean filter, the signal-adaptive L2 mean filter,

and the L2 learning vector quantizer in ultrasonic image processing and analysis,

which are reviewed in Sec. 5.6. The use of

Lp

mean filters as approximators of

max and min operators for positive and negative p, respectively, is exploited in

Sec. 5.7 to design analog implementations of sorting networks. Edge preserving

filtering by combining nonlinear means and order statistics is studied in Sec. 5.8.

5.2 Nonlinear Mean Filters

The nonlinear mean of the N numbers xi, i = 1, 2,..., N, is defined by

y=f(x~,x2,...,XN)=g-~(y'~Nla~g(x~))

2i =

i ai

(5.1)

where g (x) is, in general, a single-valued analytic nonlinear function and a i are

weights. The properties of the nonlinear mean depend on the function g (x) and

the weights ai. Equation (5.2) indicates several choices of g(x) that result in filters

playing a special role in image processing:

x

1/x

g(x) = lnx

x p,

p ~ R- {-1,0, 1}

arithmetic mean ~,

harmonic mean YH,

geometric mean YG,

Lp

mean yLp,

(5.2)

Get *Nonlinear Image Processing* now with O’Reilly online learning.

O’Reilly members experience live online training, plus books, videos, and digital content from 200+ publishers.