Autofocus in infrared microscopy
Raphael Abele, Daniele Fronte, Pierre-Yvan Liardet, Jean-Marc Boi, Jean-Luc
Damoiseaux, Djamal Merad
To cite this version:
Raphael Abele, Daniele Fronte, Pierre-Yvan Liardet, Jean-Marc Boi, Jean-Luc Damoiseaux, et
al.. Autofocus in infrared microscopy. 2018 IEEE 23rd International Conference on Emerging Technologies and Factory Automation (ETFA), Sep 2018, Turin, France.
pp.631-637,
10.1109/ETFA.2018.8502648. hal-02060828
HAL Id: hal-02060828
https://hal-amu.archives-ouvertes.fr/hal-02060828
Submitted on 7 Mar 2019
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of scientific research documents, whether they are published or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est
destinée au dépôt et à la diffusion de documents
scientifiques de niveau recherche, publiés ou non,
émanant des établissements d’enseignement et de
recherche français ou étrangers, des laboratoires
publics ou privés.
Autofocus in infrared microscopy
Raphael Abele∗† , Daniele Fronte∗ , Pierre-Yvan Liardet∗ , Jean-Marc Boi† , Jean-Luc Damoiseaux† , Djamal Merad†
∗
STMicroelectronics
Rousset, France
raphael.abele, daniele.fronte, pierre-yvan.liardet @st.com
Abstract—Autofocus (AF) is a widely investigated subject in the
fields of natural scene images, industrial assembly and biologic
microscopy. This paper proposes a new effective AF method for
infrared (IR) microscopy in the context of the Integrated Circuit
industry (IC). The proposed method operates in the wavelet
domain using a custom orthogonal wavelet for the 2D Discrete
Wavelet Transform (DWT). The quality criterion of our AF
algorithm relies on the standard deviance of the DWT coefficients,
computed per subband and per level. Tested on several optical
magnifying lenses, our method is robust time-efficient, and usable
on-the-fly in the IC location system.
Index Terms—Infrared imaging, Microscopy, Focusing, Image
analysis, Image quality, Image decomposition, Wavelet transforms, Statistics
†
Laboratoire d’Informatique et Systèmes
Aix-Marseille University
Marseille, France
raphael.abele, jean-marc.boi,
jean-luc.damoiseaux, djamal.merad @lis-lab.fr
3D positioning of the laser on the SOI is done with an infrared
(IR) camera that can see through the silicon layer. In our work,
the optical system and the laser source are interdependent,
so the optical adjustment implies laser position adjustment.
Usually, these adjustments are made manually by skilled men,
which raises two problems: time loss and non-reproducibility
of the characterization (because of potential imprecision). An
autofocus (AF) system could partially solve this problem by
automating one dimension of the 3D positioning process.
Section (II) shortly introduces some works related to the AF.
Then, section III presents our particular work environment.
Section IV describes the AF method we propose, and section
V details the overall experiment results.
I. I NTRODUCTION
Integrated Circuits (ICs1 ) are electronic components whose
applications seem unlimited: they are used in many fields such
as wearable technologies and IoT (Internet of objects). In
order to protect the privacy of users, some ICs need to be
secured, and their security needs to be validated. One way
to characterize a secure IC is to study its behavior following
a physical disruption. Such disruption may be obtained by a
laser shot targeted on the internal structures of the IC. The
accuracy of the shot relies on the precision of the laser power
calibration and 3D positioning inside the IC. In our study, the
Object Of Interest (OOI) is considered as made of three layers
of different materials: (1) the silicon, (2) a conductor material
forming the conductive tracks, and (3) an electrical insulator,
as illustrated in Fig. 1.
The internal OOI structures are visible on the surface of the
conductive tracks, which is the Surface Of Interest (SOI). The
II. A ROUND THE AUTOFOCUS
The AF mechanism is a deterministic algorithm that determines the lens position for which the system is "well focused"
on the scene/object of interest (see Fig. 2), that is, the position
for which the image of the scene or object is sharpest.
Two AF approaches exist: active and passive. The active
approach depends on an additional system measuring the
distance from the system to the scene/object. For example,
ultrasonic sound waves and IR light reflection are two ways of
measuring distances. Once the optics-scene distance is known,
the correct lens position to obtain a focused image is estimated.
