Morphological Filter Detector for Image Forensics Applications

Abstract

Mathematical morphology provides a large set of powerful non-linear image operators, widely used for feature extraction, noise removal or image enhancement. Although morphological filters might be used to remove artifacts produced by image manipulations, both on binary and graylevel documents, little effort has been spent towards their forensic identification. In this paper we propose a non-trivial extension of a deterministic approach originally detecting erosion and dilation of binary images. The proposed approach operates on grayscale images and is robust to image compression and other typical attacks. When the image is attacked the method looses its deterministic nature and uses a properly trained SVM classifier, using the original detector as a feature extractor. Extensive tests demonstrate that the proposed method guarantees very high accuracy in filtering detection, providing 100% accuracy in discriminating the presence and the type of morphological filter in raw images of three different datasets. The achieved accuracy is also good after JPEG compression, equal or above 76.8% on all datasets for quality factors above 80. The proposed approach is also able to determine the adopted structuring element for moderate compression factors. Finally, it is robust against noise addition and it can distinguish morphological filter from other filters.

Full text

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2965745, IEEE Access
Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.DOI
Morphological Filter Detector for
Image Forensics Applications
GIULIA BOATO1,3 (Member, IEEE), DUC-TIEN DANG-NGUYEN2(Member, IEEE) and
FRANCESCO G.B. DE NATALE1,3(Senior Member, IEEE)
1University of Trento, Italy (e-mail: giulia.boato@unitn.it; francesco.denatale@unitn.it)
2University of Bergen, Norway(e-mail: ductien.dangnguyen@uib.no)
3Italian Consortium of Telecommunications (CNIT)
Corresponding author: Giulia Boato (e-mail: giulia.boato@unitn.it).
This work has been supported by the PRIN PREMIER project.
ABSTRACT Mathematical morphology provides a large set of powerful non-linear image operators,
widely used for feature extraction, noise removal or image enhancement. Although morphological filters
might be used to remove artifacts produced by image manipulations, both on binary and graylevel
documents, little effort has been spent towards their forensic identification. In this paper we propose
a non-trivial extension of a deterministic approach originally detecting erosion and dilation of binary
images. The proposed approach operates on grayscale images and is robust to image compression and other
typical attacks. When the image is attacked the method looses its deterministic nature and uses a properly
trained SVM classifier, using the original detector as a feature extractor. Extensive tests demonstrate that
the proposed method guarantees very high accuracy in filtering detection, providing 100% accuracy in
discriminating the presence and the type of morphological filter in raw images of three different datasets.
The achieved accuracy is also good after JPEG compression, equal or above 76.8% on all datasets
for quality factors above 80. The proposed approach is also able to determine the adopted structuring
element for moderate compression factors. Finally, it is robust against noise addition and it can distinguish
morphological filter from other filters.
INDEX TERMS Digital Image Forensics, Media Authentication, Morphological Filter Detection
I. INTRODUCTION
In the last decade, researchers and practitioners in multi-
media forensics have been developing a substantial body of
knowledge and techniques targeted to the authentication of
multimedia objects and their processing history recovery [1]–
[5]. A recent trend tries to define universal detectors able
to reveal manipulations independently from the type of pro-
cessing applied, which could leverage media authentication
in applications like journalism or social media analysis [6].
On the other hand, many methods have been proposed to
detect different types of forgeries, which is very relevant for
diverse applications. In particular, first it is crucial in digital
investigations, given that images, audio tracks and video
sequences now play a crucial role where they often represent
digital evidences to the court [7]. Secondly in multimedia
data phylogeny, which aims at recovering and tracing back
the life-cycle of an image or a video [8]–[11].
This broad class of specific manipulation detectors in-
cludes the identification of pasted regions [12]–[16], resizing
[17], [18], re-compression [19], image enhancing [20], in-
consistencies in the geometry and illumination of the image
due to possible manipulations [21]–[23], and various types of
non-linear filtering (especially median) [24]–[35].
In the context of non-linear filtering detection very lit-
tle attention was given to morphological filters [36] often
used in image processing for artifacts removal and image
enhancement [37] [38]. The detection of this kind of filtering
is of interest in the context of both image phylogeny and
specific tampering identification in legal scenarios, but could
be very useful also to detect possible counter forensic attacks
based on morphology, where such filters, very powerful in the
removal of local noise, could be exploited at the end of the
image manipulation process to cover other types of traces.
In this paper we present a non trivial extension of a recent
work [39] which introduced a deterministic detector of ero-
sion and dilation in binary images. The proposed extension
works on grayscale images by detecting morphological filters
application in an accurate way both in uncompressed and
VOLUME 4, 2016 1

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2965745, IEEE Access
Boato et al.: Morphological Filter Detector for Image Forensics Applications
compressed images. The method also allows for erosion
versus dilation discrimination and in many cases also for the
adopted structural element identification. Robustness against
JPEG compression, noise addition and confusion with other
types of filters is also tested on various datasets.
The rest of the paper is organized as follows: Section II
provides the theoretical background for the problem for-
mulation; Section III describes the proposed methodology
for morphological filtering detection in uncompressed and
compressed images; Section IV describes the experimental
setup, datasets and scenarios adopted for testing and valida-
tion; Section V details the experimental analysis and obtained
results; finally, Section VI reports some concluding remarks.
II. THEORETICAL FORMULATION
In this section we introduce the mathematical formulation of
the problem, and we derive the proposed methodology for
morphological filter detection.
A. MATHEMATICAL FORMULATION AND PROPERTIES
Mathematical morphology defines a set of nonlinear filters
commonly employed in digital image processing to modify
the local structural content of images. All the morphological
filters are derived from the various combinations of two
basic operators, erosion and dilation, and a kernel mask (or
shortly, a kernel) called structuring element, characterized
by a shape, a size, and a reference point. The shape and
the size of the kernel are responsible for the behavior of the
operator on the image, while the reference point just defines
the shift of the filtered image with respect to the original. The
invention of such mathematical tools dates back to 1964 [36],
and was meant to the filtering of binary images for mineral
studies. Later studies [40] led to the generalization of the
theory to the case of grayscale images.
According to this theory, given a grayscale image f(x, y)
and a binary structuring element B, the two fundamental
morphological operators, erosion and dilation, are respec-
tively defined as:
fB=minB(f(x, y)∩Bxy )(1)
f⊕B=maxB(f(x, y)∩Bxy )(2)
where Bxy represents the structuring element (kernel) B
with the reference point centered at the coordinates x, y of
the image plane, while the intersection operations returns the
subset of the image pixels overlapped with the 1s of B. In
this respect, the basic grayscale operators are particular cases
of rank-order filters, and behave very similarly to min-max
(see examples in Figure 1) and median operators, except for
the shape of the mask.
As in binary morphology, the composition of erosion and
dilation allows defining more complex filters, among which
the most common are the open and close operators, respec-
tively defined as follows:
(f◦B)=(fB)⊕B(3)
FIGURE 1: Example of grayscale erosion (bottom left) and
dilation (bottom right) of an image detail (top left), using
a cross-shaped structuring element (top right). The resulting
patches show the effect of min and max operations: erosion
produces a darker version of the original image eliminating
small cross-shaped details, while dilation produces the oppo-
site effect.
(f•B)=(f⊕B)B(4)
Also the mathematical properties of morphological
grayscale operators match the ones of the corresponding
binary operators. Consequently, the theoretical background
of [39] remains valid also for in the grayscale domain.
In particular, the following properties are exploited in the
construction of the proposed detector:
(i) Translation invariance: the position of the reference
point only affects the translation of the filtered image
(ii) Dilation commutativity: A⊕B=B⊕A
(iii) Associativity: a cascade of erosions (dilations) is equal
to the erosion (dilation) with a mask generated by
dilating each other the original masks (A,Band Care
binary structuring elements)
ABC=A(B⊕C)(5)
A⊕B⊕C=A⊕(B⊕C)(6)
(iv) Open and close idempotence: iterating open and close
with the same structural element does not produce
additional changes in the image
A◦B◦B=A◦B(7)
A•B•B=A•B(8)
Additionally, it is easy to see that the two theorems intro-
duced in [39] remain valid, since their demonstrations do not
depend on Eq. 1 and Eq. 2, but only on the properties (i)-(iv),
which hold also for grayscale images.
2VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2965745, IEEE Access
Boato et al.: Morphological Filter Detector for Image Forensics Applications
12345678910 11 12 13 14 15
24 25
16
17 18 19 20 21 22 23
26 27 28 29 30 31 32 33 34 35 36
FIGURE 2: Set of structuring elements (kernels) used for simulations. In dark, structural element’s reference point.
Erosion Dilation
ErosionDilation
Structuring
Elements
XOR Σ=0
XOR Σ=0
Input
Image
Scan next element ...
Dilation
Detected
Erosion
Detected
Y
Y
N
N
FIGURE 3: Proposed detection scheme for raw grayscale images. After opening/closing with a kernel K, if the image does not
change, this means that it was previously dilated/eroded with that kernel (details in Section II.A). By applying the procedure
for a set of possible masks, the detector is thus able to reveal a perfect match between the input eroded/dilated image and the
corresponding opened/closed version.
Theorem 1: Let I0=IK, then I0•K=I0. Respectively,
if I0=I⊕K, then I0◦K=I0.
Theorem 2: Let I0=IK, then ∀Msuch that ∃E|M⊕E=
K, we have that I0•M=I0. Respectively, if I0=I⊕K,
then I0◦M=I0.
Theorem 1 can be equivalently formulated in terms of
series of erosion and dilation operators, according to the
definition of open and close operators. Theorem 2 extends
the equality of Theorem 1 to any kernel mask Mthat can
produce Kby dilation with an appropriate kernel E.
As an immediate consequence of the above theorems, an
image dilated (eroded) with a given kernel K, will remain
unchanged after applying an open (close) operator with the
same element. This provides a simple test to detect a filtered
image: apply an open (close) operator with a kernel K, if
the image does not change, this means that it was previously
dilated (eroded) with that kernel, otherwise it was not. The
detection consists then in subsequently checking with the
above procedure a set of possible masks. In [39], a set of
common kernels were proposed, characterized by some level
of symmetry (see Figure 2).
