Marine Snow Removal in Underwater Images
Bogdan Smolka and Monika Mendrela
Silesian University of Technology, Akademicka 16, Gliwice, Poland
Marine Snow, Image Enhancement, Noise Reduction.
In this paper two methods of marine snow detection in underwater images are presented. The proposed tech-
niques are based on the pixel corruption measures which enable the identification of clusters forming the
marine snow. As the detection of marine snow contaminating the images must be followed by an inpainting
step, various techniques which allow to restore the images with missing regions were evaluated. The experi-
ments revealed that the restoration quality of applied inpainting techniques is dependent on the image structure
and the size of regions needed to be restored and that their overall efficiency is comparable. Therefore, the
faster algorithms should be preferred. To asses the quality of marine snow removal techniques, a database of
images with 5 levels of contamination was created. The experiments performed on this database showed that
the proposed marine snow detection techniques coupled with fast inpainting methods yield very satisfactory
results, superior to the techniques already known from the literature.
Due to numerous factors contributing to the degrada-
tion of underwater images, the problem of their en-
hancement has received a lot of attention. Poor light-
ing conditions frequently have a negative impact on
underwater image acquisition. The underwater im-
ages can be of low contrast, they are frequently blurry
or hazy, have severely distorted chrominance chan-
nels, or even have completely lost their color. The nat-
ural occurrence of so called marine snow (MS), which
is an accumulation of biogenic material falling down
from the upper layers of the water column, is also a
significant source of underwater images distortions.
As plants and animals perish and decompose, ma-
rine snow grows as a result of the aggregation process
up to several centimeters in diameter. The particles
are made up of faecal matter, sand, soot and other
inorganic materials and the reflected light appear as
bright spots obscuring the underwater image scene
orgens et al., 2015; Alldredge and Silver, 1988).
Marine snow is typically regarded as a source of
noise that should be eliminated from the image to pre-
vent a decline in the efficiency of image processing
operations like object segmentation, classification or
recognition. Due to the fact that the floating particles
frequently resemble the seabed’s structures, such as
small stones and the textures of plants and animals,
the problem of MS detection and removal is challeng-
ing. In addition, the pixels that represent the detected
snowflakes should be eliminated in a way that pre-
vents artifacts from being introduced and rendering
image analysis systems less effective.
In this work, two effective techniques for detect-
ing marine snow are described and the efficacy of
various image inpainting methods which are used
to replace the noisy pixels are examined. The pro-
posed techniques of underwater image enhancement
affected by marine snow proved to be reliable and ef-
fective and can be used in real-time applications.
Marine snow varies in size, shape and brightness.
However, mostly it appears as bright, almost white
spots, randomly distributed over the image. The spots
can be of different size, but usually they are repre-
sented by circular clusters of up to 50 × 50 pixels, de-
pending on the image resolution. Their shape resem-
bles to some extent the Gaussian distribution, with in-
tensity decreasing with the distance to the peak loca-
tion. However, the maximum of intensity can be also
located outside of the spot center and very often the
snowflakes are surrounded by a dark area which high-
lights their intensity.
The process of marine snow removal can be di-
vided into two steps: detection of snow particles and
their replacement, preferably using a suitable image
Smolka, B. and Mendrela, M.
Marine Snow Removal in Underwater Images.
DOI: 10.5220/0011588200003332
In Proceedings of the 14th International Joint Conference on Computational Intelligence (IJCCI 2022), pages 463-471
ISBN: 978-989-758-611-8; ISSN: 2184-3236
Copyright © 2023 by SCITEPRESS Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)
inpainting method. An efficient algorithm of marine
snow removal should be accurate, computationally ef-
ficient and should not create artifacts which could af-
fect the performance of computer vision analysis.
