See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/233871684

A Learning Framework for Morphological Operators Using Counter–Harmonic

Mean

Conference Paper · December 2012

DOI: 10.1007/978-3-642-38294-9_28 · Source: arXiv

CITATIONS READS
34 449

3 authors, including:

Jonathan Masci Jesús Angulo

Dalle Molle Institute for Artificial Intelligence MINES ParisTech
50 PUBLICATIONS 7,578 CITATIONS 183 PUBLICATIONS 3,301 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Virtual Reality Navigation System for Laparoscopic Surgery View project

All content following this page was uploaded by Jesús Angulo on 04 November 2014.

The user has requested enhancement of the downloaded file.

 A Learning Framework for Morphological
Operators using Counter–Harmonic Mean

Jonathan Masci1? , Jesús Angulo2 , and Jürgen Schmidhuber1

1
IDSIA – USI – SUPSI, Manno–Lugano, Switzerland
[email protected],
arXiv:1212.2546v1 [cs.CV] 11 Dec 2012

WWW home page: http://idsia/~masci/

2
CMM-Centre de Morphologie Mathématique,
Mathématiques et Systèmes, MINES ParisTech, France

Abstract. We present a novel framework for learning morphological op-

erators using counter-harmonic mean. It combines concepts from mor-
phology and convolutional neural networks. A thorough experimental
validation analyzes basic morphological operators dilation and erosion,
opening and closing, as well as the much more complex top-hat trans-
form, for which we report a real-world application from the steel industry.
Using online learning and stochastic gradient descent, our system learns
both the structuring element and the composition of operators. It scales
well to large datasets and online settings.

Keywords: mathematical morphology, convolutional networks, online

learning, machine learning

1 Introduction

Mathematical morphology (MM) is a nonlinear image processing methodology

based on computing max/min filters in local neighborhoods defined by struc-
turing elements [16,17]. By concatenation of two basic operators, i.e., the di-
lation δB (f ) and the erosion εB (f ), on the image f , we obtain the closing
ϕB (f ) = εB (δB (f )) and the opening ϕB (f ) = γB (δB (f )), which are filters
with scale-space properties and selective feature extraction skills according to
the underlaying structuring element B. Other more sophisticated filters are ob-
tained by combinations of openings and closings, to address problems such as
non-Gaussian denoising, image regularization, etc.
Finding the proper pipeline of morphological operators and structuring el-
ements in real applications is a cumbersome and time consuming task. In the
machine learning community there has always been lot of interest in learning such
operators, but due to the non-differentiable nature of the max/min filtering only
few approaches have been found to succeed, notably one based on LMS (gradi-
ent steepest descent algorithm) for rank filters formulated with the sign function
?
This work has been supported by ArcelorMittal, Maizières Research, Measurement
and Control Dept., France.
2

[13,14]. This idea was later revisited [12] in a neural network framework combin-
ing morphological/rank filters and linear FIR filters. Other attempts from the
evolutionary community (e.g., genetic algorithms [7] or simulated annealing [21])
use black-box optimizers to circumvent the differentiability issue. However, most
of the proposed approaches do not cover all operators. More importantly, they
cannot learn both the structuring element and the operator, e.g., [11]. This is
obviously a quite important limitation as it makes very hard or even impossible
the composition of complex filtering pipelines. Furthermore, such systems are
usually limited to a very specific application and hardly generalize to complex
scenarios.
Inspired by recent work [1] on counter-harmonic mean asymptotic morphol-
ogy, we propose a novel framework to learn morphological pipelines of operators.
We combine convolutional neural networks (CNN) with a new type of layer that
permits complex pipelines through multiple layers and therefore extends this
models to a Morphological Convolutional Neural Network (MPCNN). It extends
previous work on deep-learning while making directly applicable all optimization
tricks and findings of this field.
Here we focus on methodological foundations and show how the model learns
several operator pipelines, from dilation/erosion to top-hat transform (i.e., residue
of opening/closing). We report an important real-world application from the steel
industry, and present a sample application to denoising where operator learning
outperforms hand-crafted structuring elements.
Our main contributions are:

– a novel framework where learning of morphological operators and filtering

pipelines can be performed using gradient-based techniques, exploiting recent
insights of deep-learning approaches;
– the introduction of a novel PConv layer for CNN, to let CNN benefit from
highly nonlinear, morphology-based filters;
– the stacking of many PConv layers, to learn complex pipelines of operators
such as opening, closing and top-hats.