III. PROPOSED APPROACH
In this section we introduce the proposed detector, distin-
guishing two cases: the detection of morphological filtering
on raw images, and the detection in the presence of post-
processing (e.g. compression, noise addition, filtering). We
will see that the former is a trivial extension of the binary case
but has very limited applicability, while the latter requires
further attention.
A. FILTER DETECTION ON RAW IMAGES
According to the theory stated in Section II, in the absence
of post-processing grayscale morphological operators can be
easily and deterministically detected by applying the schema
proposed in [39] and reported in Figure 3.
However, it is worth mentioning that this scenario is rarely
verified in grayscale images, which are typically stored in
compressed format after filtering. The compression (as well
as most other image processing operations) modifies the
image, thus hindering the applicability of Theorem 1 and
thus the applicability of the deterministic method depicted
in Figure 3. Indeed, the detector will never report a perfect
match between the input eroded (dilated) image and the
corresponding opened (closed) version.
In the next section, we propose an extension of the above
VOLUME 4, 2016 3

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2965745, IEEE Access
Boato et al.: Morphological Filter Detector for Image Forensics Applications
approach that is able to deal with such scenario. The new
algorithm will exploit the traces left by the morphological
filters even after post-processing, and will rely on a statistical
analysis, thus loosing its deterministic nature. We will see
however that it is sufficiently robust to many attacks, and in
particular to the JPEG compression.
B. FILTER DETECTION ON POST-PROCESSED IMAGES
As mentioned in the previous section, any further processing
on the filtered image will modify the pixel values, possibly
re-introducing structures that were eliminated by the mor-
phological operator. A further morphological filtering will
then produce a non-null effect on the image, which will be
revealed by the detector. It is therefore necessary to verify if
some traces of the original filter survive the post-processing.
Recalling that the grayscale morphology operates as a
local min-max filter, we can expect that it will produce
larger variations in the presence of high-contrast structures
that match the structural element geometry. On the contrary,
unless the post-processing is meant to be extremely visible,
it will introduce small gray level variations on the image.
Consequently, although we will not have a null difference
image at the detector output, we can expect that the local
differences will be much smaller for a filtered image than
for an original one. An experimental evidence of this fact is
provided in Figure 4. Here we plot the log-scale histograms
of the absolute differences before and after the application
of an open operator to a JPEG compressed unfiltered image,
and to the relevant JPEG compressed dilated image (the two
right most columns). Both open and dilation operators have
been applied using the kernel mask 35 in Figure 2, and the
JPEG quality factor has been set to QF = 95. It is possible
to observe that the histogram referred to the dilated image
decreases steeply, with significant bins only for low values of
the difference, while the histogram referred to the unfiltered
image shows a long tail with significant values also above 50.
This means on one side that the compression does not affect
the high-contrast structures present in the original image
(which are then removed by the following open operator),
and on the other side that it does not re-introduce in the
filtered image any high-contrast structure sensitive to the
filter itself. Similar results on the eroded version and the close
operator are shown on the two left most columns.
On this basis, we propose to modify the detection scheme
as shown in Figure 5. The core of the procedure still involves
the application of grayscale opening/closing. In this case,
however, we take into consideration the statistical properties
of the differences between the input and output images, to see
how such residual is distributed. Therefore, we calculate the
histogram of the difference image (processing block HIST in
Figure 5) and feed it into a statistical classifier to perform the
decision. As far as the classifier is concerned, we adopted a
properly trained classifier.
Finally, we should notice that not all the areas of an image
are equally affected by a morphological filter. In particular,
the filter has a negligible effect on flat areas, thus possibly
jeopardizing the results of the detector. To avoid this problem
we decided to limit the analysis to the image regions that
contain significant textures or edges. To this purpose, we
calculate the block-wise normalized local variance α,i.e.,
for each block, we normalise all the pixel values to [0,1]
and then compute the standard variance, and we restrict the
computation of the histogram to the blocks with α > αth,
where αth has been empirically set to 0.15. This task is
performed by the first processing block in Figure 5.
IV. EXPERIMENTAL SETUP
In order to assess the performance of the proposed detector,
we tested it in various scenarios and we evaluated it in
terms of accuracy. In this section we describe such scenarios
and the relevant experiments. Furthermore, we introduce
the datasets used for the testing, and we provide additional
details the relevant training procedure.
A. DATASETS
Three publicly available datasets have been used in the ex-
periments:
•The Uncompressed Colour Image Database (UCID)
[41], built with the purpose of providing a standard
set for performance assessments in image retrieval and
compression. The dataset consists of 1,338 uncom-
pressed color images, with fixed sizes of 512 ×384 or
384 ×512 pixels in uncompressed TIFF format.
•The Dresden Image Database (DRESDEN) [42], orig-
inally created for evaluation of forensic techniques re-
lated to camera-based information. From their public
web-interface, we selected the complete set of RAW
images (1,189 uncompressed images, all with fixed size
of 3008 ×2000 pixels).
•The Raw Images Dataset (RAISE) [43], consisting of
8156 raw images with resolutions ranging from 3,008×
2,000 to 4,928 ×3,264 pixels. Authors provide also
smaller subsets, among which we selected the one con-
taining 1,000 images (RAISE-1k).
The three datasets were selected to diversify the range of
resolutions in the experimental tests. In order to evaluate our
proposed schema, all the images were converted to grayscale
with a depth of 8 bits.
B. TESTING SCENARIOS
The proposed detector has been tested in various practical
scenarios:
•The first set of tests were performed in order to under-
stand the impact of the parameters as well as to select a
proper classifier. Based on the setups found on this set of
