In this work we address the problem of optimal sensor placement for a given region and task. An important
issue in designing sensor arrays is the appropriate placement of the sensors such that they achieve a predefined
goal. There are many problems that could be considered in the placement of multiple sensors. In this work
we focus on the four problems identified by Hörster and Lienhart. To solve these problems, we propose an
algorithm based on Direct Search, which is able to approach the global optimal solution within reasonable time
and memory consumption. The algorithm is experimentally evaluated and the results are presented on two real
floorplans. The experimental results show that our DS algorithm is able to improve the results given by the
most performing heuristic introduced in. The algorithm is then extended to work also on continuous solution
spaces, and 3D problems.
The choice of an appropriate illumination design is one of the most important steps in creating successful machine
vision systems for automated inspection tasks. In a popular technique, multiple inspection images are captured
under angular-varying illumination directions over the hemisphere, which yields a set of images referred to as
illumination series. However, most existing approaches are restricted in that they use rather simple patterns
like point- or sector-shaped illumination patterns on the hemisphere. In this paper, we present an illumination
technique which reduces the effort for capturing inspection images for each reflectance feature by using linear
combinations of basis light patterns over the hemisphere as feature-specific illumination patterns. The key idea
is to encode linear functions for feature extraction as angular-dependent illumination patterns, and thereby to
compute linear features from the scene's reflectance field directly in the optical domain. In the experimental
part, we evaluate the proposed illumination technique on the problem of optical material type classification of
printed circuit boards (PCBs).
Uneven illumination is a common problem in real optical systems for machine vision applications, and it contributes
significant errors when using phase-shifting algorithms (PSA) to reconstruct the surface of a moving
object. Here, we propose an illumination-reflectivity-focus (IRF) model to characterize this uneven illumination
effect on phase-measuring profilometry. With this model, we separate the illumination factor effectively, and
then formulate the phase reconstruction as an optimization problem. To simplify the optimization process, we
calibrate the uneven illumination distribution beforehand, and then use the calibrated illumination information
during surface profilometry. After calibration, the degrees of freedom are reduced. Accordingly, we develop
a novel illumination-invariant phase-shifting algorithm (II-PSA) to reconstruct the surface of a moving object
under an uneven illumination environment. Experimental results show that the proposed algorithm can improve
the reconstruction quality both visually and numerically. Therefore, using this IRF model and the corresponding
II-PSA, not only can we handle uneven illumination in a real optical system with a large field of view (FOV),
but we also develop a robust and efficient method for reconstructing the surface of a moving object.
Unconstrained environments with variable ambient illumination and changes of head pose are still challenging
for many face recognition systems. To recognize a person independent of pose, we separate shape from texture
information using an active appearance model. We do not directly use the texture information from the
active appearance model for recognition. Instead we extract local texture features from a shape and pose free
representation of facial images. We use a smooth warp function to transform the images. We compensate also
the shape information for head pose changes and fuse the results of separate classiers for shape features and
local texture features. We analyze the inuence of the individual contributions of shape and texture information
on the recognition performance. We show that fusing shape and texture information can boost the recognition
performance in an access control scenario.
Automatic industrial surface inspection methodology based on Magnetic Particle Inspection is developed from image
acquisition to defect classification. First the acquisition system is optimized, then tubular material images are acquired,
reconstructed then stored. The characteristics of the crack-like defects with respect to its geometric model and curvature
are used as a priori knowledge for mathematical morphology and linear filtering. After the segmentation and binarization
of the image, vast amount of defect candidates exist. Finally classification is performed with decision tree learning
algorithm due to its robustness and speed. The parameters for mathematical morphology, linear filtering and
classification are analyzed and optimized with Design Of Experiments based on Taguchi approach. The most significant
parameters obtained may be analyzed and tuned further. Experiments are performed on tubular materials and evaluated
by its accuracy and robustness by comparing ground truth and processed images. The result is promising with 97 % True
Positive and only 0.01 % False Positive rate on the testing set.