2 Background

Here we illustrate the foundations of our approach, introducing the Counter-

Harmonic Mean formulation for asymptotic morphology and CNN.

2.1 Asymptotic morphology using Counter-Harmonic Mean

We start from the notion of counter-harmonic mean [2], initially used [19] for
constructing robust morphological-like operators. More recently, its morpholog-
ical asymptotic behavior was characterized [1]. Let f (x) be a 2D real-valued
image, i.e., f : Ω ⊂ Z2 → R, where x ∈ Ω denotes the coordinates of the pixel
in the image domain Ω. Given a (positive) weighting kernel w : W → R+ , W
 3

being the support window of the filter, the counter-harmonic mean (CHM) filter
of order P , −∞ ≤ P ≤ ∞ is defined by,

f P +1 (y)w(x − y)dy
R
P +1
P (f ∗ w)(x) y∈W (x)
κw (f )(x) = = R , (1)
(f P ∗ w)(x) y∈W (x)
f P (y)w(x − y)dy

where f P is the image, where each pixel value of f is raised to power P , / indi-
cates pixel-wise division, and W (y) is the support window of the filter w centered
on point y. We note that the CHM filter can be interpreted as P −deformed con-
volution, i.e., κP
w (f )(x) ≡ (f ∗P w)(x). For P  0 (P  0) the pixels with largest
(smallest) values in the local neighborhood W will dominate the result of the
weighted sum (convolution), therefore morphological dilation and erosion are
the limit cases of the CHM filter, i.e., limP →+∞ (f ∗P w)(x) = supy∈W (x) f (y) =
δW (f )(x), and limP →−∞ (f ∗P w)(x) = inf y∈W (x) f (y) = εW (f )(x), where W
plays the role of the structuring element. As proven earlier [1], apart from the
limit cases (e.g., a typical order of magnitude of 5 ≤ |P | < 10), we have the
following behavior:
 
1
(f ∗P w)(x) |P 0 ≈ sup f (y) + log (w(x − y)) , (2)
y∈W (x) P
 
1
(f ∗P w)(x) |P 0 ≈ inf f (y) − log (w(x − y)) , (3)
y∈W (x) P

which can be interpreted, respectively, as the nonflat dilation (supremal convo-

lution) and nonflat erosion (infimal convolution) using the structuring function
b(x) = P1 log (w(x)). By using constant weight kernels, i.e., w(x) = 1 if x ∈ W
and w(x) = 0 if x ∈ / W , and |P |  0, we just recover the corresponding flat
structuring element W , associated to the structuring function w(x) = 0 if x ∈ W
and w(x) = −∞ if x ∈ / W.
From a precise morphological viewpoint, we notice that for finite P one can-
not guarantee that (f ∗P w)(x) yields exactly a pair of dilation/erosion, in the
sense of commutation with max/min [16,17]. Consequently, stricto sensu, we can
only name them as pseudo-dilation (P  0) and pseudo-erosion (P  0). The
asymptotic cases of the CHM filter can be also combined to approximate opening
and closing operators, i.e.,
(
P 0
((f ∗−P w) ∗P w) (x) −−−→ γW (f )(x),
P 0 (4)
((f ∗P w) ∗−P w) (x) −−−→ ϕW (f )(x).

2.2 Convolutional Neural Networks

CNN are hierarchical models alternating two basic operations, convolution and
subsampling, reminiscent of simple and complex cells in the primary visual cortex
[8]. Their main characteristic is that they exploit the 2D structure of images
via weight sharing, learning a set of convolutional filters. Certain CNN scale
4

...

CONVOLUTION POOLING FULLY CONNECTED

Fig. 1. A schematic representation of an CNN. Convolutional and pooling layers are

stacked below the fully connected layers used for classification.