tests, we kept the best setups (in terms of performance)
for the next tests.
•The second set of tests refers to the detection of mor-
phological filtering. In this case, we want to establish
the accuracy of the detector in discriminating filtered
vs. pristine (unfiltered) images, in classifying the type
4VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2965745, IEEE Access
Boato et al.: Morphological Filter Detector for Image Forensics Applications
Unprocessed, QF = 95
I1 = JPEGQF=95(I0)
Dilation, Kernel index = 35, QF = 95
I2 = JPEGQF=95(I0 K35)
Opened, Kernel index = 35
I' 1 = I1 ° K35
Opened, Kernel index = 35
I' 2 = I2 ° K35
Difference image (shown in log)
Idiff
1 = |I'1 - I 1|
Difference image (shown in log)
Idiff
2 = |I'2 - I 2|
Unprocessed, QF = 95
I1 = JPEGQF=95(I0)
Dilation, Kernel index = 35, QF = 95
I2 = JPEGQF=95(I0 K35)
Opened, Kernel index = 35
I' 1 = I1 ° K35
Opened, Kernel index = 35
I' 2 = I2 ° K35
Difference image (shown in log)
Idiff
1 = |I'1 - I 1|
Difference image (shown in log)
Idiff
2 = |I'2 - I 2|
Unprocessed, QF = 95
I1 = JPEGQF=95(I0)
Dilation, Kernel index = 35, QF = 95
I2 = JPEGQF=95(I0 K35)
Opened, Kernel index = 35
I' 1 = I1 ° K35
Opened, Kernel index = 35
I' 2 = I2 ° K35
Difference image (shown in log)
Idiff
1 = |I'1 - I 1|
Difference image (shown in log)
Idiff
2 = |I'2 - I 2|
Unprocessed, QF = 95
I1 = JPEGQF=95(I0)
Dilation, Kernel index = 35, QF = 95
I2 = JPEGQF=95(I0 K35)
Opened, Kernel index = 35
I' 1 = I1 ° K35
Opened, Kernel index = 35
I' 2 = I2 ° K35
Difference image (shown in log)
Idiff
1 = |I'1 - I 1|
Difference image (shown in log)
Idiff
2 = |I'2 - I 2|
Unprocessed, QF = 95
I1 = JPEGQF=95(I0)
Dilation, Kernel index = 35, QF = 95
I2 = JPEGQF=95(I0 K35)
Opened, Kernel index = 35
I' 1 = I1 ° K35
Opened, Kernel index = 35
I' 2 = I2 ° K35
Difference image (shown in log)
Idiff
1 = |I'1 - I 1|
Difference image (shown in log)
Idiff
2 = |I'2 - I 2|
Unprocessed, QF = 95
I1 = JPEGQF=95(I0)
Dilation, Kernel index = 35, QF = 95
I2 = JPEGQF=95(I0 K35)
Opened, Kernel index = 35
I' 1 = I1 ° K35
Opened, Kernel index = 35
I' 2 = I2 ° K35
Difference image (shown in log)
Idiff
1 = |I'1 - I 1|
Difference image (shown in log)
Idiff
2 = |I'2 - I 2|
Unprocessed, QF = 95
I1 = JPEGQF=95(I0)
Dilation, Kernel index = 35, QF = 95
I2 = JPEGQF=95(I0 K35)
Opened, Kernel index = 35
I' 1 = I1 ° K35
Opened, Kernel index = 35
I' 2 = I2 ° K35
Difference image (shown in log)
Idiff
1 = |I'1 - I 1|
Difference image (shown in log)
Idiff
2 = |I'2 - I 2|
Unprocessed, QF = 95
I1 = JPEGQF=95(I0)
Dilation, Kernel index = 35, QF = 95
I2 = JPEGQF=95(I0 K35)
Opened, Kernel index = 35
I' 1 = I1 ° K35
Opened, Kernel index = 35
I' 2 = I2 ° K35
Difference image (shown in log)
Idiff
1 = |I'1 - I 1|
Difference image (shown in log)
Idiff
2 = |I'2 - I 2|
Unprocessed, QF = 95
I1 = JPEGQF=95(I0)
Dilation, Kernel index = 35, QF = 95
I2 = JPEGQF=95(I0 K35)
Opened, Kernel index = 35
I' 1 = I1 ° K35
Opened, Kernel index = 35
I' 2 = I2 ° K35
Difference image (shown in log)
Idiff
1 = |I'1 - I 1|
Difference image (shown in log)
Idiff
2 = |I'2 - I 2|
Unprocessed, QF = 95
I1 = JPEGQF=95(I0)
Dilation, Kernel index = 35, QF = 95
I2 = JPEGQF=95(I0 K35)
Opened, Kernel index = 35
I' 1 = I1 ° K35
Opened, Kernel index = 35
I' 2 = I2 ° K35
Difference image (shown in log)
Idiff
1 = |I'1 - I 1|
Difference image (shown in log)
Idiff
2 = |I'2 - I 2|
Unprocessed, QF = 95
I1 = JPEGQF=95(I0)
Dilation, Kernel index = 35, QF = 95
I2 = JPEGQF=95(I0 K35)
Opened, Kernel index = 35
I' 1 = I1 ° K35
Opened, Kernel index = 35
I' 2 = I2 ° K35
Difference image (shown in log)
Idiff
1 = |I'1 - I 1|
Difference image (shown in log)
Idiff
2 = |I'2 - I 2|
Unprocessed, QF = 95
I1 = JPEGQF=95(I0)
Dilation, Kernel index = 35, QF = 95
I2 = JPEGQF=95(I0 K35)
Opened, Kernel index = 35
I' 1 = I1 ° K35
Opened, Kernel index = 35
I' 2 = I2 ° K35
Difference image (shown in log)
Idiff
1 = |I'1 - I 1|
Difference image (shown in log)
Idiff
2 = |I'2 - I 2|
Unprocessed, QF = 95
I1 = JPEGQF=95(I0)
Dilation, Kernel index = 35, QF = 95
I2 = JPEGQF=95(I0 K35)
Opened, Kernel index = 35
I' 1 = I1 ° K35
Opened, Kernel index = 35
I' 2 = I2 ° K35
Difference image (shown in log)
Idiff
1 = |I'1 - I 1|
Difference image (shown in log)
Idiff
2 = |I'2 - I 2|
Unprocessed, QF = 95
I1 = JPEGQF=95(I0)
Dilation, Kernel index = 35, QF = 95
I2 = JPEGQF=95(I0 K35)
Opened, Kernel index = 35
I' 1 = I1 ° K35
Opened, Kernel index = 35
I' 2 = I2 ° K35
Difference image (shown in log)
Idiff
1 = |I'1 - I 1|
Difference image (shown in log)
Idiff
2 = |I'2 - I 2|
Ie=JPEGQF =95 (I0 K35)
<latexit sha1_base64="EdhPX99H0OLJ3lxXVPjq7scSL+g=">AAACD3icbVC7SgNBFJ2NrxhfUUubwaDEJuxqgloEguIj2iRgHpCEZXYySYbMzi4zs0JY9g9s/BUbC0Vsbe38GyfJFho9cOFwzr3ce4/jMyqVaX4Zibn5hcWl5HJqZXVtfSO9uVWXXiAwqWGPeaLpIEkY5aSmqGKk6QuCXIeRhjM8H/uNeyIk9fidGvmk46I+pz2KkdKSnd4v2wQW4U3l4soOq5fF00KULdsmbHsu5YGEt3Z4VIgO7HTGzJkTwL/EikkGxKjY6c9218OBS7jCDEnZskxfdUIkFMWMRKl2IImP8BD1SUtTjlwiO+HknwjuaaULe57QxRWcqD8nQuRKOXId3ekiNZCz3lj8z2sFqnfSCSn3A0U4ni7qBQwqD47DgV0qCFZspAnCgupbIR4ggbDSEaZ0CNbsy39J/TBn5XP5aj5TOovjSIIdsAuywALHoASuQQXUAAYP4Am8gFfj0Xg23oz3aWvCiGe2wS8YH9+UsJnh</latexit>
Id=JPEGQF =95 (I0K35)
<latexit sha1_base64="4BQ8wyEreLZhUL0QA4LUZ7bttDY=">AAACDnicbVC7SgNBFJ31GeMramkzGAKxCbuaoBaBoPiINgmYByTLMjs7SYbM7iwzs0JY8gU2/oqNhSK21nb+jZNHoYkHLhzOuZd773FDRqUyzW9jYXFpeWU1sZZc39jc2k7t7NYljwQmNcwZF00XScJoQGqKKkaaoSDIdxlpuP2Lkd94IEJSHtyrQUhsH3UD2qEYKS05qUzZ8WAR3lYur524elU8KwyzZceEbR6ySMI7Jz4uDA+dVNrMmWPAeWJNSRpMUXFSX22P48gngcIMSdmyzFDZMRKKYkaGyXYkSYhwH3VJS9MA+UTa8fidIcxoxYMdLnQFCo7V3xMx8qUc+K7u9JHqyVlvJP7ntSLVObVjGoSRIgGeLOpEDCoOR9lAjwqCFRtogrCg+laIe0ggrHSCSR2CNfvyPKkf5ax8Ll/Np0vn0zgSYB8cgCywwAkogRtQATWAwSN4Bq/gzXgyXox342PSumBMZ/bAHxifP8P5mW4=</latexit>
Compressed Unprocessed
0 5 10 15 20 25 30 35 40 45 50
100
101
102
103
104
105
106
0 5 10 15 20 25 30 35 40 45 50
100
101
102
103
104
105
106
Iu=JPEGQF =95 (I0)
<latexit sha1_base64="MMO4fgdsMpR0+N+of/mqTEBl0oY=">AAAB/nicbVDLSsNAFJ34rPUVFVduBotQNyWRirooFMVHXbVgH9CGMJlO2qGTSZiZCCUU/BU3LhRx63e482+ctllo64ELh3Pu5d57vIhRqSzr21hYXFpeWc2sZdc3Nre2zZ3dhgxjgUkdhywULQ9JwigndUUVI61IEBR4jDS9wdXYbz4SIWnIH9QwIk6Aepz6FCOlJdfcr7hx6b56fesmtZvSxekoX3GtY9fMWQVrAjhP7JTkQIqqa351uiGOA8IVZkjKtm1FykmQUBQzMsp2YkkihAeoR9qachQQ6SST80fwSCtd6IdCF1dwov6eSFAg5TDwdGeAVF/OemPxP68dK//cSSiPYkU4ni7yYwZVCMdZwC4VBCs21ARhQfWtEPeRQFjpxLI6BHv25XnSOCnYxUKxVsyVL9M4MuAAHII8sMEZKIM7UAV1gEECnsEreDOejBfj3fiYti4Y6cwe+APj8wdNypPG</latexit>
I•
u=Iu•K35
<latexit sha1_base64="CWAoqlS7LQQIMUngAzNFy2BAeP8=">AAACCXicbVDLSsNAFJ3UV62vqEs3g0VwVRKt6EYourG4qWAf0MYwmU7aoZNJmIdQQrZu/BU3LhRx6x+482+ctllo9cCFM+fcy9x7goRRqRznyyosLC4trxRXS2vrG5tb9vZOS8ZaYNLEMYtFJ0CSMMpJU1HFSCcRBEUBI+1gdDnx2/dESBrzWzVOiBehAachxUgZybdh/S7tBZoxojJfn9d9DfMnvPbT45PMt8tOxZkC/iVuTsogR8O3P3v9GOuIcIUZkrLrOonyUiQUxYxkpZ6WJEF4hAakayhHEZFeOr0kgwdG6cMwFqa4glP150SKIinHUWA6I6SGct6biP95Xa3CMy+lPNGKcDz7KNQMqhhOYoF9KghWbGwIwoKaXSEeIoGwMuGVTAju/Ml/Seuo4lYr1ZtquXaRx1EEe2AfHAIXnIIauAIN0AQYPIAn8AJerUfr2Xqz3metBSuf2QW/YH18A+UrmdE=</latexit>