More recently, research on enhancing the situational awareness of pilots, especially in poor visibility flight conditions,
gains more and more interests. Since pilots may not be able to spot the runway clearly in poor visibility
conditions, such as fog, smoke, haze or dim lighting conditions, aviation landing problem can occur due to the
(unexpected) presence of objects on the runway. Complicated and trivial instruments, switches, bottoms, plus
sudden happenings are enough for the pilots to take care of during landing approach. Therefore, an automatic
hazard detection approach that combines non-linear Multi-scale Retinex (MSR) image enhancement, edge detection
with basic edge pattern analysis, and image analysis is investigated. The effect of applying the enhancement
method is to make the image of the runway almost independent from the poor atmospheric conditions. The
following smart edge detection process extracts edge information, which can also reduce the storing space, the
comparison and retrieval time, and the effect of sensor noise. After analyzing the features existing in the edge
differences occurring in the runway area by digital image processing techniques, the existing potential hazard will
be localized and labeled. Experimental results show that the proposed approach is effective in runway hazard
detection in poor visibility conditions.
Materials scientists make use of image processing tools more and more as technology advances and the data
volume that needs to be analyzed increases. We propose a method to optically measure magnetic eld induced
strain (MFIS) as well as twin boundary movement in Ni2MnGa single crystal shape memory alloys to facilitate
spatially resolved tracking of deformation. Current magneto-mechanical experiments used to measure MFIS
can measure strain only in one direction and do not provide information about the movement of individual
twin boundaries. A sequence of images captured from a high resolution camera is analyzed by a boundary
detection algorithm to provide strain data in multiple directions. Subsequent motion detection and Hough feature
extraction provide quantitative information about the location and movement of active twin boundaries.
Human actions annotation in videos has received an increase attention from the scientific community these last
years mainly due to its large implication in many computer vision applications. The current leading paradigm to
perform human actions annotation is based on local features. Local features robust to geometric transformations
and occlusion are extracted from a video and aggregated to obtain a global video signature. However, current
aggregation schemes such as Bag-of-Words or spatio-temporal grids have no or limited information about the
local features spatio-temporal localization in videos. It has been shown that local features localization can be
hepful for detecting a concept or an action. In this work we improve on the aggregation step by embedding local
features spatio-temporal information in the final video representation by introducing a point process model. We
propose an event recognition system involving two main steps: (1) local features extraction based on robust point
trajectories, and (2) a global action representation capturing the spatio-temporal context information through
an innovative point process clustering. A point process provides indeed a well-defined formalism to characterize
local features localization along with their interactions information. Results are evaluated on the HOllywood in
Human Action (HOHA) dataset showing an improvement over the state-of-art.
Thermal face recognition becomes an active research direction in human identification because it does not rely on
illumination condition. Face detection and eyeglasses detection are necessary steps prior to face recognition using
thermal images. Infrared light cannot go through glasses and thus glasses will appear as dark areas in a thermal image.
One possible solution is to detect eyeglasses and to exclude the eyeglasses areas before face matching. In thermal face
detection, a projection profile analysis algorithm is proposed, where region growing and morphology operations are used
to segment the body of a subject; then the derivatives of two projections (horizontal and vertical) are calculated and
analyzed to locate a minimal rectangle of containing the face area. Of course, the searching region of a pair of eyeglasses
is within the detected face area. The eyeglasses detection algorithm should produce either a binary mask if eyeglasses
present, or an empty set if no eyeglasses at all. In the proposed eyeglasses detection algorithm, block processing, region
growing, and priori knowledge (i.e., low mean and variance within glasses areas, the shapes and locations of eyeglasses)
are employed. The results of face detection and eyeglasses detection are quantitatively measured and analyzed using the
manually defined ground truths (for both face and eyeglasses). Our experimental results shown that the proposed face
detection and eyeglasses detection algorithms performed very well in contrast with the predefined ground truths.
This paper presents a new approach to optical material stress analysis, which eliminates the need to apply
a random dot pattern to the surface of the sample being tested. A multi-resolution hierarchical sub-division
is implemented, with a consistent polynomial decimation applied at each layer of the tree. The degree of
decimation must be selected depending on the nature of the structure of the surface of the sample being At each
layer the individual patches are registered using a modified normalized phase correlation, whereby the Fourier
basis functions are projected onto the orthogonal complement of a low degree Gram polynomial basis. This
reduces the effect of the Gibbs error on the local registration. The registration positions are then subjected to
a regularization via an entropy weighted tensor-polynomial approximation. The Gibbs polynomial basis is used
for the tensor product, since they are orthonormal and model the continuous deformation associated with an
elastic deformation. The stability of the proposed method is demonstrated in real measurements and the results
with and without the application of the random pattern are compared.
