Segmentation of Magnetic Resonance Images of the Knee using Three-dimensional Active Shape Models
Student: Klas Josephson, F00
Advisor: Anders Ericsson, Johan Karlsson, Kalle Åström, Magnus Tägil, (Ortopedi)
In cooperation with: Avdelningen för Ortopedi, Lunds universitetssjukhus
Date finished: 2004-12-16
Abstract: In this thesis a fully automated segmentation system for the bones of the knee in Magnetic Resonance Images are presented. Several datasets were segmented manually. The resulting structures are first represented by unorganized point clouds. Level set methods were then used to fit surfaces to these point clouds. The iterated closest point method was then used to establish point correspondences between different surfaces.Both surfaces and correspondences were then used to build a three dimensional statistical shape model of the bones in the knee. The resulting model is then used to automatically segment the knee in subsequent datasets through a method called Active Shape Models.The result of the segmentation is promising, but the quality of the segmentation is dependent on the initial guess.The segmentation is useful for establishing the position of other parts of the knee, for further segmentation, feature extraction and interpretation of the images.The algorithms have also been tested on SPECT images of the brain with better results.
Spectral Fingerprint Matching
Student: Magnus Wennergren, F00
Advisor: Magnus Oskarsson, Björn Nordin (Precise Biometrics AB)
In cooperation with: Precise Biometrics AB
Date finished: 2004-11-09
Abstract: This master thesis explores a new way to match fingerprint images. Today is well establishedmethods use minutiae-points or bitmap correlation for matching. The new method exploits another way of fingerprint representation. The fingerprint image is divided into a number of small cells, which all are represented by three spectral parameters: direction, wavelength and phase. These parameters give a fine representation of the fingerprint texture. Algorithms for both the extraction of the parameters and the matching of these are developed in the master thesis. The parameters are matched each type by itself, and the scores of the three different matching algorithms are also weighted together to achieve maximum matching performance.The results are very promising. They show that this new way of fingerprint matchingmight in the future be a good competitor to the established methods.
Decision Support System for Lund Cancer using PET/CT Images
Student: David Jacobsson, F99 and Fredrik Olofsson, F00
Advisor: Anders Ericsson, Johan Karlsson, Kalle Åström, Andreas Järund (Weaidu in Europé AB)
In cooperation with: Weaidu in Europé AB
Date finished: 2004-10-13
Abstract:Interpretation of medical images is often difficult and time consuming, even for experienced physicians. The aid of image analysis and machine learning can make this process easier.
This thesis presents a fully automatic decision support system for lung cancer diagnostics with images from combined Positron Emission Tomography and Computed Tomography. Algorithms for segmentation of the lung region, localisation of suspected tumours (hot spots), detection of the heart, noise reduction and feature extraction are evaluated. A graphical user interface is also developed.
A database of 99 patients with and without lung cancer diagnosis is used to train and test two learning systems, Support Vector Machines, SVM, and Artificial Neural Networks, ANN. The best result of the classification is a class rate of 94%, obtained using SVM with six features and two classes, lung cancer and no lung cancer.
Detection and Recognition of Text in Images.
Student: Johan Windmark, F99
Advisor: Rikard Berthilsson
Date finished: 2004-10-08
Abstract: In this thesis we will study automatic Document Image Analysis (DIA) under the assumption that images are obtained using an ordinary handheld camera.Using images taken by a handheld camera introduces several new problems to document image analysis compared to using images obtained from flat bed scanning. Examples of such problems are uneven lighting, perspective effects and low image quality. In this thesis, these problems and algorithms to overcome them are described in detail. Emphasis is put on image binarization and rectification. We also study how to detect the position of groups of characters in the document. Furthermore, the developed methods have successfully been implemented as a software application on a cell phone equipped with camera.The software application automatically finds and interprets all the numbers that are necessary to make a payment from an invoice and sends the result over the air to a remote server. It could be incorporated into a complete system for paying bills from cell phone, without entering numbers manually, by connecting the cell phone to a bank's Internet services. Performance tests on invoices show good results, provided that the invoices are sufficiently illuminated.The methods described here can also be applied to reading for example ISBN numbers on books and telephone numbers. These are all examples of pieces of information that are designed to be used by machines and not by humans. Unfortunately, in many situationssuch numbers need to be manipulated and entered manually, which may be tedious and potentially also error prone.