Such an active system is not available, so we consider a passive
approach relying on image analysis. This analysis determines
the most relevant object/scene image, which we then use to
find the correct focus. At this point, we are confronted with
the issue of Image Quality Assessment (IQA).
Project funded by the Association Nationale de la Recherche et de la
Technologie (ANRT).
1 A glossary is present at the end of this paper.
Fig. 1: Concept mapping of three materials composing our IC.
Fig. 2: Minimalist optical system; the image on plane 2 (green)
is well focused on the optical sensor whereas the image on
plane 1 is not.
The Human Visual System (HVS) is the most reliable tool
for IQA: the main difficulty is the interpretation of human
subjectivity with algorithms following objective rules and
criteria [1].
Some subjective methods attempt to emulate HVS using
the Mean Opinion Score, but they are time-consuming and
generally non-deterministic; these methods are partly based on
objective metrics pooled together to predict the quality score
[2].
Objective IQA methods attempt to be correlated with the
HVS. Depending on the amount of available information from
the image source, methods are categorized as Full-Reference
(FR) based if an original, non-distorted image is available [3]
or No-Reference (NR) based otherwise [4]. If just some of
the original image features are known, the method is qualified
as Reduced-Reference (RR) based. See reference [5] as an
example. RR-IQA methods are generally guided/optimized
NR-IQA methods. In the AF context, both of the NR and
RR-IQA methods are used, since the quality criterion is
hypothetical. Images are ranked according to this criterion,
and the best-ranked image is associated with the correct lens
position.
Statistics of order 1, 2 and higher are widely used in signal
processing since the 1960’s. In image analysis, each statistic
point out an image feature [6]:
• Average Gradient (AG): reflects the contrast and the
clarity of the image. It can be used to measure the spatial
resolution of a test image, where a larger average gradient
indicates better spatial resolution.
• Mean metric: indicates the image average brightness
level. For equivalent scenery, image brightness increases
with the mean.
• Auto-Correlation (AC): spotlight redundant data. If an
image is blurred or the edges are smoothed, the correlation between neighboring pixels becomes high.
• Entropy metric: measures the information quantity of an
image. If the probability of occurrence of each gray level
is low, the entropy is high, and vice versa.
• Standard Deviation (SD): reflects the contrast of the
image; the image contrast increases with the SD.
• Kurtosis metric: a statistical measure of the degree of
sharpness or flatness of a distribution (i.e., average slope
and energy concentration). Increases with the depth of
focus.
• Skewness metric: a statistical measure of the direction and
extent to which a data set deviates from a distribution. For
a standard normal distribution, high skewness indicates
asymmetry of the data. In this case, the data contains a
greater amount of information.
These statistics can be computed on the image data or on
its transform, that is, in spatial, frequency or time-frequency
domains.
Since AF algorithms rely on specific image analysis, they
cannot be generalized. Several methods exist, each adapted
to its context such as images of natural scenes [5], [7]–[9],
low-contrast images [10] or microscopic images [11]–[13]. In
paper [13], the author studied several autofocus methods in
a particular case (lipid droplets inside microscopic worms).
The focus criterion is clarity-evaluation based. He tested
the performance and accuracy of sixteen autofocus methods:
histogram, intensity, statistic, derivative and transform-based.
He found that the absolute Tenengrad algorithm had the best
performance against accuracy for its specifics. Finally, each
approach uses well justified metrics. In our proposal we define
metrics adapted to IR microscopy for the IC industry, which
implies some constraints (see Section III).
III. W ORK ENVIRONMENT
A. Available material
Observation of the SOI under the silicon layer is possible
thanks to its optical properties. Indeed, around the wavelength
1µm and higher, the silicon absorption coefficient is not
significant [14], [15]. In other words, silicon is "transparent"
to lights whose wavelengths are around 1µm and above (such
as IR light).
Our motorized optical system provided by AlphaNov2 is
composed of:
• an uncooled IR camera
• an optical microscope with 4 magnifying lenses (2.5x,
5x, 20x and 50x) allowing us to observe the OOI (around
5mm2 ) and its internal structures (micro-metric scale)
Three types of IR camera are available on the market: LongWave, Mid-Wave and Short-Wave (LWIR, MWIR and SWIR,
respectively). On the one hand, MWIR and LWIR sensors
detect thermal emissions from objects, and are efficient when
these are warmer than their surroundings. On the other hand,
SWIR cameras use the reflected light, much like the slightly
shorter wavelengths of the visible spectrum (see Fig. 3). In
our experiment we worked with the latter type of camera.