I•
e=Ie•K35
<latexit sha1_base64="L/hGoWvhTSkmiPb43dNr/HJRl/o=">AAACCXicbVDLSsNAFJ3UV62vqEs3g0VwVRKt6EYourG4qWAf0MYwmd60QycPZiZCCdm68VfcuFDErX/gzr9x2mah1QMXzpxzL3Pv8WLOpLKsL6OwsLi0vFJcLa2tb2xumds7LRklgkKTRjwSHY9I4CyEpmKKQycWQAKPQ9sbXU789j0IyaLwVo1jcAIyCJnPKFFack1cv0t7XsI5qMyF87oLOH/iazc9Pslcs2xVrCnwX2LnpIxyNFzzs9ePaBJAqCgnUnZtK1ZOSoRilENW6iUSYkJHZABdTUMSgHTS6SUZPtBKH/uR0BUqPFV/TqQkkHIceLozIGoo572J+J/XTZR/5qQsjBMFIZ195CccqwhPYsF9JoAqPtaEUMH0rpgOiSBU6fBKOgR7/uS/pHVUsauV6k21XLvI4yiiPbSPDpGNTlENXaEGaiKKHtATekGvxqPxbLwZ77PWgpHP7KJfMD6+AbKLmbE=</latexit>
I
d=IdK35
<latexit sha1_base64="vIiGQuDtB7zem+r+Tp0R/lV4UqY=">AAACBXicbVDLSsNAFJ3UV62vqEtdDBbBVUm0ohuh6MbipoJ9QBvDZDJth04mYWYilJCNG3/FjQtF3PoP7vwbp2kW2nrgwplz7mXuPV7EqFSW9W0UFhaXlleKq6W19Y3NLXN7pyXDWGDSxCELRcdDkjDKSVNRxUgnEgQFHiNtb3Q18dsPREga8js1jogToAGnfYqR0pJr7tfvkx6mAqeuf1F3fZg94I2bnJymrlm2KlYGOE/snJRBjoZrfvX8EMcB4QozJGXXtiLlJEgoihlJS71YkgjhERqQrqYcBUQ6SXZFCg+14sN+KHRxBTP190SCAinHgac7A6SGctabiP953Vj1z52E8ihWhOPpR/2YQRXCSSTQp4JgxcaaICyo3hXiIRIIKx1cSYdgz548T1rHFbtaqd5Wy7XLPI4i2AMH4AjY4AzUwDVogCbA4BE8g1fwZjwZL8a78TFtLRj5zC74A+PzByz5l7k=</latexit>
I
u=IuK35
<latexit sha1_base64="VBgH2BOo9krsTEfEySoNPx4xLwc=">AAACBXicbVDLSsNAFJ3UV62vqEtdDBbBVUm0ohuh6MbipoJ9QBvDZDpph04mYWYilJCNG3/FjQtF3PoP7vwbp2kW2nrgwplz7mXuPV7EqFSW9W0UFhaXlleKq6W19Y3NLXN7pyXDWGDSxCELRcdDkjDKSVNRxUgnEgQFHiNtb3Q18dsPREga8js1jogToAGnPsVIack19+v3SQ9TgVM3vqi7Mcwe8MZNTk5T1yxbFSsDnCd2TsogR8M1v3r9EMcB4QozJGXXtiLlJEgoihlJS71YkgjhERqQrqYcBUQ6SXZFCg+10od+KHRxBTP190SCAinHgac7A6SGctabiP953Vj5505CeRQrwvH0Iz9mUIVwEgnsU0GwYmNNEBZU7wrxEAmElQ6upEOwZ0+eJ63jil2tVG+r5dplHkcR7IEDcARscAZq4Bo0QBNg8AiewSt4M56MF+Pd+Ji2Fox8Zhf8gfH5A2J/l9s=</latexit>
Idiff
u,=|I
uIu|
<latexit sha1_base64="FWZg/Ny3Hww1fx/3G9meP/A+omA=">AAACEHicbZC7SgNBFIZn4y3G26qlzWAQLTTsSkAbIWhjugjmAsm6zM7OJkNmL8xFCJt9BBtfxcZCEVtLO9/GSbKFRn8Y+PjPOZw5v5cwKqRlfRmFhcWl5ZXiamltfWNzy9zeaYlYcUyaOGYx73hIEEYj0pRUMtJJOEGhx0jbG15N6u17wgWNo1s5SogTon5EA4qR1JZrHtbvUp8GQeam6hj2MOU4uxhrc4augiew7qqxa5atijUV/At2DmWQq+Ganz0/xiokkcQMCdG1rUQ6KeKSYkayUk8JkiA8RH3S1RihkAgnnR6UwQPt+DCIuX6RhFP350SKQiFGoac7QyQHYr42Mf+rdZUMzp2URomSJMKzRYFiUMZwkg70KSdYspEGhDnVf4V4gDjCUmdY0iHY8yf/hdZpxa5WqjfVcu0yj6MI9sA+OAI2OAM1cA0aoAkweABP4AW8Go/Gs/FmvM9aC0Y+swt+yfj4BjtinLQ=</latexit>
Idiff
d,=|I
dId|
<latexit sha1_base64="8xakbng8RMT5XL8aItEkj9CG+Jc=">AAACEHicbZDLSsNAFIYn9VbrLerSzWARXWhJpKAboejG7irYC7QxTCaTduhkEmYmQknzCG58FTcuFHHr0p1v47TNQqs/DHz85xzOnN+LGZXKsr6MwsLi0vJKcbW0tr6xuWVu77RklAhMmjhikeh4SBJGOWkqqhjpxIKg0GOk7Q2vJvX2PRGSRvxWjWLihKjPaUAxUtpyzcP6XerTIMjc1D+GPUwFzi7G2pyh68MTWHf9sWuWrYo1FfwLdg5lkKvhmp89P8JJSLjCDEnZta1YOSkSimJGslIvkSRGeIj6pKuRo5BIJ50elMED7fgwiIR+XMGp+3MiRaGUo9DTnSFSAzlfm5j/1bqJCs6dlPI4UYTj2aIgYVBFcJIO9KkgWLGRBoQF1X+FeIAEwkpnWNIh2PMn/4XWacWuVqo31XLtMo+jCPbAPjgCNjgDNXANGqAJMHgAT+AFvBqPxrPxZrzPWgtGPrMLfsn4+Abro5yB</latexit>
Idiff
u,•=|I•
uIu|
<latexit sha1_base64="3KkJOeJuW8iYHGJ0xS15YjM4dlQ=">AAACFHicbVDLSsNAFJ3UV62vqEs3g0UQ1JJIQTdC0Y3dVbAPaGOYTCft0MmDeQglzUe48VfcuFDErQt3/o3TNoK2Hrhw5px7mXuPFzMqpGV9GbmFxaXllfxqYW19Y3PL3N5piEhxTOo4YhFveUgQRkNSl1Qy0oo5QYHHSNMbXI395j3hgkbhrRzGxAlQL6Q+xUhqyTWPqndJl/p+6ibqGHY8xRiR6cVIyz8PV8ETWHXVyDWLVsmaAM4TOyNFkKHmmp+dboRVQEKJGRKibVuxdBLEJcWMpIWOEiRGeIB6pK1piAIinGRyVAoPtNKFfsR1hRJO1N8TCQqEGAae7gyQ7ItZbyz+57WV9M+dhIaxkiTE0498xaCM4Dgh2KWcYMmGmiDMqd4V4j7iCEudY0GHYM+ePE8apyW7XCrflIuVyyyOPNgD++AQ2OAMVMA1qIE6wOABPIEX8Go8Gs/Gm/E+bc0Z2cwu+APj4xvP4J6q</latexit>
Idiff
e,•=|I•
eIe|
<latexit sha1_base64="3kQqVyKhc7m2gRd2r6EPf91UMMs=">AAACFHicbVDLSgMxFM3UV62vqks3wSIIapmRgm6Eohu7q2Af0I5DJr3ThmYeJBmhTPsRbvwVNy4UcevCnX9j2o6grQcunJxzL7n3uBFnUpnml5FZWFxaXsmu5tbWNza38ts7dRnGgkKNhjwUTZdI4CyAmmKKQzMSQHyXQ8PtX439xj0IycLgVg0isH3SDZjHKFFacvJHlbukwzxv5CRwjNtuzDmo0cVQyz8PB/AJrjgwdPIFs2hOgOeJlZICSlF18p/tTkhjHwJFOZGyZZmRshMiFKMcRrl2LCEitE+60NI0ID5IO5kcNcIHWulgLxS6AoUn6u+JhPhSDnxXd/pE9eSsNxb/81qx8s7thAVRrCCg04+8mGMV4nFCuMMEUMUHmhAqmN4V0x4RhCqdY06HYM2ePE/qp0WrVCzdlArlyzSOLNpD++gQWegMldE1qqIaougBPaEX9Go8Gs/Gm/E+bc0Y6cwu+gPj4xuEoJ56</latexit>
Iu
<latexit sha1_base64="TgZFfWUJkNfPSRzblRFRKnRa1UA=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0mkoMeiF71VtB/QhrLZTtqlm03Y3Qgl9Cd48aCIV3+RN/+N2zYHbX0w8Hhvhpl5QSK4Nq777RTW1jc2t4rbpZ3dvf2D8uFRS8epYthksYhVJ6AaBZfYNNwI7CQKaRQIbAfjm5nffkKleSwfzSRBP6JDyUPOqLHSw10/7ZcrbtWdg6wSLycVyNHol796g5ilEUrDBNW667mJ8TOqDGcCp6VeqjGhbEyH2LVU0gi1n81PnZIzqwxIGCtb0pC5+nsio5HWkyiwnRE1I73szcT/vG5qwis/4zJJDUq2WBSmgpiYzP4mA66QGTGxhDLF7a2EjaiizNh0SjYEb/nlVdK6qHq1au2+Vqlf53EU4QRO4Rw8uIQ63EIDmsBgCM/wCm+OcF6cd+dj0Vpw8plj+APn8wctbI28</latexit>
Ie
<latexit sha1_base64="niiPzwIaFYz6MPyPgdKABhJz9ik=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0mkoMeiF71VtB/QhrLZTtqlm03Y3Qgl9Cd48aCIV3+RN/+N2zYHbX0w8Hhvhpl5QSK4Nq777RTW1jc2t4rbpZ3dvf2D8uFRS8epYthksYhVJ6AaBZfYNNwI7CQKaRQIbAfjm5nffkKleSwfzSRBP6JDyUPOqLHSw10f++WKW3XnIKvEy0kFcjT65a/eIGZphNIwQbXuem5i/Iwqw5nAaamXakwoG9Mhdi2VNELtZ/NTp+TMKgMSxsqWNGSu/p7IaKT1JApsZ0TNSC97M/E/r5ua8MrPuExSg5ItFoWpICYms7/JgCtkRkwsoUxxeythI6ooMzadkg3BW355lbQuql6tWruvVerXeRxFOIFTOAcPLqEOt9CAJjAYwjO8wpsjnBfn3flYtBacfOYY/sD5/AEVLI2s</latexit>
Id
<latexit sha1_base64="KfLuyoMIWrD8vL6MhjUlHMFwMFg=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0mkoMeiF71VtB/QhrLZbNqlm03YnQil9Cd48aCIV3+RN/+N2zYHbX0w8Hhvhpl5QSqFQdf9dgpr6xubW8Xt0s7u3v5B+fCoZZJMM95kiUx0J6CGS6F4EwVK3kk1p3EgeTsY3cz89hPXRiTqEccp92M6UCISjKKVHu76Yb9ccavuHGSVeDmpQI5Gv/zVCxOWxVwhk9SYruem6E+oRsEkn5Z6meEpZSM64F1LFY258SfzU6fkzCohiRJtSyGZq78nJjQ2ZhwHtjOmODTL3kz8z+tmGF35E6HSDLlii0VRJgkmZPY3CYXmDOXYEsq0sLcSNqSaMrTplGwI3vLLq6R1UfVq1dp9rVK/zuMowgmcwjl4cAl1uIUGNIHBAJ7hFd4c6bw4787HorXg5DPH8AfO5w8TqI2r</latexit>
Compressed Eroded Compressed Dilated
Top-left corner of size 1000 x 1000
Histograms of Differences
Difference Images (shown in log)
Closed Images
Unprocessed, QF = 95
I1 = JPEGQF=95(I0)
Dilation, Kernel index = 35, QF = 95
I2 = JPEGQF=95(I0 K35)
Opened, Kernel index = 35
I' 1 = I1 ° K35
Opened, Kernel index = 35
I' 2 = I2 ° K35
Difference image (shown in log)
Idiff
1 = |I'1 - I 1|
Difference image (shown in log)
Idiff
2 = |I'2 - I 2|
Iu
<latexit sha1_base64="TgZFfWUJkNfPSRzblRFRKnRa1UA=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0mkoMeiF71VtB/QhrLZTtqlm03Y3Qgl9Cd48aCIV3+RN/+N2zYHbX0w8Hhvhpl5QSK4Nq777RTW1jc2t4rbpZ3dvf2D8uFRS8epYthksYhVJ6AaBZfYNNwI7CQKaRQIbAfjm5nffkKleSwfzSRBP6JDyUPOqLHSw10/7ZcrbtWdg6wSLycVyNHol796g5ilEUrDBNW667mJ8TOqDGcCp6VeqjGhbEyH2LVU0gi1n81PnZIzqwxIGCtb0pC5+nsio5HWkyiwnRE1I73szcT/vG5qwis/4zJJDUq2WBSmgpiYzP4mA66QGTGxhDLF7a2EjaiizNh0SjYEb/nlVdK6qHq1au2+Vqlf53EU4QRO4Rw8uIQ63EIDmsBgCM/wCm+OcF6cd+dj0Vpw8plj+APn8wctbI28</latexit>
Opened Images
FIGURE 4: Histograms of differences before and after application of the close and open operators to an unprocessed image
(grayscale version of ‘r3ba1827ft.TIF’ from RAISE-1k) and its eroded and dilated versions, respectively. Both images have
been JPEG compressed with QF = 95. Kernel mask 35 and the analysed windows size of 1000 ×1000 are used. The chart
shows the first 50 bins in logarithmic scale.