A straightforward method of marine snow re-
moval can rely on the reduced vector ordering. Us-
ing this approach, for each pixel in the local filter-
ing window, the distance in a chosen color space, to
other samples from this window is calculated. Then
the cumulated distances assigned to each pixel are
sorted and the pixel for which the sum of distances
is minimized, replaces the central sample of the fil-
tering window (Astola et al., 1990). Defined in this
way Vector Median Filter, (VMF) is able to remove
outliers and clusters of pixels like marine snow parti-
cles, which are treated as impulsive noise. However,
as the central pixel is uniformly replaced by a pixel
from the local window, this procedure leads to loss
of details and overall image blurring. Additionally,
if large spots of pixels are to be filtered out, a suffi-
ciently large processing window is needed to ensure
that the vector median, which is replacing the central
pixel of the window, does not belong to the unwanted
Quite often, instead of the vectorial processing of
the image structures, channelwise (marginal) methods
are applied. In this way, instead of the vector median,
the pixel whose components are medians of the re-
spective channel values of the samples from the pro-
cessing window is declared as the filter output. As
the correlation between color channels is neglected,
the marginal median filter output generally does not
belong to the set of pixels from the filtering window,
which means that a new color is created. Such a pro-
cedure can generate artifacts, especially at the edges
of image objects, however it can be advantageous in
the case of Gaussian noise which affects all image
To decrease the amount of pixels which are unnec-
essarily changed by the median filter, a switching pro-
cedure can be applied. First, the outlying pixels are
detected and then they are replaced by the local me-
dian. A simple measure of pixel corruption is the dif-
ference between the intensity of the central pixel of a
filtering window and the median or weighted median
value (Sun and Neuvo, 1994). If the absolute differ-
ence exceeds a predefined threshold, the pixel is de-
clared as noisy. When all image pixels are analyzed,
the noisy pixels in a filtering window are replaced
with the median of uncorrupted samples in the pro-
cessing window, otherwise they remain unchanged.
This procedure can be applied iteratively, especially if
the noise contamination intensity is high (Zhou Wang
and Zhang, 1999).
(a) Test image 1 (b) Test image 2
(1) (2) (3) (4)
(c) Test images with missing regions
Figure 1: Color test images used for the evaluation of the
efficiency of inpainting methods.
To alleviate the problems caused by direct appli-
cation of image denoising algorithms, filtering meth-
ods dedicated to marine snow removal were proposed.
In (Banerjee et al., 2014) a patch-based Adaptive
Probabilistic Approach (APA) was presented. First,
the image is converted from RGB into YCbCr color
space. Then the luminance channel is processed and
the pixels which intensity exceeds an adaptive thresh-
old, based on local mean and variance, are detected
and the probability of MS occurrence is computed. In
the sequel, to avoid object misclassification, the pro-
cessing patch is enlarged and the calculations are re-
peated. If the existence of marine snow spot within
the processing window is confirmed, the central pixel
is replaced with the local marginal median. Finally,
the image with reconstructed luminance channel is
converted back into the RGB color space.
Another approach, based on a Supervised Median
Filtering scheme (SMF), is focused on removing ma-
rine snow utilizing the characteristics of snowflakes
and applying a voting mechanism (Farhadifard et al.,
2017a). In the first step, the algorithm detects the pix-
els belonging to a snowflake, analysing the distance
in the RGB color space to the remaining pixels of a
processing window. When the initial candidates are
found, then their local density is determined. The de-
tected pixels, which are grouped in clusters of low-
saturated pixels are declared as parts of a snowflake.
Pixel classification is repeated for different window
ROBOVIS 2022 - Workshop on Robotics, Computer Vision and Intelligent Systems
sizes and finally the color pixels recognized as marine
snow are replaced with the marginal median of the
samples from the local area detected as background.
When comparing SMF to APA, this procedure of-
fers slightly better results, but still it does not solve the
problem of removing circle-shaped light reflections
caused by artificial illumination reflected from marine
snow particles. It also requires predefining thresh-
olds and is time consuming due to window resizing
and voting scheme used for the pixels classification.
Therefore, an improved method of marine snow re-
moval was proposed in (Farhadifard et al., 2017b).
The detection step remained unchanged, however to
replace the pixels detected as belonging to the ma-
rine snow, instead of median filtering, an inpainting
method based on the Field of Experts concept (Roth
and Black, 2005) was utilized.
The aim of image inpainting methods is the recon-
struction of its lost or damaged parts. Inpainting is
being used in many applications including removing
inscriptions and logos from images, eliminating the
red eye effect and restoration of old, damaged pho-
tographs, among many others. In this work we use
various inpainting methods to replace the image pix-
els detected as marine snow and we also validate their
performance in terms of objective quality measures.
The most computationally complex, but very ef-
fective technique, is the Annihilating filter based
Low-rank Hankel structured matrix completion ap-
proach, known as ALOHA (Jin and Ye, 2015). It
is a patch-based method which exploits the annihila-
tion property between a shift-invariant filter and im-
age data, observed in many inpainting algorithms.
Among the existing methods, there are many
which are based on Partial Differential Equations
(PDEs) of different order. An example is an optical
flow based inpainting method called Absolutely Min-
imizing Lipschitz Extension (AMLE) (Caselles et al.,
1997). This technique solves PDE on the Riemannian
manifold for recovering missing values. Harmonic In-
painting via a Discrete Heat Flow algorithm (HARI)
(Shen and Chan, 2002) is also a method based on
solving a PDE.