Reduction of herbicide spraying is an important key to environmentally and economically improve weed
management. To achieve this, remote sensors such as imaging systems are commonly used to detect weed plants.
We developed spatial algorithms that detect the crop rows to discriminate crop from weeds. These algorithms
have been thoroughly tested and provide robust and accurate results without learning process but their detection
is limited to inter-row areas. Crop/Weed discrimination using spectral information is able to detect intra-row
weeds but generally needs a prior learning process.
We propose a method based on spatial and spectral information to enhance the discrimination and overcome the
limitations of both algorithms. The classification from the spatial algorithm is used to build the training set for
the spectral discrimination method. With this approach we are able to improve the range of weed detection in the
entire field (inter and intra-row). To test the efficiency of these algorithms, a relevant database of virtual images
issued from SimAField model has been used and combined to LOPEX93 spectral database.
The developed method based is evaluated and compared with the initial method in this paper and shows an
important enhancement from 86% of weed detection to more than 95%.
Quality control of clams considers the detection of foreign objects like shell pieces, sand and even parasites.
Particularly, Mulinia edulis clams are susceptible to have a parasite infection caused by the isopoda Edotea
magellanica, which represents a serious commercial problem commonly addressed by manual inspection. In this
work a machine vision system capable of automatically detect the parasite using a clam image is presented. The
parasite visualization inside the clam is achieved by an optoelectronic imaging system based on an transillumination
technique. Furthermore, automatic parasite detection in the clam's image is accomplished by a pattern
recognition system designed to quantitatively describe parasite candidate zones. The extracted features are used
to predict the parasite presence by means of a binary decision tree classifier. A real sample dataset of more than
155000 patterns of parasite candidate zones was generated using 190 shell-off cooked clams from the Chilean
south pacific coasts. This data collection was used to train a test the classifier using cross-validation. Primary
results have shown a mean parasite detection rate of 85% and a mean total correct classification of 87%, which
represent a substantive improvement to the existing solutions.
We develop a methodology for automatic in-line flaw detection in industrial woven fabrics. Where state of the
art detection algorithms apply texture analysis methods to operate on low-resolved (~200 ppi) image data, we
describe here a process flow to segment single yarns in high-resolved (~1000 ppi) textile images. Four yarn
shape features are extracted, allowing a precise detection and measurement of defects. The degree of precision
reached allows a classification of detected defects according to their nature, providing an innovation in the field of
automatic fabric flaw detection. The design has been carried out to meet real time requirements and face adverse
conditions caused by loom vibrations and dirt. The entire process flow is discussed followed by an evaluation
using a database with real-life industrial fabric images. This work pertains to the construction of an on-loom
defect detection system to be used in manufacturing practice.
In order to enhance turboshaft lifespan and increase thermal efficiency, aeronautical manufacturers have to optimize the
temperature of engine components in operation. Dedicated combustion tests are undertaken and specific techniques are
developed to measure surface temperatures. Thermal paints have been used for several years, associated with skilled
operator observations, as a valuable means to get peak temperature profiles. This article describes major advances in the
analysis process based on color to temperature supervised classification and digitization of the outer shape of the
components in order to get 3D temperature maps.
A non-contact scanner enables to acquire both a color image and a 3D mesh of the component. The color image is
processed with a classification algorithm to give a temperature image. Different colorimetric distances are tested to
compare each pixel to the database and find the best matching temperature, which is then associated to a node of the 3D
mesh. The relevance of the method is to increase temperature resolution and robustness and to allow more reliable
comparisons between numerical simulation and bench test measurement. This system is currently implemented in the
engine development process at Turbomeca.
We suggest a local approximation of perceptual color differences in
a device dependent color space, e.g. an RGB space. The approximation is efficiently
computed from measuring Euclidean color distance in the device dependent
color space combined with an associate memory data structure. Established
measures of color difference are considered. Results for small perceptual color
differences in a color inspection setup are given.
We have modified the Fuzzy C-Means algorithm for an application related to segmentation of hyperspectral images.