VOLUME 4, 2016 5

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2965745, IEEE Access
Boato et al.: Morphological Filter Detector for Image Forensics Applications
Erosion Dilation
ErosionDilation
Structuring
Elements
Classifier
Input
Image
Scan next element ...
Dilation
Detected
Erosion
Detected
1
1
0
0
Classifier
HIST
HIST
Block
Selection
Block
Selection
+
+
-
-
FIGURE 5: Proposed detection scheme for attacked grayscale images. After opening/closing, we select image regions that
contain significant textures or edges and there we calculate the histogram of the differences between the input and output
images. Such histograms are fed into a statistical classifier to perform the decision (details in Section III.B).
of operator (erosion vs. dilation), and finally in deter-
mining the exact structuring element used for filtering,
both in uncompressed and JPEG compressed images
with various quality factors (from QF = 100 indicating
highest quality images, down to QF = 70 which still
corresponds to compressed images of high quality).
•The third set of tests concerns robustness against noise.
The test image is contaminated with random noise and
then compressed, and we evaluate the accuracy of the
detector in determining the presence and type of filtering
and the structuring element used.
•Finally, we want to determine the capability of our
approach in distinguishing between morphological fil-
ters and other filters that produce similar results. In
particular, we considered Gaussian lowpass and median
filtering. In both cases, the filtered images are com-
pressed and passed to the detector to reveal possible
false alarms.
It is worth noting that, due to the properties analysed in
Section II, in the presence of a cascade of different basic op-
erators (erosion and dilation), the detector will reveal the last
operator applied. Accordingly, when processing an opened
or a closed image, the detector will reveal the last dilation
or erosion, respectively. Furthermore, in the presence of a
cascade of the same basic operator, the detector will reveal
a single erosion or dilation with the composed structuring
element. Therefore, in the experimental section we will just
consider erosion and dilation detection, even if the image
may have been potentially processed with more complex
combinations of filters.
V. EXPERIMENTAL RESULTS AND DISCUSSION
In this section, all experimental tests carried out are de-
scribed. All tests are performed using Python 3.6 and libSVM
on a standard machine (Macbook Pro 2016 2,3GHz 4-kernel
Intel Core i5, 8GB ram).
A. THE SENSIBILITY AND THE CHOICE OF THE
CLASSIFIERS
In order to understand the impact of the parameters as well
as to select a proper classifier, we ran some experiments on
a subset of images. After that, we empirically determined
the values for these parameters. Table 1 summarizes these
parameters, their meaning and how their values were decided.
The classifier receives in input the histogram of the dif-
ferences between the input image and the relevant opened
(closed) image and returns a binary decision. An important
aspect is the training of the classifier. In fact, there is a
clear dependency of the image statistics on the level of
compression applied, which is reflected on the characteristics
of the histogram. As an example, Figure 6 (a) shows the
same situation of Figure 4 at a lower quality factor. It can
be observed that the two distributions are still well separable,
but the histogram related to the filtered image shows a longer
tail, due to the larger artifacts introduced by the compression.
Accordingly, we decided to train a set of classifiers for
varying JPEG quality factors, from 100 to 70. During the
test phase, we tested with different classifiers: k-NN, deci-
sion trees, naive Bayes classifier, and SVM, and empirically
decided to use SVM with the radial basis function (RBF)
kernel.
We also note that given the appropriate quality factor,
the histograms of differences show a peculiar behavior only
if the detector parameters match the input filter in both
the operator (erosion or dilation) and the structural element
(kernel), whereas such behavior is never found in all the other
combinations. It is important to notice that the application of
the detector to an unfiltered image or to an image filtered with
a different combination of operator and/or kernel produces
similar results, i.e., only the matching detector responds to
the filtered image. As an example, in Figure 6 (b) we show
the histograms deriving from the application of an erosion
detector to an eroded and a dilated image, both with the same
kernel. It can be observed that the dilated image responds
6VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2965745, IEEE Access
Boato et al.: Morphological Filter Detector for Image Forensics Applications
TABLE 1: Parameters selection.
Parameter/ Classifier Meaning Range Empirical Value/Selection
Image resolution The size of the analyzed region of interest 384 ×512 to 4928 ×3264 Good results for resolution from 1000 ×
1000.
Block Selection Threshold
αth and Block Size s
Erosion and dilation have a negligible effect
on flat areas, thus these two parameters allow
the method to effectively work on non-flat
blocks only.
αth ≥0,α= 0 means all pixels
in the block has the same values.
s∈ {3×3,5×5, ...}
αth = 0.15 and s= 3 ×3.
nNumber of bins of the analyzed histograms. 0≤n≤256 Even if for n > 30 we already achieve 90%
accuracy, we set nat the highest value, i.e.,
considering all bins.
The classifiers SVM, KNN, Linear Regression,
and others
SVM with RBF kernel and grid search pro-
vides the best performance.
(a) (b)
(c) (d)
Compressed unprocessed image
Compressed eroded image
Compressed dilated image
Compressed eroded image
Compressed eroded image (kernel K*)
Compressed eroded image (kernel K)
Compressed eroded image
Compressed eroded image
(same kernel)
(different kernels)
Histograms of Differences (QF = 80)
Histograms of Differences (QF = 95)
Histograms of Differences (QF = 95)
Histograms of Differences (QF = 95)
Differences
Differences
Differences
Differences
Number of Pixels
Number of Pixels
Number of Pixels
Number of Pixels
FIGURE 6: Analysis of different cases in erosion detection. (a) Histograms resulting from original image and its dilated
version, after compression with QF = 80; (b) Histograms resulting from dilated and eroded images produced with the same
kernel, after compression at QF = 95; (c) Histograms resulting from eroded images produced with unrelated kernels, after
compression at QF = 95; and (d) Histograms resulting from eroded images produced with kernel Kand its dilated version
K∗, after compression at QF = 95
.
VOLUME 4, 2016 7

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2965745, IEEE Access
Boato et al.: Morphological Filter Detector for Image Forensics Applications
TABLE 2: Morphological filtering detection results on all three datasets using SVM with RBF kernel, 10-fold cross validation.
The numbers are in percentage and each of them is the average performance (for distinguishing between pristine versus filtered
images) over all 36 kernels. Detailed results are reported in Table 3.
Dataset Operator QF = 100 95 90 85 80 75 70
UCID Erosion 100 90.86 82.62 79.63 77.45 67.41 64.42
Dilation 100 90.16 82.86 79.95 76.80 68.77 64.94
Dresden Erosion 100 97.56 96.94 95.20 88.73 79.85 78.34
Dilation 100 97.93 96.82 96.11 88.74 81.08 75.08
Raise Erosion 100 98.66 96.25 94.82 89.00 80.42 78.55
Dilation 100 97.98 97.43 93.72 91.02 80.49 78.53
similarly to the uncompressed image in Figure 6 (a). Analo-
gously, in Figure 6 (c), we compare the histograms deriving
from the erosion detector applied to two eroded images, one
with the same kernel and the other with a different one.
Also in this case, the image filtered with a different kernel
responds as unfiltered. Finally, in Figure 6 (d) we show the
case of application of the erosion detector to two images
filtered with kernels belonging to the same group (i.e., one
can be obtained from the other by dilation). As expected from
Theorem 2, the two histograms are almost overlapped, since
both show the statistical properties of a filtered image.
For some other choices, for example the block size or αth,
all values are empirically selected and are reported in Table 1.
B. DISTINGUISHING BETWEEN PRISTINE VERSUS
MORPHOLOGICAL FILTERED IMAGES
1) Experimental Results on Raw Images
We ran a test on uncompressed images in order to confirm
the deterministic nature of the proposed approach on raw
images. Since the detection strategy in this case is the very
same proposed in [39], we followed their general approach
for deriving results relative to raw grayscale images. Each
image has been processed with erosion and dilation operators
considering all the 36 kernels. All images, along with their
unprocessed versions, were fed to both dilation and erosion
detectors. For each image, all 36 kernels are tested, returning
either the largest kernel with perfect match between input and
output or no detection if all kernels fail. All three datasets
result in 100% accuracy in discriminating the presence and
the type of morphological filter.
2) Experimental Results on JPEG Compressed Images
In order to test the approach on JPEG compressed images,
we considered a set of seven different quality factors with
QF ∈ {100,95,90,85,80,75,70}, with 14 (7 ×2) binary
classifiers, respectively, using a Gaussian kernel with grid-
search for the parameters. We apply k-Fold validation with
k= 10. Shown in Table 2 are the average results (over all
kernels) of the proposed method on three datasets at the seven
quality factors. The detailed results for all kernels and both
morphological operators on the three datasets are reported in
Table 3. According to the results, we can observe that the
proposed method can provide high detection performance on
low compression level (i.e., quality factor QF ≥80) with
accuracy equal or above 76.8% on all datasets. At stronger
0
10
20
30
40
50
60
70
80
90
100
Full
3000x3000
2000x2000
1000x1000
500x500
300x300
FIGURE 7: Precision (computed as percentage) on erosion
detection under different resolutions. Dataset: Raise, QF =
90. All 36 kernels are used, results are shown in the error
bars.
compression levels, the performance reduces significantly.