Another technique called INAS belongs to a set
of algorithms based on sparse linear algebra and like
the mentioned algorithms it solves discrete PDEs
(D’Errico, 2004). Using this approach only horizon-
tal and vertical neighbours of the processed pixel are
considered. In the case of small missing elements,
this inpainting method works fast, however the pro-
cessing of each of the image channels is performed
Inpaitning via Iterative Process (INPAI) is a
method based on Discrete Cosine Transform (DCT)
and was created for automatic smoothing of multidi-
mensional incomplete data. It was adopted for the
purpose of supplementing large datasets of medical
or satellite images (Garcia, 2010). This procedure
is using the penalized least square (PLS) regression
formulated in terms of DCT. A statistical model is
created and the modeling process is completely con-
trolled by one smoothing parameter.
(1) (2) (3) (4)
Test image 1
(1) (2) (3) (4)
Test image 2
Figure 2: Comparison of inpainting results in terms of
PSNR quality measure using corrupted versions of test im-
age 1 and 2 shown in Fig. 1 c).
Modified Planar Rotator Model for Missing Data
Prediction (MPR) (
c and Hristopulos, 2018) is
anticipating the unknown values in the spatial data us-
ing Gibbs Markov Random field. This model assumes
the use of spin interaction between closest neighbors
and is suitable for data sets of big sizes.
The Mumford-Shah Inpainting with Ambrosio-
Tortorelli Approximation (MSI) (Esedoglu and Shen,
2002) belongs to a group of variational models sim-
ulating the unique macro-inpainting mechanism. The
reconstruction of damaged parts of an image is per-
formed iteratively solving the Euler-Lagrange equa-
To increase the computational efficiency of the
inpainting procedure, a fast algorithm which recur-
sively fills in the missing pixels with the average of
the pixels with known values contained in the sliding
processing window was proposed (Smolka and Men-
drela, 2020). The elaborated Mean based Method
(MEAM) can be seen as a generalization of the
fast image inpainting algorithm proposed in (Oliveira
et al., 2001).
Marine Snow Removal in Underwater Images
Test Image 1 Corrupted Image (1) ALOHA AMLE HARI
Test Image 1 Corrupted Image (1) ALOHA AMLE HARI
Figure 3: Efficiency of the inpainting results using the color test images corrupted by pattern (1) shown in Fig. 1.
We conducted a study based on two images shown
in Fig. 1 of size 300 × 300, from which some parts
were removed. The regions of removed pixels, form-
ing 4 various patterns are marked with black color in
Fig. 1c). The most important, potentially decisive
factors were preservation of the natural appearance
and recovery of image details. In addition to the vi-
sual assessment, the inpainting results were analyzed
using the widely used PSNR image quality measure.
The efficiency of the evaluated methods in terms
of PSNR has been summarized in Fig. 2. It is worth
noticing that the differences in objective restoration
measures depend not only on the image structure but
also on the shape of the missing regions. The restora-
tion efficiency of the proposed method can be also as-
sessed visually in Fig. 3.
As can be observed very good results were ob-
tained using the ALOHA inpainting when restoring
the test images. The differences in the restoration
quality were smaller in case of the second test image,
which indicates that the efficiency of inpainting meth-
ods is dependent on the image structure and also on
the size of image regions which have to be filled. The
drawback of ALOHA is its huge computational com-
plexity which renders this method unsuitable for real
time applications. In contrast, the very fast MEAM
and INAS techniques proved to yield results compara-
ble with those offered by much more computationally
expensive methods.
ROBOVIS 2022 - Workshop on Robotics, Computer Vision and Intelligent Systems
Figure 4: Exemplary images with increasing Marine Snow Intensity Level (MSIL).
Figure 5: Marine Snow test images.
As the ground truth images corrupted with marine
snow are not available, we marked manually all pix-
els depicting the MS particles in 15 carefully chosen,
representative images with resolution of 640 × 480
pixels. Figure 5 shows 3 exemplary images from the
dataset. These images were extracted from a video se-
quence from the Dataset of Marine Snow available at, which contains
more than 1000 frames. In this way a set of 15 im-
ages with carefully marked marine snow particles was
To obtain the ground truth images needed to per-
form experiments which aim was to reveal the effi-
ciency of the various inpaiting methods, the missing
regions were restored using the ALOHA algorithm,
choosing parameter settings offering visually satisfy-
ing results. In this way, the restored images were
used as a reference set, which enabled to perform both
the objective and visual evaluation of the available in-
painting techniques.