Student: Johan Klang, E00
Advisor: Magnus Oskarsson, Fredrik Kahl
Date finished: 2004-09-29
Abstract: The overall goal of this work is to be able to obtain camera position and orientation from a single picture in a limited and modelled environment. With the aid of pictures the application will be able to be used in different scenarios where the need for navigation and/orposition determination is important. One thinkable application is to use a mobile phone with built-in camera and have the cellular display pointing out the position on the street of the city the user is currently in.The primary goal is to determine position and horizontal orientation with mapping of window pictures and matching of these against a database containing 2D-information of a limited area with buildings.
Automatic Vessel Detection and Three-dimensional Vessel Reconstruction from Biplane Angiograms
Student: Karl Johansson, D99
Advisor: Rikard Berthilsson, Kalle Åström, Henrik Jönsson (Thoraxkirurgiska kliniken, Lunds Universitetssjukhus) ,Anders Lundin (Röntgenavdelningen, Lunds Universitetssjukhus)
In cooperation with: Thoraxkirurgiska kliniken, Lunds Universitetssjukhus and Röntgenavdelningen, Lunds Universitetssjukhus
Date finished: 2004-09-24
Abstract: This thesis examines the possibilities of automatically detecting and tracking blood vessels from biplane angiograms and reconstructing them in a three-dimensional space. A three-dimensional model would facilitate the work of the physicians and serve as a complement to the manual analysis of biplane angiograms. It could, for instance, be helpful in diagnosing cardiovascular diseases and work as a complement to the planning of surgical operations. A method that utilizes the directional information of the Hessian to locate the coronary artery is proposed. The maximal change in contrast along the tangential direction of the vessel, and optimization techniques are used to achieve sub-pixel precision. The extracted and labelled tree is tracked in each sequence of images, by looking for points that are maximally correlated with each other within windows surrounding each point. Only points that correlate most strongly with each other in "both" directions are chosen. Three-dimensional reconstruction is performed by determining the epipolar geometry with a linear least-square method. To improve scene coherence, the error between the 2D points and the projected 3D reconstructed points is minimized by using a classic bundle adjustment technique. The results of the combined detection and tracking algorithm are promising and the artery tree is extracted with high accuracy in the angiograms. However, the 3D reconstruction still suffers from uncertainties concerning the geometry of the scene and the metric correctness of the reconstructed vessel segments, thus further improvements are necessary in this area.
Image Association Support for Mobile Phones
Student: Magnus Hölén, D 99
Advisor: Handledare: Rikard Berthilsson
Date finished: 2004-08-27
Abstract: The goal of this master's thesis is to develop methods for image association support for mobile phones. The problem can be formulated as, given two images of an object, conclude whether they represent the same object or not. We assume that both images are taken under approximately the same conditions. Furthermore, since the target platform is a mobile telephone, with limited computational power and memory, the algorithms that should solve the problem, must be designed with this fact in mind.Several methods for association support are studied in this thesis, all of which are based on finding landmarks in images followed by matching of landmarks. The accuracy of the program varied with the choice of object to recognize. A repeating pattern in the object would stop any chance of finding a good match, while an object with distinct and well localized landmarks almost always gave good matches.An example where association support could be used is pin code elimination, where the user takes a picture of him self to unlock his camera equipped mobile phone. Other applications includes having the phone automatically call a company if the user photographs the company logo, or safe codeword saving, where photographing a login screen or electronic door lock gives the code needed to pass.