The resolution also plays an important role as it can be
grasped from the results. Indeed, UCID has lower perfor-
mance with respect to Dresden and Raise. It is also interesting
to notice that there is little difference between Dresden and
Raise. To have a deeper understanding of the impact of the
image resolution, we ran another test on different image
resolutions: the proposed method achieves good performance
(over 90%) for resolution from about 1,000 ×1,000. Shown
in Figure 7 are results over 36 kernels on erosion detection
for RAISE-1k images.
C. DETECTOR ROBUSTNESS ANALYSIS
The second test in this set is to analyse the impact of different
kernels. We applied all kernel masks and then tried to detect
them. Shown in Figure 8 is the confusion matrix of the
erosion detection (similar results were obtained on dilation
detection). According to the results, the detector was able
to recognize small size kernels (kernels from 1 to 14), or
larger kernels in special shape (kernels 20 to 25), but poorly
classified the others. This is understandable as a consequence
of Theorem 2.
The third set of tests concerns analysis of robustness
against noise addition. In the first test, morphological filter-
8VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2965745, IEEE Access
Boato et al.: Morphological Filter Detector for Image Forensics Applications
TABLE 3: Detailed results on erosion and dilation detection on UCID, Dresden and RAISE-1k datasets. K indicates the kernel
mask used (from 1 to 36). Average results are summarised in Table 2.
QF K = 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
95 90.42 90.89 90.34 89.61 90.15 90.56 91.2 92.2 92.69 92.89 93.09 93.48 94.47 93.59 94.4 88.41 88.49 88.78
UCID 90 83.42 82.73 82.57 83.06 83.14 82.49 82.67 82.62 83.59 84.08 83.41 83.63 83.62 84.32 84.02 82.22 83.18 83.72
Erosion 85 79.06 79.7 80.69 80.74 80.91 81.42 81.03 80.17 80 80.65 79.88 80.87 80.73 80.25 79.3 80.45 81.17 81.99
80 75.83 76.23 77.19 76.99 77.62 78.42 78.66 78.19 79.1 79.67 80.26 80.08 79.93 79.59 80.32 76.24 76.4 76.76
75 64.25 65.2 66.06 65.31 65.09 65.55 66.22 65.24 65.9 65.78 65.59 64.86 64.35 64.49 65.07 69.13 69.07 69.08
70 63.34 64.23 64.91 63.92 62.95 63.56 64.53 65.45 64.87 64.5 65.28 65.41 64.61 65.6 64.72 62.8 63.66 63.69
100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
95 89.24 88.4 88.1 88.11 89.09 89.99 89.28 89.9 90.46 91.03 90.76 91.28 91.44 90.49 91.33 90.05 90.66 90.5
UCID 90 83.18 82.29 81.66 82.34 82.67 82.85 82.35 83.04 83.42 82.55 83.13 83.21 83.85 83.33 84.22 82.22 82.24 82.21
Dilation 85 80.05 81.01 81.54 81.51 82.35 82.43 83.15 82.9 82.3 81.76 80.9 81.69 81.01 81.7 82.61 81.15 81.36 81.39
80 75.29 76.16 75.77 75.33 74.89 75.42 75.81 75.06 75.99 75 75 75.84 76.02 76.56 75.65 77.25 77.96 78.31
75 65.35 65.08 65.49 65.17 65.95 65.94 65.53 66.12 66.64 67.59 67.69 67.9 68.6 68.58 69.13 70.23 70.63 70.22
70 63.34 63.93 64.77 65.4 66.36 67.35 67.89 68.67 67.95 68.91 68.3 67.33 66.94 66.35 66.05 62.8 62.47 63.29
100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
95 99.03 99.04 98.32 97.41 97.12 97.74 98.66 98.18 97.72 96.76 95.9 96.28 96.15 97 96.34 98.95 98.2 97.58
Dresden 90 98.85 97.91 97.83 97.41 96.92 97.28 97.74 97.22 96.58 96.45 95.9 95.93 96.15 96.73 96.34 97.93 97.8 96.91
Erosion 85 97.01 97.24 97.67 97.41 96.75 95.78 95.15 94.63 94.76 93.76 94.61 95.17 94.33 95.24 95.8 96.52 97.01 96.81
80 87.43 88.17 89.16 88.31 88.29 87.49 87.41 87.22 87.66 87.88 88.45 88.11 87.41 86.53 86.72 89.32 89.85 90.15
75 78.63 78.3 78.64 79.59 78.84 79.73 79.33 78.57 79.42 78.93 78.02 78.05 78.76 78.35 79.28 80.03 79.32 79.56
70 76.19 76.7 76.22 76.2 76.3 76.66 77.39 77.18 76.73 77.3 78.02 77.9 78.35 78.35 79.02 82.23 81.79 82.78
100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
95 98.32 97.86 98.41 97.59 96.9 97.26 96.35 95.9 96.47 97.34 96.93 97.7 97.37 98.35 97.95 98.83 99.81 99.26
Dresden 90 98.32 97.86 98.16 97.59 96.9 97.26 96.35 95.9 96.18 97.14 96.93 96.52 96.36 95.99 95.26 97.58 98 98.06
Dilation 85 96.74 96.67 97.06 96.72 96.5 96.03 96.35 95.87 96.18 96.73 96.75 96.39 95.62 95.99 95.26 96.86 97.47 98.06
80 87.78 88.29 89.27 88.79 88.14 88.85 88.18 88.43 87.66 87.65 87.2 87.23 87.53 86.79 87.78 90.24 89.36 88.43
75 79.34 79.2 79.79 80.75 80.9 80.69 79.84 79.66 80.51 79.95 80.72 81.69 81.27 80.45 79.62 80.21 80.72 81.13
70 73.81 74.7 74.9 73.98 73.64 74.33 74.31 74.23 74.3 73.4 72.66 72.22 72.15 72.65 72.27 75.01 74.82 74.74
100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
95 98.78 98.97 99.03 98.65 99.08 98.16 98.17 97.99 97.97 97.96 97.27 97.42 97.82 96.87 95.97 98.55 98.8 99.74
RAISE-1k 90 97.34 97.53 97.16 96.34 96.15 95.42 95.96 95.47 96.4 95.68 96.43 97.29 97.46 96.87 95.97 98.04 98.03 98.96
Erosion 85 94.23 94.54 95.24 94.41 94.82 94.05 93.78 94.61 95.26 94.33 93.35 94.21 94.98 95.76 95.97 96.94 97.65 96.69
80 88.17 88.53 87.96 87.19 86.43 86.89 85.98 86.4 86.05 86.19 85.98 85.16 84.73 84.48 85.26 91.52 91.63 90.8
75 78.26 78.57 77.6 78.56 78.8 77.9 77.18 77.76 78.21 78.35 78.16 78.02 77.67 77.95 77.11 82.99 83.14 82.75
70 77 77.63 76.94 76.42 76.53 76.45 77.18 76.68 77.42 78.25 78.16 78.02 77.67 77.95 77.11 82.22 81.96 81
100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
95 98.22 98.07 98.25 99.04 99.04 99.12 99.75 99.25 99.02 98.94 98.48 97.59 96.83 96.17 95.26 99.36 99.49 99.92
RAISE-1k 90 97.81 98.07 97.33 98.33 99.04 99.12 99.72 99.25 98.36 97.53 97.41 97.59 96.83 96.17 95.26 97.12 97.96 98.49
Dilation 85 94.74 93.79 93.83 94.55 95.13 94.24 94.73 93.81 94.58 95.42 94.56 94.25 95.24 94.95 94.62 96.13 95.71 96.06
80 88.46 88.97 89.76 90.07 90.22 91.18 91.09 91.07 91.61 91.52 90.57 89.65 90.63 91.07 90.53 92.52 93.43 92.55
75 77.85 76.87 76.98 76.83 76.34 75.63 75.31 74.31 73.89 74.45 74.17 73.83 74.53 74.61 74.11 82.68 82.58 83.06
70 77.46 76.73 76.14 75.66 74.74 75.42 75.31 74.31 73.89 74.45 74.17 73.83 73.67 73.2 72.3 81.93 81.13 80.91
K= 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36
100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
95 89.09 89.07 88.87 89.13 89.99 90.35 90.19 89.87 90.83 90.44 90.33 89.98 90.61 90.67 91.31 91.03 91.34 92.26
UCID 90 83.29 83.67 83.86 84.17 84.82 84.95 84.52 80.42 81.14 81.34 80.88 80.55 80.2 80.56 80.39 80.48 80.29 80.36
Erosion 85 82.21 82.96 83.18 83.86 83.4 83.55 84.19 74.24 74.92 75.37 76.1 75.74 75.41 76.02 75.64 76.41 77.04 77.5
80 76.7 77.01 76.62 76.12 76.52 77.34 78.11 75.44 75.64 75.18 75.88 75.71 76.34 76.72 77.12 77.92 77.78 78.46
75 68.8 69.23 69.86 69.63 70.3 70.59 70.92 68.14 68.08 67.81 68.61 68.24 67.99 68.44 68 68.24 68.29 69.2
70 63.53 64.52 64.71 65.69 65.67 65.73 65.93 62.58 62.74 62.5 62.69 62.64 63.05 63.03 63.58 63.61 64.04 64.03
100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
95 90.14 89.72 89.68 89.94 90.06 89.97 90.89 89.91 89.73 89.99 89.58 89.76 90.08 90.73 90.41 91.27 91.62 92.09
UCID 90 82.77 83.18 83.39 83.58 83.57 83.96 84.61 81 80.83 80.75 81.46 82.18 82.47 82.92 82.88 83.3 84.22 85.22
Dilation 85 81.87 81.64 81.82 82.17 82.45 83.34 83.67 73.87 74.19 74.78 75 74.7 75.24 76.12 75.96 76.42 77.26 76.96
80 78.69 78.59 78.39 78.19 77.88 78.36 77.88 75.78 76.59 76.2 76.21 76.52 76.92 77.72 78.03 78.67 78.44 78.57
75 70.71 70.46 70.2 70.05 69.65 70.41 70.18 69.02 69.83 70.08 70.63 70.24 70.01 69.69 69.94 70.55 70.67 71.66
70 63.16 63.26 63.25 63.96 64.52 64.91 64.91 62.27 63.22 63.66 63.38 63.19 63.77 63.46 64.2 64.98 64.63 65.01
100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
95 97.84 98.17 98.73 98.09 98.17 98.03 97.07 98.03 97.22 97.95 97.84 98.19 97.97 97.16 96.67 96.15 96.68 95.76
Dresden 90 97.59 98.17 97.78 97.98 97.66 96.72 97.07 96.65 95.72 96.57 96.58 96.44 96.06 96.71 96.27 96.15 96.25 95.69
Erosion 85 96.51 97.21 96.96 96.56 95.69 95.69 95.62 93.89 94.15 94.19 93.67 94.08 93.95 93.5 93.16 92.9 92.03 91.64
80 90.05 89.45 89.25 88.97 88.55 89 88.3 89.3 88.9 89.56 89.71 90.29 90.16 89.64 88.65 89.32 89.67 89.83
75 79.4 80.18 79.99 79.43 79.61 78.99 78.83 81.44 82.37 82.33 81.64 81.11 81.26 81.41 80.83 81.14 81.31 81.89