The images with marked regions of marine snow
were used to simulate pictures with increasing con-
tamination intensity. To this end, the marked regions
(marine snow flakes) from all images from the dataset
were superimposed on each of the 15 pictures restored
with ALOHA. In this way we created 15 × 15 = 225
images with added marine snowflakes.
To further increase the intensity of the marine
snow, we created a database consisting of images with
artificially superimposed MS regions taken from 2
different images from the dataset. The sources of
MS were randomly chosen 3 times for each image, so
that 45 noisy images were prepared. Additionally, for
each reference image, 3 corrupted sources containing
noise from 3, 4 and 5 other images from the dataset
were composed. In this way, we created a dataset of
5 marine snow contamination intensity levels, com-
posed of 225 + 4 × 45 = 405 images, with overlaying
marine snow. These images can be treated as contami-
nated by MS noise with available ground truth, so that
various denoising techniques can be evaluated. Figure
5 depicts examples of images with increasing marine
snow intensity levels. Below, for each noisy image
the superimposed snowflake regions are marked with
white color.
The results of the restoration of images corrupted
by marine snow obtained using the analyzed inpaint-
ing techniques are depicted in Fig. 6, which shows the
PSNR values of the enhanced images with increasing
amount of the marine snow. Again, the fast INAS and
Marine Snow Removal in Underwater Images
MEAM technique proved to yield results comparable
with the competitive inpainting techniques.
Having a database of images corrupted by artifi-
cially created marine snow, we were able to create
and evaluate methods of its detection. The first of the
elaborated MS detection techniquess is based on the
Ranked Order Absolute Difference (ROAD) statistic
(Garnett et al., 2005), which is used for determining
the degree of image pixels corruption. This method
will be referred to as ROAD based Algorithm for Ma-
rine Snow detection - RAMS.
The ROAD technique calculates the Euclidean
distance in the RGB color space between a processed
pixel and its neighbors contained in a filtering win-
dow. Then the distances are sorted and a sum of
a specified number of smallest distances serves as a
measure of pixel similarity to its local neighborhood.
In this way the outlying pixels can be easily detected
and increasing the size of the filtering window, clus-
ters of outliers can be identified. In this way, the
pixels which compose the marine snow are treated as
small pixel clusters which can be localized analyzing
the ROAD values.
1 2 3 4 5
Figure 6: Efficiency of the inpaiting methods applied to re-
store the underwater images corrupted with increasing Ma-
rine Snow Intensity Levels (MSIL) expressed with PSNR.
To increase the discriminating properties of the
proposed detector, the obtained map of pixel contam-
ination was subjected to the morphological Top-Hat
operation, which enables to diminish the influence of
image texture on the final marine snow detection re-
sult. The detection step needs a thresholding value
which was obtained using the triangle binarization
method (G. Zack, 1977).
In this way, if the result of the Top-Hat operation
performed on the image depicting the ROAD values
exceeds the automatically determined threshold, then
a pixel is considered as belonging to a cluster of pixels
that make up a snowflake, which should be restored
using an inpainting technique. Otherwise, the pixel is
classified as background and is left unchanged. The
proposed method needs 2 parameters: the size r of
the processing window containing r × r pixels and the
number of smallest distances denoted as α, taken for
the calculation of ROAD based dissimilarity. The per-
formed experiments showed that satisfactory results
were obtained using r = α = 7.
The second method used for the detection of ma-
rine snow is based on the distance transformation and
will be called Distance Transform based Algorithm
for Marine Snow - DITAMS. This method determines
for every image pixel the minimal cost of a digital
path joining it with the boundary of a sliding process-
ing window.
The path cost is calculated as the sum of transi-
tions between adjacent pixels expressed through the
Euclidean distance in the RGB color space. Pixels
belonging to a cluster of pixels which differ signif-
icantly from their surrounding are not connected by
a low-cost path and therefore they can be easily de-
tected using the triangle binarization technique. To
speed up the process of distance transform calcula-
tion, a two-pass algorithm proposed by Rosenfeld and
Pfaltz has been applied (Rosenfeld and Pfaltz, 1968).
Figure 7 shows a test image and the binarization
result using the triangle algorithm. The results ob-
tained using the RAMS method are visually very sim-
ilar. The performed experiments revealed that sat-
isfying efficiency of the proposed DITAMS scheme
was obtained for the window size r = 7, which makes
the proposed scheme fast enough to be applied in real
time MS detection systems.
