ERSIP Assignment

Rabina Twayana — 07.12.2025

Mathematical Morphology (MM) from the 1980s

The main purpose of this report is to review the development of mathematical morphologgy from
1980s over the time by analyzing papers from the 1980s and 2000s. This report discusses the
promises, challenges of MM, its evolutions over time and evaluation of the limitations of proposed
approaches with current knowledge.

Mathematical Morphology in 1980s: Promises and Limitations

Destival (1986) introduced the mathematical morphology concept, a numerical treatment for auto-
matic detection of relevant shapes and structures from the remote sensing imageries. This method
analyze the shapes in global way unlike classical method which are based on linear transforma-
tion and limited to punctual treatment (pixel-by-pixel, using small neighborhood around the pixels).
Whereas in MM, objects under study is represented by sets and analyze using the set operations
(union, intersections) etc. The main aim of this work was to detect linear networks (roads and vil-
lages) precisely and automatically as possible.

Flouzat (1989) provides comprehensive analysis on how mathematical morphology contributed to

extract information from an image. It shows the the application of MM from low-level image pro-
cessing to high-level structural analysis.

Promises of Mathematical Morphology in the 1980s

a. Linear feaures such as ridge, valley and roads can be exracted using complex morphological
operations appropriately and sequentially to well adapted spatial resolution images (Destival, 1986;
Flouzat, 1989).

b. Mathematical morphology at the low level of image processing offers several use cases, as high-
lighted by Flouzat (1989). These include:

• Enhancing early vision: MM supports in image analysis by enhancing the observed phenomenon
representation. For instance, different Alternated Sequential Filters (ASF) families, such as
multidirectional and comparative filters were used to effectively remove speckle noise in radar
imagery caused by coherent illumination by preserving local mean effectively.
• Extracting textural information: Morphological filters can capture the texturual information
indirectly from images. In addition to this, it can be used to encnhance flat grey-level segments
after conventional textural analysis.
• Edge detection: Morphological gradient can be used to detect edges using erosion and dilation
operators .
• Spectral classification: Morphological methods (SKIZ or distance computation) can be applied
for classification of the points (pixels) represented in a 2D spectral feature space.

c. The benefits that can be obtained from mathematical morphology at the connected-component
level, as described by Flouzat (1989), are as follows:

• Mathematical morphology can recostruct and segment an image.

• MM supports enhancement through agglomeration and size filtering based on closing followed
by increasing openings.
• MM is very efficient for principal direction analysis.

1
 ERSIP Assignment
Rabina Twayana — 07.12.2025

• MM supports filtering based on length of objects using skeleton.

d. Flouzat (1989) claims that mathematical morphology provides is a powerful methods for quanti-
fying and describing image based on the structures characteristics. This includes:

• Directional filtering method provides information about angular distribution

• Intercept measurement indicate global spatial arrangement
• Morphological compacity index provides global information about shapes and edges
• MM also enables symbolic graph construction which represents spatial arrangement of objects
by the measurement of adjacency lengthes. It makes possible to make a comparision of graphs
derived from the whole image or part of an image.

Advantages
• Morphology is a complementary method with other digital image processing steps that improves
overall image analysis (Flouzat, 1989).
• Morphological algorithms can be broken down into elementary operators which enable imple-
mentation with very high speed of execution time (Flouzat, 1989).

Limitations of proposed approaches

• It is difficult to define criteria that is valid for entire image for extracting information (Destival,
1986).
• It is also hard to implement the approaches that are fully automatic as human knowledge and
decisions play a crucial role in interpreting ambiguous cases or cleaning up results throughoy
the operations (Destival, 1986).
• Computers have limited ability as compared to human eye to interpret spatial features or patterns
in an image as human has ability to keep in sight whole image considering numerous criteria
(Destival, 1986).

Comparison with Modern Approaches

• However, detection of objects/features is dominated by deep learning in recent days, they are
still not fully automatic. Specially, MM is still used massively in post-processing step to refine
detection output such as removing noise, filling gaps, and smoothing object boundaries and
many more.
• In 1980s, symbolic graph construction methods was manual or semi-automatic, but now fully
automatic graph creation is possible with the rise of advanced image processing techniques
which can be feed into Graph Neural Network (GNN) pipeline for further analysis.

2
 ERSIP Assignment
Rabina Twayana — 07.12.2025

Mathematical Morphology in the 2000s: Challenges, Evolution, Promises and

Limitations
Pesaresi and Benediktsson (2001) proposed new segmentation method based on morphological char-
acteristics (morphological leveling, morphological spectrum, and morphological profiles) of con-
nected components in images.

Soille and Pesaresi (2002) discusses how mathematical morphology can be used to improve the
analysis of geoscience and remote sensing data for various task such as filtering, simplification,
directional segmentation, and crest line extraction.

Aptoula and Lefèvre (2007) provides the comparision of different ordering strategy while applying
multivariate mathematical morphological operators in color or multi-spectral images, particularly
for noise reduction, texture classification, and pixel-wise classification.