Classical fuzzy c-means algorithm uses Euclidean distance for computing sample membership to each cluster. We have
introduced a different distance metric, Spectral Similarity Value (SSV), in order to have a more convenient similarity
measure for reflectance information. SSV distance metric considers both magnitude difference (by the use of Euclidean
distance) and spectral shape (by the use of Pearson correlation). Experiments confirmed that the introduction of this
metric improves the quality of hyperspectral image segmentation, creating spectrally more dense clusters and increasing
the number of correctly classified pixels.
In this paper we present an algorithm for the recognition of 1D barcodes using the Hough transform,
which is highly robust regarding the typical degraded image. The algorithm addresses various typical
image distortions, such as inhomogeneous illumination, reflections, damaged barcode or blurriness
etc. Other problems arise from recognizing low quality printing (low contrast or poor ink
receptivity). Traditional approaches are unable to provide a fast solution for handling such complex
and mixed noise factors. A multi-level method offers a better approach to best manage competing
constraints of complex noise and fast decode. At the lowest level, images are processed in gray
scale. At the middle level, the image is transformed into the Hough domain. At the top level, global
results, including missing information, is processed within a global context including domain
heuristics as well as OCR. The three levels work closely together by passing information up and
down between levels.
This paper proposes an algorithm for the detection of pillars or posts in the video captured by a single camera
implemented on the fore side of a room mirror in a car. The main purpose of this algorithm is to complement
the weakness of current ultrasonic parking assist system, which does not well find the exact position of pillars or
does not recognize narrow posts. The proposed algorithm is consisted of three steps: straight line detection, line
tracking, and the estimation of 3D position of pillars. In the first step, the strong lines are found by the Hough
transform. Second step is the combination of detection and tracking, and the third is the calculation of 3D
position of the line by the analysis of trajectory of relative positions and the parameters of camera. Experiments
on synthetic and real images show that the proposed method successfully locates and tracks the position of
pillars, which helps the ultrasonic system to correctly locate the edges of pillars. It is believed that the proposed
algorithm can also be employed as a basic element for vision based autonomous driving system.
In this paper, a novel support vector machine (SVM) tree is proposed for gesture recognition from the silhouette
images. A skeleton based strategy is adopted to extract the features from a video sequence representing any
human gesture. In our binary tree implementation of SVM, the number of binary classifiers required is reduced
since, instead of grouping different classes together in order to train a global classifier, we select two classes for
training at every node of the tree and use probability theory to classify the remaining points based on their
similarities and differences to the two classes used for training. This process is carried on, randomly selecting
two classes for training at a node, thus creating two child nodes and subsequently assigning the classes to the
nodes derived. In the classification phase, we start out at the root node. At each node of the tree, a binary
decision is made regarding the assignment of the input data point to either of the group represented by the left
and right sub-tree of the node which may contain multiple classes. This is repeated recursively downward until
we reach a leaf node that represents the class to which the input data point belonging. Finally, the proposed
framework is tested on various data sets to check its efficiency. Encouraging results are achieved in terms of
Small devices used in our day life are constructed with powerful architectures that can be used for industrial
applications when requiring portability and communication facilities. We present in this paper an example of
the use of an embedded system, the Zeus epic 520 single board computer, for defect detection in textiles using
image processing. We implement the Haar wavelet transform using the embedded visual C++ 4.0 compiler
for Windows CE 5. The algorithm was tested for defect detection using images of fabrics with five types of
defects. An average of 95% in terms of correct defect detection was obtained, achieving a similar performance
than using processors with float point arithmetic calculations.
In this paper, we propose an orthophotoplan segmentation method based on watershed algorithm combined with
an efficient region merging strategy for roof detection. The preliminary segmentation is obtained by the watershed
algorithm with an optimal couple of colorimetric invariant/color gradient optimized for the application. The use
of the appropriate couple of invariant/gradient permits to limit illumination changes (shadows, brightness, etc)
affecting the images. Even if the watershed based results are good, the images are over-segmented. That is why,
a region merging procedure is proposed. This procedure uses a merging criteria based on 2D modeling of roof
ridges and region features adapted to the orthophotoplan particularities. The proposed strategy is evaluated on
100 real roof images with a ground truth image segmentation in order to demonstrate the effectiveness and the
reliability of the proposed approach.