Monocular depth from occluding edges
Student: Stefan Karlsson, F98
Advisor: Kalle Åström
Date finished: 2004-08-20
Abstract: This report describes some experiments with algorithms that determine depth from occluding edges as well as performing completion. First Gaussian diffusion is applied to the problem, with continuous depth effects as result. After that several algorithms are developed for modal and amodal completion from T-junctions and line ends, and the resulting contours are sorted in depth. The problem of completion is approached by extracting features and building graphs from these, resulting in a number of hypotheses. Bézier splines are used as connecting elements between features. Further an attention based mechanism is experimented with as a way to solve the emerging search problem. The various stimuli presented are hand drawn contour images.
Shape Analysis of MRI Data of the Brain and Cranial Structures of Schizophrenic and Healthy Subjects
Student: Nils Maltesson, D99 och Karin Wickström , F99
Advisor: Anders Eriksson, Johan Karlsson, Kalle Åström
Date finished: 2004-08-19
Abstract: Schizophrenia is a mental disease whose causes have up until today remained mostly unknown. One theory is that the disease may be caused by physical changes in the brain, possibly already during the pre-natal development of the fetus. An investigation of the origins of schizophrenia would yield a deeper understanding of the disease. Therefore, the aim of this thesis is to find areas in the brain and surrounding skull that are affected by the disease, and how these are changed when comparing to healthy controls. Landmarks representing different areas in the brain and skull have been extracted from MRimages. The landmarks are aligned using Procrustes analysis, making shape comparisons between subjects possible. These are investigated using different shape analysis and statistical techniques such as Likelihood Ratio Tests, Principal Component Analysis and Partial Least Squares combined with t-tests. By using these methods statistically significant landmarks that differ between the two groups have been found. Despite small groups of subjects, the results are promising and show that some key landmarks are different when comparing schizophrenic and healthy subjects. This indicates that the theory of pre-nataldevelopment is interesting for further studies.
Annotation Support for Image Databases
Student: Anna Persson, F99
Advisor: Rikard Berthilsson, Kalle Åström
Date Finished: 2004-06-09
Abstract: During the last years digital cameras have become increasingly popular. This together with the fact that photos can be taken at almost no cost, leads to people taking many photos. These pictures can be stored on a computer, which after a while may lead to a huge library of pictures. With that many pictures, it might be difficult to find the one searched for, like for example a certain person, or a picture of a beach. In this thesis, we study the problem of grouping pictures together according to what faces they contain. Methods for this can be implemented as software on different platforms, such as PC's, digital cameras and mobile phones. In a possible future user scenario the user marks a face in one picture, and the software presents all pictures in the library containing that face. In order to compare and group pictures according their content, some pre-processing is required, e.g. band pass filtering and so called feathering. By using affine transformations of the pictures, it is possible to allow also for some deformations, such as different facial expressions, position and size of the face in the picture, without affecting the recognition. A series of different methods for preprocessing and classification is presented in this thesis.
Realtime vision applications for mobile phone platforms
Student: Pär Sarbäck, F99 and Anders Görtz, E99
Advisor: Rikard Berthilsson, Magnus Oskarsson, Lars Halling (Softhouse Consulting AB)
In cooperation with: Softhouse Consulting AB
Date Finished: 2004-05-27
Abstract: In this master's thesis three different methods are evaluated for realtime vision applications, where the goal is to detect motion in image sequences. The first method is the KLT-tracker which tracks good features from frame to frame in image sequences. The second is a differential technique for optical flow computation. The third and final method is a simplified version of optical flow using a differentiated Gaussian filter applied to image sequences for computing the time derivative. For evaluating these computer vision methods, two different kinds of realtime motion detection alarms are developed, where both types later on are implemented on a mobile phone using the Java 2 Platform, Micro Edition (J2ME). Finally the performance of the implemented algorithms on a mobile phone are validated showing good results, both regarding accuracy and low computational time.