Challenges faced by the MM community in the 2000s

• Watershed segmentation (based on edge-detection strategy) causes severe over segmentation
due to presence of local minima, local maxima and textural effects. To overcome this issue,
marker-based watershed was introduced as a standard nonlinear solution but marker detection
was very hard in this approach. Moreover, watershed-plus-marker appraoch assumes large and
homogeneous objects but in fine-resolution remote sensing images, where features are smaller
and irregular (for instance, segmenattion of urban scenes), it becomes highly unstable (Pesaresi
and Benediktsson, 2001).
• Within opening and closing reconstruction, leveling approach requires preliminarily partition-
ing of the image into two sets (one for opening by reconstruction and another for closing by
reconstruction) which is the main problem in this approach. Following leveling approach, ob-
jects are defined based on morphological characteristics such as flat, concave or convex, thus
avoiding the problem of preliminary partitioning into separate sets (Pesaresi and Benediktsson,
2001).
• In real application, single-SE-size approach appears to me very simplistic as some structures
may have high response for a givel SE size and lower response for other SE size. Thus, mul-
tiscale processing is proposed to overcome this problem. This ensures the adaptive, robust
segmentation of structures with varying structure sizes (Pesaresi and Benediktsson, 2001).
• The extension of MM from binary and grey scale images to multivariate (e.g., color, multispec-
tral) images not straightforward. This is because, classical framework of MM were based on
complete lattices, which require a total ordering for extrema computation. However, multival-
ued vectors do not have a natural ordering, making it difficult to extend MM operators without
introducing ordering schemes. As a result, most of the avalilable vector ordering approaches
were not widely accepted. Therefore the multiple experiment were carried out to compare the
ordering approaches on various application (noise reduction, texture classification and pixel
classification).

Evolution of state-of-the-art since the 1980s

• Edge detection method does not performs well on fine resolution satellite imageries where thin,
nested, or low-contrast regions must be retained. The emergence of morphological and multi-
scale methods helps to solve this problem by analyzing the structural behavior of connected
components across multiple scale (Pesaresi and Benediktsson, 2001).
• Application of morphological approach expanded across DEM (extraction of catchment basins,

3
 ERSIP Assignment
Rabina Twayana — 07.12.2025

watersheds, and drainage network), masking of sea-ice, road network extraction using SAR,
segmentation of SAR images, extraction of built-up areas in SPOT-P images, detection of grav-
ity anamolies using directional morphological filters, and determination of road orientation,
crest line extraction etc (Soille and Pesaresi, 2002).
• The progression of the MM framework from univariate (Grayscale) images to multivariate (vec-
tor) images with more focus on defining ordering of vectors such as marginal, bit mixing, lexi-
cographical, conditional, reduced etc (Aptoula and Lefèvre, 2007).

New promises in the 2000s

• The authors show segmentation examples for satellite images and demonstrate that their method
works well in low-contrast scenes and low spatial resolution (Pesaresi and Benediktsson, 2001).
• For the segmentation of complex scene like urban areas, morphological reconstruction by open-
ing and closing operations will detect structures more effectively as compared to traditional
edge-based methods (Pesaresi and Benediktsson, 2001).
• Self-dual and self complementary filters are more efficient for EO data based application, as
it avoid dependence in terms of local contrast of searched image objects (Soille and Pesaresi,
2002).
• Operators such as rank openings and dual rank closings is very efficient for reducing speckle
noise in SAR while retaining linaer features (Soille and Pesaresi, 2002).
• Self-dual level based on area openings and closings is useful in removing irrelevant features
depending on their shape and size before applying segmentation to avoid oversegmentation
problem. levelings simplify images while preserving significant edges and elongated structures,
aiding visualization, compression, and reducing over-segmentation (Soille and Pesaresi, 2002).

Limitations of proposed approaches

• The drawback of the proposed multi-scale morphological segementation method is the necessity
of applying a range SE sizes of increasing opening and closing by reconstruction operations,
which may cause a heavy computational burden (Pesaresi and Benediktsson, 2001).
• The self-complementarity operators such as gradient-based processing and top-hat tranforma-
tion is only applicatble if the search regions of the image is thick enough to have core with low
gradient values (Soille and Pesaresi, 2002).
• Different ordering approaches have its own strengths and weaknesses. The choice of ordering
is highly data and application dependent (Aptoula and Lefèvre, 2007).

Comparison with Modern Approaches

• Despite advances, there is still no universally accepted ordering scheme that works well for all
types of multivariate images and and different applications. The best choice often depend on
number of experimentation.
• The performance of morphological operators is highly sensitive to the choice of parameters
(e.g., reference vectors, distance metrics) and requires numerous experiments to tune and achieve
desired results.
• Many modern approaches, especially those based on distance-based orderings, require extensive
calculations, which can be computationally expensive, particularly when working with high-
dimensional data (e.g., hyperspectral images which has 100s of bands).
• Although object detection and segmentation tasks improved with modern deep learning tech-
niques, multi-scale object representation challenges still persists.

4
 ERSIP Assignment
Rabina Twayana — 07.12.2025

References
Aptoula, E. and Lefèvre, S. (2007). A comparative study on multivariate mathematical morphology.
Pattern Recognition, 40(11):2914–2929.

Destival, I. (1986). MATHEMATICAL MORPHOLOGY APPLIED TO REMOTE SENSINGt.

Flouzat, G. (1989). Review On Image Analysis With Mathematical Morphology In Remote Sensing.
In 12th Canadian Symposium on Remote Sensing Geoscience and Remote Sensing Symposium,,
volume 4, pages 2424–2429, Vancouver, Canada. IEEE.

Pesaresi, M. and Benediktsson, J. (2001). A new approach for the morphological segmentation
of high-resolution satellite imagery. IEEE Transactions on Geoscience and Remote Sensing,
39(2):309–320.

Soille, P. and Pesaresi, M. (2002). Advances in mathematical morphology applied to geoscience and
remote sensing. IEEE Transactions on Geoscience and Remote Sensing, 40(9):2042–2055.