70 83.08 83.64 83.46 84.22 84.65 84.18 84.58 73.26 73.62 73.98 74.6 75.4 75.12 75.53 75.87 76.29 76.6 76.76
100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
95 99.05 99.96 99.53 99.12 98.32 97.5 98.05 97.84 98.45 98.65 98 97.64 97.06 97.75 97.67 98.33 97.44 96.55
Dresden 90 97.06 97.5 97.17 97.52 96.7 95.83 96.75 97.51 97.07 96.95 96.53 97.21 96.36 95.98 95.95 95.31 96.01 95.67
Dilation 85 97.06 97.21 97.17 97.52 96.7 95.83 95.38 94.81 94.71 95.65 96.16 95.28 95.41 95.35 95.21 94.61 94.76 93.8
80 89.2 88.78 88.51 88.51 88.83 89.82 89.64 89.69 90.24 90.14 89.31 90.04 89.09 89.21 88.83 88.82 89.09 89.14
75 81.46 81.36 81.68 81.56 81.32 81.12 80.21 82.31 81.63 81.76 81.92 82.78 82.69 82.52 82.3 81.63 82.55 81.79
70 75.02 75.73 76.2 75.7 75.99 76.95 77.03 74.12 74.83 75.04 75.79 75.83 76.52 77.19 77.55 77.79 78.31 79.17
100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
95 98.83 98.58 98.88 98.85 99.06 99.47 99.94 99.32 99.55 99.84 99.31 99.09 99.22 99.76 98.82 98.21 98.2 99.12
RAISE-1k 90 98.83 98.58 97.66 96.8 97.35 96.72 96.1 95.18 95.64 94.67 94.48 93.48 93.21 93.58 94.35 94.94 94.98 95.93
Erosion 85 96.37 96.4 97.19 96.8 96.13 95.33 95.21 95.18 94.86 94.67 94.48 93.48 93.21 92.95 92.19 92.48 93 92.68
80 90.57 90.51 90.98 90.39 90.51 90.19 90.71 91.67 92.37 91.68 91.66 91.1 90.75 90.37 90.34 90.3 90.04 90.56
75 83.32 82.36 81.42 80.9 80.11 80.86 79.97 80.78 81.04 81.96 82.47 83.07 83.6 83.68 83.25 82.73 82.23 82.54
70 81.53 81.39 81.13 80.9 80.04 79.3 78.6 78.27 79.24 80.19 79.48 78.72 77.83 77.71 76.79 77.3 77.45 77.42
100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
95 99.92 98.15 97.56 97.97 97.49 98.19 98.14 97.62 96.9 96.95 96.87 96.87 96.3 96.97 97.66 97.25 97.73 96.86
RAISE-1k 90 98.23 98.15 97.56 96.76 97.09 96.81 97.07 97.62 96.9 96.95 96.02 96.87 96.3 96.39 97.08 96.44 96.94 96.86
Dilation 85 96.35 96.87 96.04 95.07 94.21 93.28 92.32 91.68 90.72 90.25 90.94 91.06 90.89 90.12 91.1 91.9 92.24 92.68
80 91.65 92.21 92.54 92.03 92.03 91.36 90.7 91.52 90.72 90.25 90.86 90.65 90.44 90.12 91.1 91.9 91.36 90.48
75 83.39 83.2 83.23 84.19 84.86 84.09 84.6 83.68 82.92 83.39 84.3 83.8 84.72 84.97 85.53 86.06 85.91 86.87
70 80.53 80.4 80.29 80.78 81.53 81.77 82.03 81.25 80.97 81.19 80.67 80.4 81.05 82.03 81.23 81.1 82.08 82.59
VOLUME 4, 2016 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2965745, IEEE Access
Boato et al.: Morphological Filter Detector for Image Forensics Applications
TABLE 4: Results on erosion detection on Raise under
different attacks, using SVM with RBF kernel, 10-fold cross
validation.
QF Pepper & Salt Gaussian (3x3) Median Filter (3x3)
100 97.92 81.21 87.94
95 95.72 81.14 86.44
90 95.19 81.01 84.27
85 91.79 79.16 80.67
80 86.99 78.76 80.57
75 83.04 75.06 76.42
70 81.23 70.24 71.49
ing is applied to each image and then this is contaminated
with a noise. Five different attacks - pepper & salt noise,
Gaussian filtering with window size 3×3, median filtering
with window size 3×3, scaling at 1.1x and scaling at
0.9x - were applied and average results (on all kernels) for
dilation detection using SVM with RBF kernel on RAISE1k
are correspondingly: 97.92%, 81.21%, 87.94%, 67.44%, and
67.81% (the results are summarized in Figure 9. According to
the results, we can claim that the proposed method is robust
against noise addition but not against processing involving
interpolation (e.g., resizing).
We applied a further test to understand if the proposed
method can still detect the morphological filter after two level
of processing where the second one is a compression. Thus,
each image after a morphological filter is first contaminated
with a noise, and then is compressed under different quality
factors. The results in Table 4 show that the proposed method
can still detect the morphological filter, however, its perfor-
mance is getting worst as the compression level increases.
5 10 15 20 25 30 35
Kernel
5
10
15
20
25
30
35
Kernel
0
10
20
30
40
50
60
70
80
90
FIGURE 8: Confusion matrix on erosion detection for mul-
tiple kernels. Dataset: Raise, QF = 90, full resolution. Values
are normalized and are in percentage.
0
10
20
30
40
50
60
70
80
90
100
Pepper & Salt
Gaussian 3x3
Median 3x3
Scale 1 .1x
Scale 0 .9x
FIGURE 9: Precision (computed as percentage) on dilation
detection under different attacks. Dataset: RAISE-1k, full
resolution, QF = 100, using SVM with RBF kernel, 10-fold
cross validation. All 36 kernels are computed and the results
are shown in the error bars.
D. MORPHOLOGICAL FILTERS VERSUS OTHER
FILTERS
Finally, we want to determine the capability of our approach
in distinguishing between morphological filters from other
filters that produce similar results. In particular, we con-
sidered three filters: pepper & salt, Gaussian lowpass and
median filtering. In all cases, the filtered images are uncom-
pressed and passed to the detector to reveal possible false
alarms. In this experiment, we trained the erosion versus pris-
tine classifier with 800 uncompressed images from RAISE in
full resolution. The trained classifier is then applied to the
rest of the 200 images, filtered with pepper & salt, Gaussian
lowpass and median filters. The average number of false
positives (over 200 images) for pepper & salt, Gaussian with
window size 3×3, Gaussian with window size 5×5, median
with window size 3×3, and median with window size 5×5
are 3.51%, 5.72%, 5.28%, 14.11%, and 13.17%, respectively.
The results are summarized in Figure 10. This confirms that
the proposed method can distinguish between morphological
filter and other filters. Only in the case of median filters the
confusion increases, but this is very reasonable, since in the
case of grayscale images the effect of morphological filters
and median filters is very similar.
VI. CONCLUSIONS
In this work we propose an effective detection strategy to
assess the use of morphological filtering in a grayscale con-
text. We deal with uncompressed images proposing a deter-
ministic approach, based on mathematical properties enjoyed
by basic morphological operators. We additionally propose
a modified pipeline to detect morphological processing in
compressed images, by exploiting the difference histogram
information as feature for classification. We present a testing
phase in which both uncompressed and compressed scenarios
10 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2965745, IEEE Access
Boato et al.: Morphological Filter Detector for Image Forensics Applications
0
10
20
30
40
50
60
70
80
90
100
Pepper & Salt
Gaussian (3x3)
Gaussian (5x5)
Median Filter (3x3) Median Filter (5x5)
FIGURE 10: False Positive (computed as percentage) on 200
images from RAISE on erosion detection over all 36 filters.
Images are in raw format, full resolution. The classifiers were
trained on 800 images (eroded vs. non-filtered) and tested on
the rest 200 images (filtered with pepper & salt, Gaussian and
median).
are taken into consideration. Results show the effectiveness
of our proposed approach in both cases. Moreover, the pro-
posed approach is able to determine the adopted structuring
element for moderate compression factors, and is robust
against a number of attacks.
REFERENCES
[1] M. Stamm, M. Wu, and K. Liu, “Information forensics: an overview of the
first decade,” IEEE Access, pp. 167–200, 2013.
[2] H. Farid, Photo forensics, M. Press, Ed., 2016.
[3] A. T. Ho, Handbook of digital forensics of multimedia data and devices.
John Wiley & Sons, 2015.
[4] A. Piva, “An overview on image forensics,” ISRN Signal Processing, 2013.
[5] H. Sencar and N. Memon, Eds., Digital Image Forensics - There is more
to a picture than meets the eye. Springer, 2013.
[6] D. Cozzolino and L. Verdoliva, “Noiseprint: A cnn-based camera model
fingerprint,” IEEE Transactions on Information Forensics and Security,
vol. 15, pp. 144–159, 2020.
[7] A. D. Rosa, A. Piva, M. Fontani, and M. Iuliani, “Investigating multimedia
contents,” in IEEE International Carnahan Conference on Security Tech-
nology (ICCST), 2014, pp. 1–6.