well to real-sized images and excel in many object recognition [3,4,6,10] and
segmentation [5,18] benchmarks. We refer to a state-of-the-art CNN as depicted
in Figure 1. It consists of several basic building blocks briefly explained here:
– Convolutional Layer : performs a 2D filtering between input images {fi }i∈I
and a bank of filters {wk }k∈K , producing another set of images {hj }j∈J
denoted as maps. Input-output correspondences are specified through a con-
nection table CT (inputImage i, filterId k, outputImage j). Filter responses
from inputs connected to the same output image P are linearly combined. This
layer performs the following mapping: hj (x) = i,k∈CTi,k,j (fi ∗wk )(x), where
∗ indicates the 2D valid convolution. Then, a nonlinear activation function
(e.g., tanh, logistic, etc.) is applied to hj . Recently relu activations have
been found to excel. They are the units we use for all our models. A relu
unit operates as relu(x) = max(lb, min(x, ub)). It is common choice to use 0
as lower bound (lb) and ∞ as upper bound (ub).
– Pooling Layer : down-samples the input images by a constant factor keeping
a value (e.g. maximum or average) for every non overlapping subregion of
size p in the images. Max-Pooling is generally favorable, as it introduces
small invariance to translation and distortion, leads to faster convergence
and better generalization [15].
– Fully Connected Layer : this is the standard layer of a multi-layer network.
It performs a linear multiplication of the input vector by a weight matrix.
Note the striking similarity between the max-pooling layer and a dilation
transform. The former is in fact a special case of dilation, with a square struc-
turing element of size p followed by downsampling (sampling one out of every p
pixels). Our novel layer, however, does not any longer limit the pooling operation
to simple squares, but allows for a much richer repertoire of structuring elements
fine-tuned for given tasks. This is what makes MCNN so powerful.

3 Method
Now we are ready to introduce the novel morphological layer based on CHM
filter formulation, referred to as PConv layer. For a single channel image f (x)
 5

and a single filter w(x) the PConv layer performs the following operation
(f P +1 ∗ w)(x)
P Conv(f ; w, P )(x) = = (f ∗P w)(x) (5)
(f P ∗ w)(x)
It is parametrized by P , a scalar which controls the type of operation (P < 0
pseudo-erosion, P > 0 pseudo-dilation and P = 0 standard linear convolution),
and by the weighting kernel w(x), where the corresponding asymptotic structur-
ing function is given by w(x) = log (w(x)). Since this formulation is differentiable
we can use gradient descent on these parameters.
The gradient of such a layer is computed by back-propagation [20,9]. In
minimizing a given objective function L(θ; X), where θ represents the set of
parameters in the model, back-propagation applies the chain rule of derivatives
to propagate the gradients down to the input layer, multiplying them by the
Jacobian matrices of the traversed layers. Let us introduce first two partial results
of back-propagation
f (x) −f (x) · (f P +1 ∗ w)(x)
∆U (x) = ; ∆D (x) = . (6)
(f P ∗ w)(x) (f P ∗ w)(x)
The gradient of a PConv layer is computed as follows
∂L
= f˜P +1 ∗ ∆U + f˜P ∗ ∆D (7)
∂w
∂L f
= f P +1 · log(f ) · ( P ∗ w̃) + f P · log(f ) · (∆D ∗ w̃) (8)
∂P f ∗w
where f˜, w̃ indicate flipping along the two dimensions and · indicates element-
wise multiplication. The partial derivative of the PConv layer with respect to its
input (to back-propagate the gradient) is instead
∂(f ∗P w)(x)
= ∆U (x) + ∆D (x). (9)
∂f
Learning the top-hat operator requires a short-circuit in the network to allow
for subtracting the input image (or the result of an intermediate layer) from the
output of a filtering chain. For this purpose we introduce the AbsDiffLayer
which takes two layers as input and emits the absolute difference between them.
Partial derivatives can still be back-propagated.

3.1 Learning Algorithm

Minimizing a PConv layer is a non-convex, highly non-linear operation prone to
local convergence. Deep-learning findings tell us that stochastic gradient descent
is the most effective algorithm to train such complex models. In our experiments
we use its full online version where weights are updated sample by sample. The
learning rate decays during training. To further avoid bad local minima we use
a momentum term. For the opening/closing tasks we also alternate between
learning P , keeping w fixed, and vice-versa. This is common in online dictionary
learning and sparse coding. We also constrain w ≥ 0.
6