Figure 8 compares the Accuracy, Precision, Re-
call and Specificity of the APA, SMF, RAMS and
DITAMS marine snow detection techniques. To de-
scribe the efficiency measures, the following notation
is used: Q - total number of image pixels, P - num-
ber of noisy pixels, N - number of undisturbed pixels
(Q=P+N), TP - pixels correctly classified as noise, TN
- pixels correctly classified as uncorrupted, FP - pixels
which were not correctly classified as noise.
Accuracy is a basic measure, which refers to the
ratio of the number of correctly classified pixels in an
image to their total number - (TP+TN)/Q. The Accu-
racy values are high for all of the analyzed methods
but the best results were obtained using DITAMS and
Precision specifies the ratio of the number of pix-
els correctly classified to the class of noisy pixels in
relation to all pixels classified as noise, which in-
cludes misclassifications as well - TP/(TP+FP). The
experiments showed that the DITAMS was the most
ROBOVIS 2022 - Workshop on Robotics, Computer Vision and Intelligent Systems
(a) Test image
(b) Binarized distance map
Figure 7: Binarization of the output of the distance trans-
form based marine snow detector.
precise among the evaluated methods. It can be also
concluded that the APA and SMF methods are not the
best choice if high precision of MS detection is re-
Next measure describing the efficiency of the pro-
posed techniques is Recall, which specifies the ratio
of the number of pixels correctly classified as noise
to all noisy pixels in the image - TP/P. The highest
value of this parameter is obtained for the RAMS al-
gorithm. The APA method is not much worse and is
followed by SMF and DITAMS. The lower values of
this parameter is caused by the high number of incor-
rectly classified pixels, which means that in the detec-
tion stage, there are a lot of pixels which belong to the
marine snowflakes but are classified as background.
Last measure of the efficiency of the compared
methods is the Specificity, (true negative rate) - TN/N,
which specifies the ratio of the pixel number correctly
classified as not belonging to the MS to the number of
background pixels. For all the compared techniques,
specificity measure achieves very high values. For
the DITAMS detection method it is almost equal to
1 which indicates that in this respect the algorithm
1 2 3 4 5
1 2 3 4 5
Figure 8: Boxplots summarizing the performance of 4 de-
tection methods for the increasing marine snow intensity
levels (MSIL).
Marine Snow Removal in Underwater Images
1 2 3 4 5
Figure 9: Efficiency of MS removal evaluated using PSNR
quality measure for increasing marine snow intensity levels.
works almost perfectly.
Figure 9 depicts the efficiency of the MS detec-
tion and removal through the application of MEAMS
and INAS inpainting methods in terms of PSNR mea-
sure. As can be observed the proposed RAMS and DI-
TAMS methods yield comparable results, which are
much better than those offered by the APA and SMF.
Additionally, the plots show that the fast MEAM in-
painting algorithm provides results similar to those
achieved using the INAS interpolation, which make
the proposed algorithms very attractive, especially in
the case of real-time applications.
Figure 10 exhibits results of marine snow removal
using the APA and SMF methods compared with
those obtained with the combination of RAMS and
DITAMS detection methods and MEAM and INAS
inpainting techniques. As can be observed the ma-
rine snow particles are removed and the image is
well restored using the novel, proposed techniques.
As both the MS detection techniques and inpainting
methods are computationally efficient, the new en-
hancement frameworks can be applied for real time
imaging tasks.
In this paper two novel methods of marine snow de-
tection and their removal with the use of fast inpaint-
ing methods have been proposed. The first one is
based on the ROAD measure of pixel impulsiveness
and the second is determining the cost of a connec-
tion between the central pixel and the boundary of
the filtering window. The quality of inpainting tech-
Noisy image MSIL 5
APA, 38.9 dB SMF, 40.6 dB
RAMS+MEAM, 44.0 dB RAMS+INAS, 44.2 dB
Figure 10: Result of marine snow removal expressed with
PSNR obtained for exemplary MSIL 5 image.
niques, which were applied to restore the underwater
images with detected clusters of pixels forming the
marine snow particles, was evaluated using the objec-
tive PSNR quality measure and the results were also
assessed visually. Extensive experiments confirmed
that the developed marine snow techniques coupled
with the fast MEAM and INAS inpainting methods
offer better image restoration results than the algo-
rithms already known from the literature. The devel-
oped techniques are computationally inexpensive and
can be applied in real time applications. Although the
elaborated algorithms are intended for marine snow
removal, they can be applied in various applications
in which the detection and inpainting of small clus-
ters of pixels is required.