With the unique capabilities of our SWIR camera, we are
able to see the SOI through the silicon layer. To propose an
automated method for focusing on the SOI, our algorithm must
take in account some specifics arising from the camera and
from the SOI (see Section III-B).
B. Constraints and special features
Concerning the camera:
2 www.alphanov.com
Fig. 3: Infrared wavelength in the electromagnetic spectrum.
2
The IR camera sensor is an InGaAs, sensitive to a large
IR wavelength range (900nm to 1700nm); the embedded
Wide Dynamic Range technology (WDR) dynamically detects
the relevant wavelengths to take into account for best image
quality. Despite its accuracy, this camera stays sensitive to
external thermal interference. Since the sensor is uncooled
(thermally not insulated), the thermal instability of the environment produces thermal noise, which affects the image
acquisition (see Fig. 4).
Cooled IR camera
Fig. 6, everything is orthogonally disposed, in regard to the
rectangular shape of the component. This criterion is a good
candidate for our AF algorithm.
However, a difficulty arises from the SOI image texture. An
IC is built by microscopic material deposition, which implies
a textural grain visible with a high magnifying lens (20x and
50x). Considering the SOI images, this grain can be compared
to a noise in the images: structures are clusters of points, more
or less dense (illustrated in Fig. 7). In the worst case, detection
of the structure boundaries turns out to be difficult and the
images could need preprocessing to be well treated.
To sum up, we may state that:
a. Sharpness is not a good criterion to make the difference
between the silicon surface and the SOI.
b. On the SOI, horizontal and vertical structures boundaries
are visible.
c. Because of textural grain and noise, topological information is not pertinent (e.g. line detection).
d. The AF method has to be effective for each magnifying
lens (2.5x, 5x, 20x and 50x).
Since we know which features are expected in the targeted
image (b.), our IQA can be qualified as RR-based. To avoid
preprocessing such as image denoising (c.), we propose a
transform-based analysis in the next section.
Uncooled IR camera
Fig. 4: An example of images acquired with cooled and
uncooled infrared cameras (source: flir.com).
Concerning the object of interest:
Two distinct surfaces are considered for the autofocus
algorithm: the silicon and the SOI. As said in III-A, silicon is
transparent to our IR light and should not be visible. However,
in many cases, contaminants (e.g. dusts) are ingrained on the
silicon, which reveals the surface to any light. This implies a
possible conflict of interest, considering the sharpness criterion: the silicon surface with its contaminants may be sharper
than the SOI. This point underlines that sharpness is not, in
practice, the appropriate criterion to exclusively characterize
an SOI. Fig. 5 shows to the notable states during the autofocus
process.
We may compare the conductive layer of an IC to a
city, and its internal structures to buildings. As shown in
IV. AUTOFOCUS METHOD PROPOSAL
The criterion analyzed for an AF system has to be strong
enough to prevent false results. In that sense, for a robust AF
algorithm, we need a robust quality metric.
A. Analysis domain proposal
As previously described, image preprocessing should be performed before any image-based analysis (e.g. denoising). But
it could break information continuity in our images (e.g. spatial
information), and could be time-consuming. To preserve this
continuity, our approach is transform-based, which may also
save computation time. This transform is described below.
Wavelet transform:
Several image transforms exists. One of the most commonly
used is wavelet decomposition, but has limitation due to its orthogonality (i.e. there is little orientation information). Others
exist such as the steerable pyramid and curvelet transforms:
Fig. 5: Images acquired at several focal point positions (green
points) according to the Z-axis. Component out of focus (left),
focused on the silicon surface (middle) and on the conductive
track surface (right).
Fig. 6: A simple representation of an IC and its internal
structures.