4 Experiments

We thoroughly evaluate our MCNN on several tasks. First we assess the quality of
dilation/erosion operators, which require a single PConv layer. This gives a good
measure of how well training can be performed using the CHM derivation. Then
we investigate a two-layer network learning openings/closings. This is already a
challenging task hardly covered in previous approaches.
We then learn the top-hat transform for a very challenging steel industry ap-
plication. Using 2 filters per layer we learn to simultaneously detect two families
of defects without resorting to multiple training. Our implementation allows for
learning multiple filters for every layer, thus producing a very rich set of filtered
maps. Subsequent convolutional layers can learn highly nonlinear embeddings.
(We believe that this will also dramatically improve segmentation capabilities
of such models to be investigated in future work.) We also show that a simple
CNN does not learn well pipelines of morphological operators. This is actually
expected a priori due to the nature of conventional convolutional layers, and
shows the value of our novel PConv layer.
As final benchmark we consider denoising. Our MCNN shows the superiority
of operator learning over hand-crafting structuring elements for non-Gaussian
(binomial and salt-and-pepper) noise. We also show that our approach performs
well on total variation (TV) approximation for additive Gaussian noise.
In all our experiments we use stochastic gradient descent and a filter size of
11 × 11 unless otherwise stated. The per-pixel mean-squared error loss (MSE) is
used.

4.1 Learning dilation and erosion

In this first set of experiments we create a dataset as follows: for every input
image fi we produce a target image tj using a predetermined flat structuring
element Bk and a predetermined operator: tj = δBk (fi ) or tj = εBk (fi ). We
train until convergence. Overfitting is not an issue in such a scenario. The net
actually learns the true function underlying the data. In fact, for an image of
512×512 and a structuring element with support of 11×11 there are 5022 patches,
way more than the 112 elements in the structuring element. A CNN with equal
topology fails, producing mainly Gaussian blurred images, illustrating the need
for a PConv layer to handle this kind of nonlinearities. Figure 2 shows the results
of a dilation with three structuring elements: a line of 15 pixels and 45◦ , a square
of size 5 pixels and a diamond of side 5 pixels. Figure 3 shows similar results
for the erosion transform. Note that the learned weighted kernels w(x) are
not exactly uniformly equal to 1. The corresponding morphological structuring
function w(x), obtained after applying the logarithm on the weights, produces
a rather flat shape. In practice, we observed that learning an erosion is slightly
more difficult than learning the dual dilation. This is related to the asymmetric
numerical behavior of CHM for P > 0 and P < 0. Nevertheless, in all cases the
learned operator has excellent performance.
 7

Input Target Net

Fig. 2. Examples of learning a dilation with three different structuring elements. The
target and net output are slightly smaller than the original image due to valid convo-
lution. The obtained kernel w(x) for each case is also depicted.

Target Net
Input

Fig. 3. Examples of learning an erosion with three different structuring elements along
with the learned kernel w(x).
8

4.2 Learning opening and closing

In this set of experiments we train our system to learn openings γBk (fi ) and
closings ϕBk (fi ). Learning such functions is extremely difficult. To the best of
our knowledge, we are the first to do this in a flexible and gradient-based frame-
work without any prior. For instance, in classical approaches [13] or more recent
ones [11], the operator needs to be fixed a priori.
Figure 4-top shows an example of a closing with a line of length 10 and an
orientation of 45◦ , whereas Figure 4-bottom shows an example of an opening
with a square of size 5. In both cases, the obtained kernel for the first L1 and
second L2 PConv layers are depicted. We see that the associated structuring
element is learned with a good approximation. On the other hand, however, we
also start to see that learning a flat opening/closing is remarkably hard and
that the network output starts to be slightly “blurry”. Why? On the one hand,
the obtained values of P , e.g., in the closing PL1 = 6.80 and PL2 = −8.85, in
the opening PL1 = −7.64 and PL2 = 7.07 are in the interval of asymptotically
unflat behavior. On the other hand, they are not totally symmetric. We intend
to further study these issues in ongoing work.

Input Target Net

L1

L2

L1

L2

Fig. 4. Top: an example of learning a closing operator where a line of length 10 and
orientation 45◦ is used. Bottom: opening with a square structuring element of size 5.
The network closely matches the output and almost perfectly learns the structuring
elements in both PConv layers.