ROBOVIS 2022 - Workshop on Robotics, Computer Vision and Intelligent Systems
This work was supported by the Silesian Univer-
sity of Technology, Poland, (grant BK 2022) and
the Polish National Science Centre under the project
Alldredge, A. L. and Silver, M. W. (1988). Characteristics,
dynamics and significance of marine snow. Progress
in Oceanography, 20(1):41–82.
Astola, J., Haavisto, P., and Neuvo, Y. (1990). Vector me-
dian filters. Proceedings of the IEEE, 78(4):678–689.
Banerjee, S., Sanyal, G., Ghosh, S., Ray, R., and Shome,
S. N. (2014). Elimination of marine snow effect from
underwater image-an adaptive probabilistic approach.
In Students’ Conference on Electrical, Electronics and
Computer Science, pages 1–4. IEEE.
Caselles, V., Morel, J.-M., and Sbert, C. (1997). An ax-
iomatic approach to image interpolation. In Proceed-
ings of International Conference on Image Process-
ing, volume 3, pages 376–379. IEEE.
D’Errico, J. (2004). Inpaint nans. MATLAB Central File
orgens, A. L., Ahmerkamp, S., M
ussig, J., Stocker, R.,
Kuypers, M. M., Khalili, A., and Kindler, K. (2015).
A laboratory model of marine snow: Preparation and
characterization of porous fiber particles. Limnology
and Oceanography: Methods, 13(11):664–671.
Esedoglu, S. and Shen, J. (2002). Digital inpainting based
on the Mumford–Shah–Euler image model. European
Journal of Applied Mathematics, 13(4):353–370.
Farhadifard, F., Radolko, M., and U. Freiherr von Lukas
(2017a). Single Image Marine Snow Removal based
on a Supervised Median Filtering Scheme. In VISI-
GRAPP (4: VISAPP), pages 280–287.
Farhadifard, F., Radolko, M., and von Lukas, U. F. (2017b).
Marine snow detection and removal: Underwater im-
age restoration using background modeling. In Proc.
of WSCG 2017, pages 81–89.
G. Zack, W. E. Rogers, S. L. (1977). Automatic mea-
surement of sister chromatid exchange frequency.
The Journal of Histochemistry and Cytochemistry,
Garcia, D. (2010). Robust smoothing of gridded data in one
and higher dimensions with missing values. Compu-
tational statistics & data analysis, 54(4):1167–1178.
Garnett, R., Huegerich, T., Chui, C., and Wenjie He (2005).
A universal noise removal algorithm with an impulse
detector. IEEE Transactions on Image Processing,
Jin, K. H. and Ye, J. C. (2015). Annihilating filter-
based low-rank Hankel matrix approach for image in-
painting. IEEE Transactions on Image Processing,
Oliveira, M. M., Bowen, B., McKenna, R., and Chang, Y.-S.
(2001). Fast digital image inpainting. In Proc. of the
Int. Conf. on Visualization, Imaging and Image Pro-
cessing (VIIP 2001, pages 261–266.
Rosenfeld, A. and Pfaltz, J. L. (1968). Distance functions
on digital pictures. Pattern Recognition, 1(1):33–61.
Roth, S. and Black, M. J. (2005). Fields of Experts: A
framework for learning image priors. In IEEE Conf.
on Computer Vision and Pattern Recognition, vol-
ume 2, pages 860–867.
Shen, J. and Chan, T. F. (2002). Mathematical models for
local nontexture inpaintings. SIAM Journal on Ap-
plied Mathematics, 62(3):1019–1043.
Smolka, B. and Mendrela, M. (2020). On the efficiency of
inpainting methods for marine snow removal in under-
water images. In Proc. of Int. Conf. on Advances in
Signal Processing and Artificial Intelligence, ASPAI,
pages 185–190.
Sun, T. and Neuvo, Y. (1994). Detail-preserving median
based filters in image processing. Pattern Recognition
Letters, 15(4):341–347.
Zhou Wang and Zhang, D. (1999). Progressive switching
median filter for the removal of impulse noise from
highly corrupted images. IEEE Transactions on Cir-
cuits and Systems II: Analog and Digital Signal Pro-
cessing, 46(1):78–80.
c, M. and Hristopulos, D. T. (2018). Gibbs Markov
random fields with continuous values based on the
modified planar rotator model. Physical Review E,
Marine Snow Removal in Underwater Images