Facedetection and tracking
Student: Andreas Persson, D99 and Martin Persson, D99
Advisor: Magnus Oskarsson, Håkan Ardö, Kalle Åström
Date Finished: 2004-04-28
Abstract: Today, surveillance with cameras is common at both private and public places. Traditionally these cameras have been monitored by a human, but thanks to technologys progress today its possible to have computers do the work. A camera can also be mounted at the entrance of a building and work as a replacement for a magnetic card and instead use an image of a persons face as a mean of identification. Optimally one would not even need to stop at the entrance since the camera detects, tracks and identifies occurring faces during their approach to the entrance. A robust real-time system is a good replacement for the human operator. The system described in this master thesis is designed to work with a computer and a digital video camera, either with the camera connected direct or through a network. The system is able to detect multiple faces and then track them while they remain in the cameras view, in an arbitrary environment. The system uses adaptive background algorithms, skin colour extraction, support vector machines and Kalman tracking. The system is able to work at a speed of 2 images per second. The system has shown good results during tests in an indoor environment, both as a doorway passing control and as general surveillance system overviewing a lobby. The bottleneck in the developed system is currently the speed of the system and the accuracy of the face detection. Suitable future works may be to implement parts of the system in hardware to increase the number of images per second that the system processes.
Physical Reflection: methods to improve inaccurate depth maps of objects observed by a 3D camera
Student: Astrid Wåhlin, E98
Advisor: Fredrik Kahl, Kalle Åström, Thomas Broomé (Interactive Institute, SMART)
In cooperation with: Interactive Institute, SMART
Date Finished: 2004-04-28
Abstract: Physical Reflection is a project based on the idea that a mechanical mat of nails should reflect an observed object's surface. A 3D camera serves as an observer, having the ability of calculating the depth to each point in the scene. However this ability is poor and the resulting depth maps are noisy and inaccurate. The main purpose of this thesis is to improve the quality of these depth maps. An idea to improve inaccurate depth maps of a certain class of object is to apply a statistical model to them. Faces are an interesting class and they are given special attention in this thesis. A well known method called the Appearance Model is tested and the facial depth maps are approximated by the model to resemble it.
Another method to improve inaccurate depth maps is to fill-in all surface holes of missing data. The advantage is that the method works on all types of objects. A number of different techniques capable of achieving this are tested in this thesis, for example, normalized convolution.
The results showed that the statistical model did not fulfill the requirements of Physical Reflection, but the filling-in surface holes techniques achieved its aim. The thesis demonstrates that a mechanical nail mat can give the impression of a physical reflection of a real-world object and the constellation can be the basis for an interesting art exhibit.
Shape Optimization for Incompressible Laminar Flows
Student: Carl Olsson, E01
Advisor: Magnus Fontes,Anders Hedenström (Animal Ecology)
In cooperation with: Animal Ecology, Lund University
Date Finished: 2004-04-23
Abstract: The problem to minimize the drag on a body that moves in an incompressible fluid is considered. We show that the difference between two solutions is small if the difference between the domains of the flow is small. This is then used to show that the energy functional describing the drag is continuous on a compact set, which implies that there exists a minima. An attempt to find the minima is then made by using steepest decent methods in MATLAB and FEMLAB.
Map Generation for Autonomous Vehicles
Student: Gustaf Brännberg, F98
Advisor: Henrik Stewénius, Kalle Åström
Date Finished: 2004-04-01
Abstract: Vision is useful for the autonomous navigation of vehicles. In this thesis the case of a vehicle equipped with multiple cameras with non-overlapping views is considered. The main topic of the thesis is the implementation and testing of algorithms for automatic map generation. Three rigidly mounted cameras are placed on a vehicle that follows an unknown trajectory in a room. Given the image sequences of these cameras, it is shown how to generate a map of the area and to reconstruct the trajectory of the vehicle. A point detector and a simple tracker was designed for the point generation. Structure and motion algorithms were used to get initial solutions to the problem. Clssical algorithms such as intersection, resection and bundle adjustment are extended to this new situation and used to refine the solution. The theory has been tested on real data with promising results, but more work is needed before the system can be used in practice.