4.3 Learning top-hat transform

Delegating the learning to a neural network allows for easily constructing com-
plex topologies by linking several simple modules. We recall that the white top-
hat is the residue of the opening, i.e., %+
Bk (fi ) = fi − γBk (fi ), and the black
 9

top-that the residue of the closing, i.e., %−

Bk (fi ) = ϕBk (fi ) − fi . Thus, to learn a
top-hat transform we introduce the AbsDiff layer. It takes two layers as input
and emits their absolute difference in output. Backpropagation is performed as
usual.
Top-hat is particularly relevant in real applications such as steel surface qual-
ity control. It is a powerful tool for defect detection. Here we first show that our
framework can learn such a transform. Increasing the number of filters per layer
from 1 to 2, we show that our MCNN is also much more powerful when jointly
learning two transforms. Figure 5 shows the results for a single top-hat. We cre-

Input Target Net

L1

L2

Fig. 5. Learning a top-hat transform. The defected input image has bright spots to be
detected. The network performs almost perfectly on this challenging task.

ate our training set by applying a white top-hat %+ B , where B is a disk of size
5 pixels. This operator extracts only the structures of size smaller than B and
brighter than the background. We clearly see that the network performs almost
perfectly. To further assess the advantages of a PConv layer over a conventional
convolutional layer, we also train a CNN with identical topology. The discrep-
ancy between the two models in terms of losses (MSE) is large: we have 1.28E-3
for our MCNN and 1.90E-3 for the CNN. More parameters are required for a
CNN to reach better performance. This clearly establishes the added value of
our MCNN.
The steel industry requires many top-hat operations. Tuning them one by
one is a cumbersome process. Furthermore, several models’ outputs need to be
considered to obtain the final detection result. Figure 6 shows that by simply
increasing the number of filters per layer we can simultaneously learn two top-hat
transforms and address this problem. We learn a white top-hat with a disk of size
5 and a black top-hat with a line of size 10 and orientation of 0◦ . A convolutional
layer is used to combine the output of the two operators. The architecture is as
follows: 2 PConv layers, Conv layer, AbsDiff layer with the input. The network
is almost perfect from our viewpoint. This opens the possibility of using such a
setup in more complex scenarios where several morphological operators should
be combined. This is of great interest in multiple class steel defect detection.

4.4 Learning denoising and image regularization

Finally we compare our MCNN to conventional morphological pipelines in the
denoising task. Morphological filters are recommended for non-Gaussian denois-
10

Input Target Net Input Target Net

Fig. 6. Learning two top-hat transforms. On the left, bright spots need to be detected.
On the right, a dark vertical line. The network performs almost perfectly in this chal-
lenging task.

ing. The purpose of this evaluation, however, is not to propose a novel noise
removal approach, but to show the advantages of a learnable pipeline over a
hand-crafted one. We start with binomial noise where 10% of the image pixels

Input Target Net Morph. Transf.

Closing

PSNR 15.63dB PSNR 36.07dB PSNR 33.49dB

PSNR 15.47dB PSNR 29.40dB PSNR 28.77dB

Fig. 7. Top: Binomial noise removal task. The learned nonlinear operator performs
better than the hand-crafted one. Learning uses noisy and original images—there is
no prior on the task. Bottom: Salt’n’pepper noise removal task. Even here, the learned
operator performs better than the corresponding morphological one.

are switched-off. The topology in this case is: 2 PConv layers and filter size of
5 × 5. We compare to a closing with a square of size 2, empirically found to
deliver best results. We make the task even harder by using larger-than-optimal
support. Training is performed in fully on-line fashion. While the target images
 11

are kept fixed, the input is generated by adding random noise sample by sample.
So the network never sees the same pattern twice. Figure 7–top compares the
two approaches, and shows the noisy image. We see that learning substantially
improves the PSNR measure.
We continue with an even more challenging task, a 10% salt’n’pepper denois-
ing. The network is made of 4 PConv layers, a very long pipeline. We compare
to an opening with a square of size 2 × 2 on a closing with the same structuring
element. Training follows the same protocol as the one for the binomial noise.
Images are generated online. This creates a possibly infinite dataset with very
small memory footprint. Figure 7–bottom shows results. Although we can ob-
serve some limitations of our approach, it still exhibits the best PSNR also in
this application.
Finally we consider the case of total variation (TV) restoration from an image
corrupted by 6% additive Gaussian noise. Morphological filtering does not excel
at this task. A MCNN is trained to learn the mapping from noisy image to TV
restored image. How well can it approximate any target transformation with a
pseudo-morphological pipeline? The architecture is composed of 2 PConv layers
with 2 filters each plus an averaging layer. Results are shown in Figure 8.

Input Target Net

PSNR 24.46dB PSNR 30.15dB PSNR 29.77dB

Fig. 8. Total Variation (TV) task. The network has to learn to approximate the TV
output (target) by means of averaging two filtering pipelines.