3
Fig. 7: Two lines delimiting
structures boundary (50x).
the steerable pyramid transform overcome this orientation
limitation and the curvelet transform is an extension of the
wavelet concept, with less redundancy and more orientation
information [16]–[18]. However, according to the features
described in section III-B, the main information needed for
our analysis is the horizontal and vertical components of
images. Therefore, we only consider wavelet decomposition
for our IQA. Moreover, it may permit skipping some useless
information related to material texture or even noise. For this
purpose, the wavelet choice is important, considering that each
wavelet matches specific signal information.
3. Cohen-Daubechies-Fauveau (CDF 9/7) 9/7 wavelet [21],
having a great capacity to extract textures, is illustrated
in Tab. IV.
4. An Optimized CDF 9/7 9/7 wavelet proposed in [22] and
illustrated in Tab. V.
Our goal is to find which wavelet match the most the
relevant information in the image, that is, structures salience.
Tab. VI shows obtained sharpness maps. We visually estimate
that COW seems to match the interesting information, that
is, relative to structures’ salience. Contrary to CDF 9/7 9/7
wavelet, COW seems to avoids a great part of the image
information relative to thermal noise and textural grain.
Wavelet choice:
Wavelets are used to decompose signal information, depending on their characteristics and topology. Here we need to find
the best wavelet matching to the interesting information in our
IR images, that is, the structures contours. Vu and Chandler
[19] described a method to construct an image sharpness map
based on 2D wavelet decomposition. The implementation of
such a sharpness map could allow us to visually evaluate how
image information is decomposed by a wavelet.
To that end, we took three photos of the SOI with 5x,
20x and 50x magnifying lenses. For each image, we built a
sharpness map considering the following: two orthogonal and
two bi-orthogonal wavelets, respectively:
B. Our IQA method
In [19], Vu and Chandler also proposed a wavelet-based
algorithm for estimating both the global and local image sharpness (Fast Index SHarpness - FISH). Their IQA comes from
the log-energy calculation of a three-level separable Discrete
Wavelet Transform (DWT) using CDF 9/7 9/7 wavelet. The
choice of this wavelet can be justified by its effectiveness to
match textures, and thus to extract a large range of data. The
energy evaluation brings information about the overall data
values. The sharper an image is, the more its DWT contains
high frequencies; and the more there are sparse frequencies,
the higher the energy is. Here the sharpest image could be
either a well focused image, or an unfocused one, because
each contains sparse high frequencies.
We generalize a DWT analysis as follows:
1) 2D DWT of grayscale image using a given wavelet, on N
levels. Let XYn denote the DWT subbands at each level
n, where XY is either LH, HH or HL DWT subbands.
2) Given a statistic function F , FXYn is F computed on
each subband XY at each decomposition level n.
3) For each decomposition level n, Fn is the pounding sum
of each FXYn :
1. Haar wavelet [20], since the image can be interpreted as
binary: reflexive versus non-reflexive materials, structures
versus background (I).
2. A Custom Orthogonal Wavelet (COW) built from a father
wavelet whose scaling coefficients are defined in (1) and
and from which filter bank coefficients are computed in
Tab. III. See its illustration in Tab. II.
φ=
1
k
1
k
1
k
1
k
(1)
Fn = α FLHn + β FHLn + γ FHHn
TABLE I: Haar wavelet illustration
Scaling function (φ)
TABLE III: COW filter bank
Wavelet function (ψ)
TABLE II: Custom Orthogonal Wavelet (COW) illustration
Scaling function φ
k
Analysis LPF
Analysis HPF
Synthesis LPF
Synthesis HPF
0
0.35355339
-0.35355339
0.35355339
0.35355339
1
0.35355339
0.35355339
0.35355339
-0.35355339
2
0.35355339
-0.35355339
0.35355339
0.35355339
3
0.35355339
0.35355339
0.35355339
-0.35355339
TABLE IV: CDF 9/7 9/7 wavelet illustration
Wavelet function ψ
Analysis scaling function φ
4
Analysis wavelet function ψ
(2)
TABLE V: Optimized CDF 9/7 9/7 wavelet illustration
Analysis scaling function φ
magnifying lenses (2.5x, 5x, 20x and 50x), at 80 FPS. From a
video, images are extracted so that each image correspond to
a lens position, and are evaluated with our proposed quality
metric. In this way, the highest-ranked image gives the best
matching focus according to our quality criterion.
Analysis wavelet function ψ
A. Results
We firstly compare our algorithm to FISH [19]. Fig. 8
presents the results.