Motion Segmentation, Object Tracking and Event Detection in Image Sequences
Student: Kalle Wittenmark, E98
Advisor: Håkan Ardö, Kalle Åström, Daniel Elvin, (Axis Communications AB)
In cooperation with: Axis Communications AB
Date Finished: 2004-02-17
Abstract: This thesis describes a system for motion segmentation, object tracking, and event detection in an office environment. Surveillance with cameras is common at, for instance, airports, highways, public areas, schools, ports and private buildings. There is a need for automatic systems that can analyse the data from these cameras. It is now possible to create an inexpensive robust automatic real time surveillance system due to fast computers and inexpensive network cameras. The system described in this thesis takes images with a network camera and the program runs on a standard desktop PC with Linux. The system is a prototype of an surveillance system. Moving objects are segmented from the background with a bakground segmentation algorithm. The movements of the objects are tracked based on position and color features and can handle interaction between objects, such as occlusion from stationary objects, and occlusion between moving objects. The tracks are presented in a movement flow image. Some predefined events, for instance, entering, exiting, and occlusion are detected and written into a log file. The log file together with the motion flow image describes the events that occur in the image sequence in a small data space. The system runs in real time with a frame rate of 3 frames/second and can be placed and accessed anywhere with an internet connection. The system is adaptive and robust to lighting changes and can run for a long time without restarting or changing of any parameters. The system has been tested in different office environments. The same system can sucessfully track objects and detect events in the office environments without changing any parameters. The system has also been tested in an outdoor environment over a parking area, but some improvements must be made to get the same performance as in the indoor environment. The system can be used as an indoor surveillence system, help for surveillance of several monitors at the same time, traffic analysis, and analysing common movement paths in public areas.
Kronecker Tensor Products, Differentiation Theory for Matrices and Applications
Student: Christian Sohl, F00
Advisor: Sergei Silvestrov
Date Finished: 2004-02-06
Abstract: The aim of this work is to present a theory of Kronecker tensor products, matrix differentiation and their interrelation and applications. The work consists primarily of three parts. The first part is concerned with the Kronecker tensor products and sums, while the second part deals with differentiation theory for matrices in general. In particular differentiation of scalar functions of a matrix with respect to the matrix and vector functions with respect to a vector or a matrix is presented. Further, matrix differentiation with respect to a matrix is introduced and its fundamental properties such as the product and chain rules are discussed. Some methods for evaluating the derivative of specific important matrix functions are also given. The third and last part consists of some applications of Kronecker tensor products and matrix derivatives, all of them in strong connection with the theory presented earlier in this work. Among the applications, Carleman linearization procedure and tensor invariance transformations of Heisenberg systems, can be found.
Segmentation of Velocity Encoded Magnetic Resonance Image Sequences of the Human Heart
Student: Ylva Aspenberg, F99
Advisor: Gunnar Sparr, Håkan Arheden (Biomedicinskt Centrum), Erik Bergvall (Biomedicinskt Centrum)
In cooperation with: Biomedicinskt Centrum
Date Finished: 2004-01-30
Abstract: This thesis develops a method for segmenting the heart muscle from the background in a velocity encoded magnetic resonance image sequence. These images contain information about both anatomy and velocity of the tissues in the body, and provide a unique possibility to examine the function of the heart. Unfortunately velocity encoded images often have weak contrast between muscle and blood, and that complicates the segmentation of the heart muscle. In this work we used a deformable model, a snake, to delineate the heart muscle. The deformable model needs so called attractor images, external forces that guide the model towards the boundaries in the image. We have calculated four new attractor images from the velocity data and we investigated their ability to guide a snake against the edges. By integrating the velocity field the trajectory of a particle can be followed during a heart cycle. This has been done in previous work and the particle trace data was used in the calculation of two of our attractor images. These images worked best and were found to be powerful attractor images especially for delineation of the inner heart wall.
Tillbaka till huvudsidan
Examensarbeten under 2009
Examensarbeten under 2008
Examensarbeten under 2007
Examensarbeten under 2006
Examensarbeten under 2005
Examensarbeten under 2004
Examensarbeten under 2003
Examensarbeten under 2002
Examensarbeten under 1999-2001