5 Conclusion and Perspectives

Our MCNN for learning morphological operators is based on a novel PConv
convolutional layer and inherits all the benefits of gradient-based deep-learning
algorithms. It can easily learn complex topologies and operator chains such as
white/black top-hats and we showed its application to steel defect detection. In
future work we intend to let MCNN simultaneously learn banks of morphological
filters and longer filtering pipelines.

References
1. Angulo, J.: Pseudo-morphological image diffusion using the counter-harmonic
paradigm. In: Proc. of ACIVS2010, LNCS Vol. 6474, Part I. pp. 426–437. Springer
(2010)
 12

2. Bullen, P.: Handbook of Means and Their Inequalities, 2nd Edition. Springer (1987)
3. Cireşan, D.C., Meier, U., Masci, J., Schmidhuber, J.: A committee of neural net-
works for traffic sign classification. In: International Joint Conference on Neural
Networks (IJCNN2011) (2011)
4. Cireşan, D.C., Meier, U., Masci, J., Schmidhuber, J.: Flexible, high performance
convolutional neural networks for image classification. In: International Joint Con-
ference on Artificial Intelligence (IJCAI2011) (2011)
5. Ciresan, D.C., Giusti, A., Gambardella, L.M., Schmidhuber, J.: Deep neural net-
works segment neuronal membranes in electron microscopy images. In: NIPS (2012)
6. Ciresan, D.C., Meier, U., Gambardella, L.M., Schmidhuber, J.: Convolutional neu-
ral network committees for handwritten character classification. In: ICDAR. pp.
1250–1254 (2011)
7. Harvey, N.R., Marshall, S.: The use of genetic algorithms in morphological filter
design. Signal Processing: Image Communication 8(1), 55–71 (1996)
8. Hubel, D.H., Wiesel, T.N.: Receptive fields and functional architecture of monkey
striate cortex. The Journal of physiology 195(1), 215–243 (March 1968)
9. LeCun, Y.: Une procédure d’apprentissage pour réseau à seuil asymétrique. Pro-
ceedings of Cognitiva 85, Paris pp. 599–604 (1985)
10. Masci, J., Meier, U., Cireşan, D.C., G., F., Schmidhuber, J.: Steel defect classi-
fication with max-pooling convolutional neural networks. In: International Joint
Conference on Neural Networks (2012)
11. Nakashizuka, M., Takenaka, S., Iiguni, Y.: Learning of structuring elements for
morphological image model with a sparsity prior. In: IEEE International Confer-
ence on Image Processing, ICIP 2010. pp. 85–88 (2010)
12. Pessoa, L.F.C., Maragos, P.: Mrl-filters: a general class of nonlinear systems and
their optimal design for image processing. IEEE Transactions on Image Processing
7(7), 966–978 (1998)
13. Salembier, P.: Adaptive rank order based filters. Signal Processing 27, 1–25 (1992)
14. Salembier, P.: Structuring element adaptation for morphological filters. J. of Visual
Communication and Image Representation 3(2), 115–136 (1992)
15. Scherer, D., Müller, A., Behnke, S.: Evaluation of pooling operations in convolu-
tional architectures for object recognition. In: International Conference on Artificial
Neural Networks (2010)
16. Serra, J.: Image analysis and mathematical morphology. Academic Press, London
(1982)
17. Soille, P.: Morphological Image Analysis: Principles and Applications. Springer-
Verlag Berlin (1999)
18. Turaga, S.C., Murray, J.F., Jain, V., Roth, F., Helmstaedter, M., Briggman, K.,
Denk, W., Seung, H.S.: Convolutional networks can learn to generate affinity
graphs for image segmentation. Neural Comput. 22(2), 511–538 (Feb 2010)
19. Vliet, L.J.v.: Robust local max-min filters by normalized power-weighted filtering.
In: Proc. of IEEE ICPR’04, Vol. 1. pp. 696–699 (2004)
20. Werbos, P.J.: Beyond Regression: New Tools for Prediction and Analysis in the
Behavioral Sciences. Ph.D. thesis, Harvard University (1974)
21. Wilson, S.S.: Training structuring elements in morphological networks. In: Math-
ematical Morphology in Image Processing, Chapter 1, pp. 1–42. Marcel Dekker
(1993)

View publication stats