The two approaches have different behaviors. With every
magnifying lens, our algorithm effectively catch the SOI focus
position (the star in Fig. 8); however, we observe unexpected
behaviors around the focus point, which could correspond to
optical distortions. The FISH algorithm catch the SOI focus
position, but is largely more sensitive to distortions from
the optics, the camera and other non-relevant information.
This reflects the importance of the textural overcoming the
structural information. This sensitivity bias the result, which
is inoperable for 2.5x and 5x magnifying lenses. These results
show that our method is adapted and robust to our specifics.
In order to better describe the image characteristics, we
compute other statistics on the same scheme as our algorithm:
entropy, Auto-Correlation (AC), skewness and kurtosis. See
Fig. 9.
Considering the interpretation of each statistic applied to
images, these measures reveal a complex image information.
Dis-focused or not, images are saturated: in natural scenes,
dis-focused images are considered as blurred. In this case
the image data are smoothed and contains little information.
In our study the textural part of the data is strong, and
light distortions implied by the optical system add unexpected
features to the images. Then, even using the DWT with our
COW, spatial information is too noisy to be interpreted without
preprocessing. The SD measure has a greater abstraction level,
enough to remain a reliable quality criterion.
where α, β and γ are coefficients with α + β + γ = 1.
4) The total Ftot of the DWT is the sum of each Fn weighted
as follows:
N
X
2N −n Fn
(3)
Ftot =
n=1
Our algorithm follows these four steps, with the specifics
listed below:
• According to the study in section IV-A, COW is used to
match the relevant image information.
• We consider the Standard Deviation (SD) (4), reflecting
the image data variability or diversity, since we cannot
rely on statistics based on spatial arrangement information, or on its quantity/quality estimation (see section
III-B).
v
u
C
u1 X
(4)
(SXYn (c) − µ)2
SDXYn = t
C c=1
•
where µ is the mean of the considered subband, and c is
the coefficient number of this subband from 1 to C.
Only the HL and LH subbands (5) are used: horizontal
and vertical components are kept, leaving the diagonal
information held by HH (see section III-B).
SDLH + SDHL
(5)
2
The weighting to obtain SDtot is reversed compared to
Ftot described below:
SDn =
•
SDtot =
N
X
2n SDn
B. Runtime
As studied in [23] and [24], the computational complexity
of DWT using filter banks implementation is O(n). Our algorithm threat the wavelets decomposition linearly per subband.
Since only two subbands per level are considered, and for a
three-level DWT, the global complexity is around:
n
n
n
(7)
O(n) + 2O( ) + 2O( ) + 2O( ) ≈ O(n)
2
4
8
(6)
n=1
For each decomposition level, some information are
dropped due to the sub-sampling. Then, in the highest
level, the root image information remains.
With this IQA we are then able to rank images by their
estimated quality. Our AF method is based on this criterion:
the best ranked image in a video corresponds to the focused
image, from which the focus position can be estimated (see
next section).
We implemented our architecture in Python language, on an
64-bit PC with an Intel Core i5-6200U at 2.30GHz and 8Gb
of RAM. The camera resolution is 320x256 pixels. Additional
packages are used for matrix operations (NumPy) and for
DWT (PyWavelet) using filter banks.
For a video containing 1500 images, our implementation run
in around 3.37 seconds. Then, an image quality is computed
in 2.24 ms. In our application, this time is not disturbing: the
focus research can be reduced if a previous AF is already
done. This solution may considerably reduce video length and
so the image stack.
V. E XPERIMENT AND DISCUSSION
Considering a linear movement given to the optical system
toward the OOI, we take a video. The initial position is the
farthest position from the OOI, the last is the nearest possible
(the physical limits). We make such an acquisition for each
5
TABLE VI: Sharpness map in function of magnifying factor and wavelet in use
Sharpness map depending on the wavelet used
Magnifying
lens
Original view
Haar
COW
CDF 9/7 9/7
OCDF 9/7
5x
20x
50x
Focus measure
Image index
Image index
(d)
Focus measure
(c)
Focus measure
(b)
Focus measure
(a)
Image index
Image index
Fig. 8: Graphs of focus measures of images extracted from videos, based on FISH (red) and on ours (black), for different
magnifying factors: 2.5x (a), 5x (b), 20x (c) and 50x (d); the stars mark the real SOI focus positions.