[8] P. Bestagini, M. Tagliasacchi, and S. Tubaro, “Image phylogeny tree re-
construction based on region selection,” in IEEE International Conference
on Acoustics, Speech and Signal Processing, 2016.
[9] Z. Dias, A. Rocha, and S. Goldenstein, “Image phylogeny by minimal
spanning trees,” IEEE Transactions on Information Forensics and Security,
vol. 7, no. 2, pp. 774–788, 2012.
[10] A. de Oliveira and et al., “Multiple parenting phylogeny relationships in
digital images,” IEEE Transactions on Information Forensics and Security,
vol. 11, no. 2, pp. 328–343, 2016.
[11] F. de O. Costa and et al., “Image phylogeny forests reconstruction,” IEEE
Transactions on Information Forensics and Security, vol. 9, no. 10, pp.
1533–1546, 2014.
[12] E. Ardizzone, A. Bruno, and G. Mazzola, “Copy-move forgery detection
by matching triangles of keypoints,” IEEE Transactions on Information
Forensics and Security, vol. 10, no. 10, pp. 2084–2094, 2015.
[13] D. Cozzolino, G. Poggi, and L. Verdoliva, “Efficient dense-field copy-
move forgery detection,” IEEE Transactions on Information Forensics and
Security, vol. 10, no. 11, pp. 2284–2297, 2015.
[14] I. Amerini, L. Ballan, R. Caldelli, A. D. Bimbo, and G. Serra, “A SIFT-
based forensic method for copy-move attack detection and transformation
recovery,” IEEE Transactions on Information Forensics and Security,
vol. 6, no. 3, pp. 1099–1110, 2011.
[15] D. Cozzolino, G. Poggi, and L. Verdoliva, “Splicebuster: a new blind
image splicing detector,” in IEEE Workshop on Informations Forensics
and Security (WIFS), 2015, pp. 1–6.
[16] X. Zhao, S. Wang, S. Li, and J. Li, “Passive image-splicing detection
by a 2-D noncausal Markov model,” IEEE Transactions on Circuits and
Systems for Video Technology, vol. 25, no. 2, pp. 185–199, 2015.
[17] D. Vàzquez-Padìn, F. Pèrez-Gonzàlez, and P. Comesaña-Alfaro, “A ran-
dom matrix approach to the forensic analysis of upscaled images,” IEEE
Transactions on Information Forensics and Security, vol. 12, no. 9, pp.
2115–2130, 2017.
[18] C. Pasquini and R. Böhme, “Information-theoretic bounds for the foren-
sic detection of downscaled signals,” IEEE Transactions on Information
Forensics and Security, vol. 14, no. 7, pp. 1928–1943, 2019.
[19] C. Pasquini, G. Boato, and F. Pèrez-Gonzàlez, “Statistical detection of
JPEG traces in digital images in uncompressed formats,” IEEE Transac-
tions on Information Forensics and Security, vol. 12, no. 12, pp. 2890–
2905, 2017.
[20] W. Shan, Y. Yi, R. Huang, and Y. Xie, “Robust contrast enhancement
forensics based on convolutional neural networks,” Signal Processing:
Image Communication, vol. 71, pp. 138–146, 2019.
[21] V. Conotter, H. Farid, and G. Boato, “Detecting photo manipulation
on signs and billboards,” in IEEE International Conference on Image
Processing (ICIP), 2010.
[22] E. Kee, J. O’Brien, and H. Farid, “Exposing photo manipulation from
shading and shadows,” ACM Transactions on Graphics, vol. 33, no. 165,
pp. 1–21, 2014.
[23] M. Iuliani, G. Fabbri, and A. Piva, “Image splicing detection based
on general perspective constraints,” in IEEE Workshop on Informations
Forensics and Security (WIFS), 2015, pp. 1–6.
[24] K. Bahrami, A. Kot, L. Li, and H. Li, “Blurred image splicing localization
by exposing blur type inconsistency,” IEEE Trans. on Information Foren-
sics and Security, vol. 10, no. 5, pp. 999–1009, 2015.
[25] G. Cao, Y. Zhao, R. Ni, and X. Li, “Contrast enhancement-based forensics
in digital images,” IEEE Trans. on Information Forensics and Security,
vol. 9, no. 3, pp. 515–525, 2014.
[26] G. Cao, Y. Zhao, R. Ni, and A. Kot, “Unsharp masking sharpening
detection via overshoot artifacts analysis,” IEEE Signal Processing Letters,
vol. 18, no. 10, pp. 603–607, 2011.
[27] M. Kirchner and J. Fridrich, “On detection of median filtering in digital
images,” in SPIE, vol. 7541, (2010), pp. 101–112.
[28] G. Cao, Y. Zhao, R. Ni, L. Yu, and H. Tian, “Forensic detection of median
filtering in digital images,” in IEEE Int. Conf. on Multimedia and Expo,
ICME2010, (2010), pp. 89–94.
[29] H.-D. Yuan, “Blind forensics of median filtering in digital images,” IEEE
Trans. Information Forensics and Security, vol. 6, no. 4, pp. 1335–1345,
2011.
[30] X. Kang, M. C. Stamm, A. Peng, and K. J. R. Liu, “Robust median fil-
tering forensics using an autoregressive model,” IEEE Trans. Information
Forensics and Security, vol. 8, no. 9, pp. 1456–1468, 2013.
[31] C. Chen, J. Ni, and J. Huang, “Blind detection of median filtering in
digital images: A difference domain based approach,” IEEE Trans. Image
Processing, vol. 22, no. 12, pp. 4699–4710, 2013.
[32] Y. Zhang, S. Li, S. Wang, and Y. Q. Shi, “Revealing the traces of median
filtering using high-order local ternary patterns,” IEEE Signal Processing
Letters, vol. 21, no. 3, pp. 275–280, 2014.
[33] J. Chen, X. Kang, Y. Liu, and Z. J. Wang, “Median filtering forensics
based on convolutional neural networks,” IEEE Signal Processing Letters,
vol. 22, no. 11, pp. 1849–1853, 2015.
[34] C. Pasquini, G. Boato, N. Anjalic, and F. D. Natale, “A deterministic
approach to detect median filtering in 1d data,” IEEE Transactions on
Information Forensics and Security, vol. 11, no. 7, pp. 1425–1437, 2016.
[35] H. Gao, M. Hu, T. Gao, and R. Cheng, “Robust detection of median filter-
ing based on combined features of difference image,” Signal Processing:
Image Communication, vol. 72, pp. 126–133, 2019.
[36] A. Haas, G. Matheron, and J. Serra, “Morphologie mathématique et
granulométries en place,” in Annales des mines, vol. 11, no. 736-753,
1967, pp. 7–3.
[37] N. M. Al-Shereefi, “Morphological filter is an active tool for edge detec-
tion in noisy image,” British Journal of Science, vol. 3, no. 2, pp. 148–156,
2012.
[38] P. Maragos, “Morphological filtering for image enhancement and feature
detection,” in The Image and Video Processing Handbook, 2004.
VOLUME 4, 2016 11

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2965745, IEEE Access
Boato et al.: Morphological Filter Detector for Image Forensics Applications
[39] F. G. De Natale and G. Boato, “Detecting morphological filtering of
binary images,” IEEE Transactions on Information Forensics and Security,
vol. 12, no. 5, pp. 1207–1217, 2017.
[40] Y. Nakagawa and A. Rosenfeld, “A note on the use of local min and max
operations in digital picture processing,” Maryland University College
Park Computer Science Center, Tech. Rep., 1977.
[41] G. Schaefer and M. Stich, “Ucid: An uncompressed color image database,”
in Storage and Retrieval Methods and Applications for Multimedia 2004,
vol. 5307. International Society for Optics and Photonics, 2003.
[42] T. Gloe and R. Böhme, “The dresden image database for benchmarking
digital image forensics,” Journal of Digital Forensic Practice, vol. 3, no.
2-4, pp. 150–159, 2010.
[43] D.-T. Dang-Nguyen, C. Pasquini, V. Conotter, and G. Boato, “Raise: a raw
images dataset for digital image forensics,” in ACM Multimedia Systems
Conference, 2015.
GIULIA BOATO is Associate Professor at the
Department of Information Engineering and Com-
puter Science (DISI) of the University of Trento
(Italy). She is Associate Editor of the IEEE Trans-
actions on Image Processing, the ELSEVIER Sig-
nal Processing: Image Communication. Her re-
search interests are focused on image and signal
processing, with particular attention to multimedia
data security and digital image and video foren-
sics, but also intelligent multidimensional data
management and analysis. She is author of 120 papers in international
conferences and journals, with H-index 22 (Google Scholar). She is elected
member of the IEEE Multimedia Signal Processing Technical Committee
(MMSP TC), the IEEE Information Forensics and Security Technical Com-
mittee (IFS TC) and the EURASIP Special Area Teams Biometrics, Data
Forensics, and Security.
DUC-TIEN DANG-NGUYEN is currently an As-
sociate Professor at the Department of Informa-
tion Science and Media Studies (Infomedia), Uni-
versity of Bergen (Norway). His main area of
expertise is on multimedia forensics, lifelogging,
multimedia retrieval, and computer vision. He is
author of 90 papers in international conferences
and journals, with H-index 18 (Google Scholar,
December 2019). He is member of IEEE Signal
Processing, ACM, SIGIR, IAPR and INSTICC.
He has organised and chaired over 20 workshops and special sessions,
has been served as program committee for numerous international confer-
ences and workshops, and has peer-reviewed for many top journals and
conferences in the fields of lifelogging, multimedia forensics, and pattern
recognition.
FRANCESCO G.B. DE NATALE was the Head
of the Department of Information Engineering and
Computer Science (DISI), University of Trento
(Italy) from 2007 to 2010, where he currently
leads the MMLab Research Laboratory. He is also
a Professor of telecommunications with the Uni-
versity of Trento. His research interests include
multimedia communications, where he published
over 200 works in the major international journals
and conferences. He is also a member of the Board
of Governors of the CNIT Consortium. He has been an Associate Editor of
the IEEE Transactions on Multimedia and the IEEE Transactions on Circuits
and System for Video Technologies, and currently associate editor of IEEE
Transaction on Image Processing.
12 VOLUME 4, 2016