Other approaches may be adopted to study our IR images, as
steerable pyramid transforms, or any image-based approach,
but they require some preprocessing. Note that data such as
the diagonal details of the DWT decomposition stay untapped,
and may also be used for IR images denoising. Despite the
optical system limitations arising from the uncooled camera,
the proposed approach does not need any image enhancement;
its time efficiency makes our autofocus method usable for onthe-fly characterizations.
C. Evaluation of silicon thickness
As said in Section I, two adjustments are necessary before
a laser shot: the 3D positioning, and the laser power. The
laser power setting depends on the silicon thickness on the
SOI. A simple way to measure the silicon thickness consists
in subtracting the positions of the SOI (1) and the silicon
surface (2). The SOI position is available with the proposed
algorithm. Changing our AF criterion may permit to focus on
(2) instead of (1). For example, in the case of contaminated
silicon surface, the sharpness could be a good criterion.
R EFERENCES
VI. C ONCLUSION
[1] Z. Wang, A. C. Bovik, and L. Lu, “Why is image quality assessment so
difficult?” in IEEE International Conference on Acoustics Speech and
Signal Processing. IEEE, may 2002.
[2] W. S. Streijl R and Hands D, “Mean opinion score (MOS) revisited:
methods and applications, limitations and alternatives.” Multimedia
Systems, 2016.
[3] H. Sheikh and A. Bovik, “Image information and visual quality,” IEEE
Transactions on Image Processing, vol. 15, no. 2, pp. 430–444, feb 2006.
[4] V. Kamble and K. M. Bhurchandi, “No-reference image quality assessment algorithms: A survey,” Optik - International Journal for Light and
Electron Optics, vol. 126, no. 11-12, pp. 1090–1097, jun 2015.
We proposed a robust passive autofocus algorithm. Our
quality criterion allows us to efficiently focus on the conductive track surface inside an integrated circuit, independently
of the magnifying lens. This proposal responds to a need for
accuracy and reproducibility improvement in the context of
the security characterization of integrated circuits using laser
shots. The proposed algorithm is based on wavelet decomposition analysis, using a wavelet designed for our specifics.
6
Focus measure
[12]
(a)
[13]
[14]
Focus measure
[15]
(b)
[16]
[17]
Focus measure
[18]
(c)
[19]
[20]
Focus measure
[21]
(d)
[22]
[23]
Image index
Image index
[24]
Fig. 9: Graphs of statistics measures of images extracted
from videos, following our method, for different magnifying
factors: 2.5x (a), 5x (b), 20x (c) and 50x (d); the AC, kurtosis,
entropy, AG and SD are in green, magenta, blue, red and black,
respectively. The entire graphs are on the left, whereas zoomed
in views around the correct focus index are on the right; the
stars mark the real SOI focus positions.
for automatic detection of mycobacterium tuberculosis,” Cytometry Part
A, vol. 81A, no. 3, pp. 213–221, jan 2012.
S. Podlech, “Autofocus by bayes spectral entropy applied to optical
microscopy,” Microscopy and Microanalysis, vol. 22, no. 01, pp. 199–
207, jan 2016.
X. Zhang, C. Jia, and K. Xie, “Evaluation of autofocus algorithm for
automatic dectection of caenorhabditis elegans lipid droplets,” Progress
in Biochemistry and Biophysics (PBB), vol. 43, pp. 167–175, 2016, 11.
M. A. Green and M. J. Keevers, “Optical properties of intrinsic silicon
at 300 k,” Progress in Photovoltaics: Research and Applications, vol. 3,
no. 3, pp. 189–192, 1995.
M. A. Green, “Self-consistent optical parameters of intrinsic silicon at
300 k including temperature coefficients,” Solar Energy Materials and
Solar Cells, vol. 92, no. 11, pp. 1305–1310, 2008.
E. Simoncelli and W. Freeman, “The steerable pyramid: a flexible
architecture for multi-scale derivative computation,” in Proceedings.,
International Conference on Image Processing, vol. 3. IEEE Comput.
Soc. Press, 1995, pp. 444 – 447.
P. Pad, V. Uhlmann, and M. Unser, “VOW: Variance-optimal wavelets
for the steerable pyramid,” in 2014 IEEE International Conference on
Image Processing (ICIP). IEEE, oct 2014, pp. 2973–2977.
P. Hill, A. Achim, M. E. Al-Mualla, and D. Bull, “Contrast sensitivity of
the wavelet, dual tree complex wavelet, curvelet, and steerable pyramid
transforms,” IEEE Transactions on Image Processing, vol. 25, no. 6, pp.
2739–2751, jun 2016.
P. V. Vu and D. M. Chandler, “A fast wavelet-based algorithm for global
and local image sharpness estimation,” IEEE Signal Processing Letters,
vol. 19, no. 7, pp. 423–426, jul 2012.
A. Haar, “Zur theorie der orthogonalen funktionensysteme,” Mathematische Annalen, vol. 69, no. 3, pp. 331–371, 1910.
A. Cohen, I. Daubechies, and J.-C. Feauveau, “Biorthogonal bases of
compactly supported wavelets,” Communications on Pure and Applied
Mathematics, vol. 45, no. 5, pp. 485–560, jun 1992.
S. Zhang, G. Yang, Z. Cheng, and H. van de Wetering, “A novel 9/7
wavelet filter banks for texture image coding,” International Journal of
Advanced Research in Artificial Intelligence, vol. 1, no. 6, 2012.
O. Rioul and M. Vetterli, “Wavelets and signal processing,” IEEE signal
processing magazine, vol. 8, no. 4, pp. 14–38, Oct 1991.
K. K. Shukla and A. K. Tiwari, Efficient algorithms for discrete wavelet
transform: with applications to denoising and fuzzy inference systems.
Springer Science & Business Media, 2013.
G LOSSARY
AC Auto-Correlation. 2, 5
AF Autofocus. 1, 2, 3, 5
AG Average Gradient. 2
CDF 9/7 Cohen-Daubechies-Fauveau 9/7. 4
COW Custom Orthogonal Wavelet. 4, 5, 4
[5] H. Kim, S. Kang, H. Kang, Y. Cho, N. Park, and J. Kim, “Automatic
focus control for assembly alignment in a lens module process,” in 2009
IEEE International Symposium on Assembly and Manufacturing. IEEE,
nov 2009.
[6] V. Kumar and P. Gupta, “Importance of statistical measures in digital
image processing,” International Journal of emerging technology and
advanced engineering, vol. 2, no. 8, pp. 56–62, Aug. 2012.
[7] J. F. Schlag, A. C. Sanderson, C. P. Neuman, and F. C. Wimberly,
“Implementation of automatic focusing algorithms for a computer vision
system with camera control,” CARNEGIE-MELLON UNIV PITTSBURGH PA ROBOTICS INST, Tech. Rep., 1983.
[8] E. Krotkov, “Focusing,” International Journal of Computer Vision,
vol. 1, no. 3, pp. 223–237, oct 1988.
[9] X. Zhang, H. Wu, and Y. Ma, “A new auto-focus measure based on
medium frequency discrete cosine transform filtering and discrete cosine
transform,” Applied and Computational Harmonic Analysis, vol. 40,
no. 2, pp. 430–437, mar 2016.
[10] X. Xu, Y. Wang, J. Tang, X. Zhang, and X. Liu, “Robust automatic
focus algorithm for low contrast images using a new contrast measure,”
Sensors, vol. 11, no. 12, pp. 8281–8294, aug 2011.
[11] J. M. Mateos-Pérez, R. Redondo, R. Nava, J. C. Valdiviezo, G. Cristóbal,
B. Escalante-Ramírez, M. J. Ruiz-Serrano, J. Pascau, and M. Desco,
“Comparative evaluation of autofocus algorithms for a real-time system
DWT Discrete Wavelet Transform. 1, 4, 5, 6
FR Full-Reference. 2
IC Integrated Circuit. 1, 2, 3
IQA Image Quality Assessment. 1, 2, 3, 4, 5
IR Infrared. 1, 2, 3, 4, 6
NR No-Reference. 2
OOI Object Of Interest. 1, 2, 5
RR Reduced-Reference. 2, 3
SD Standard Deviation. 2, 5
SOI Surface Of Interest. 1, 2, 3, 4, 5
SWIR Short Waves Infrared. 2
7