ÄLDRE EXAMENSARBETEN VID AVDELNINGEN FÖR MATEMATIK, LTH

Segmentation and Classification of Immature Blood Cells using Digital Images of Bone Marrow Smears

• Student: Björn Nordin
• Advisor: Anders Heyden, Johan Håkansson
• In cooperation with: CellaVision AB
• Date finished: 98-12-10

Multiscale image analysis and image enhanchment based on partial differential equations

• Student: Henrik Malm
• Advisor: Gunnar Sparr, Stefan Diehl
• Date finished: 980422
• Abstract: Inspired by the use of partial differential equations to describe physical phenomena, the examination of their use in image processing has been a fast growing research topic during the last decade. Their main applications are restoration, segmentation and the generation of multiscale image representations, so-called scale-spaces. It is this last topic that is focused on in this paper, but it will be shown that many scale-space methods also have very interesting enhancement properties. Other properties that are interesting for a scale-space generating equation are invariancy under different kinds of transformations. A review of different methods will be given, including the basic theory, their scale-space properties and some numerical aspects. Visual examples of the use of each method will also be given. Finally, the results of some experiments, that investigates if the invariancy properties of the continuous theories can be carried over to discrete implementations of the methods, are presented.

Computation of geometric structure from an uncalibrated camera on a robot arm

• Student: Peter Lindström
• Advisor: Gunnar Sparr, Anders Heyden, Kalle Åström.
• Date finished: 98-04-21
• Abstract: Today, we try to make robots and other machines more flexible and more human. One way is to equip a robot with a system that functions like the human vision. By taking images with one camera moving between images, we get depth perception.\\[3mm] The purpose of this work is to see how far it is possible to perform a hand-eye calibration on an industrial robot with an uncalibrated camera attached to it. Algorithms for reconstruction of a 3D-object from a sequence of 2D-images using the concepts of shape and depth have been developed. With the use of a known focalpoint the reconstruction can be made quite exact.\\[3mm] The algorithms have been tried on a series of randomly generated cases with the results analyzed. A real image-sequence has also been used for testing.

Controlling a Ball and Plate Process Using Computer Vision.

• Student: Martin Rentsch
• Advisor: Bo Bernhardsson (regler), Kalle Åström, Anders Robertsson (regler).
• Date finished: 1998-03-13
• Abstract: This thesis describes the setup of a Ball and Plate process. The aim is to make the ball roll along a predefined trajectory by tilting the plate appropriately. The position of the ball is sensed by means of a video camera and the plate is actuated by an industrial robot. An analog video camera acquires pictures to a host computer, where they are digitized. The visual information is processed to give as a result the location of the ball on the plate. Control signals for the two degrees of freedom of the ball are derived and transmitted to the robot system. The exact matehmatical model of the process is linearized and discretized for the design of an onbserver based digital controller. Position control of the robot is implemented on a detached computer system.

Computer vision

• Student: Jesper Göransson
• Advisor: Andrew Zisserman (Oxford), Gunnar Sparr
• Date finished:

Automated Identification of Fingerprints

• Student: Jerker Bergenek, Linus Wiebe
• Advisor: Anders Heyden, Fredrik Kahl, Mårten Öbrink
• In cooperation with: Identifier AB
• Date finished: 98-04-03

Controlling the Ball and Beam Process Using a Video Camera.

• Student: Fredrik Emilsson
• Advisor: Bo Bernhardsson (regler), Johan Eker (regler), Kalle Åström
• Date finished: 1997-11-27
• Web site: Controlling the Ball and Beam Process Using a Video Camera.
• Abstract: In some control situations it is difficult to use ordinary sensors. The rapid development in camera technology and frame grabbers has made it possible to use visual information in the control loop. The goal of this masters thesis was to control "the ball and beam" process using these ideas. An ordinary video camera was placed in front of the process. The camera was connected to a computer through a frame grabber. The images of the process were then analyzed in real time. The angle of the beam and the position of the ball were calculated and used as input in a traditional control algorithm. A sampling frequency of 8 Hz was obtained in the final implementation. The thesis contains a thorough description of the image analysis, the control algorithm, the hardware as well as a performance analysis of the system.

Segmentation and classification of Papsmear using image analysis and neural networks

• Student: Johan Håkansson
• Advisor: Kalle Åström, Christer Fåhreus.
• In cooperation with: CellaVision
• Date finished: 1997-11-14
• Abstract:The Papanicolaou or Pap smear that test women for cervical cancer has saved tens of thousands of lives since it was introduced in the early 1940s. Since about 100 millions of Pap smears are prepared and judged manually each year, it is desirable to develop an automated screening system.This work examines the possibility of using image analysis and artificial neural networks (ANN) in the screening process. It focuses on finding and segmenting cell nucleuses in the image as the shape and texture of these are a strong indicator of a beginning neoplasm. A segmentation algorithm using local thresholding and a procedure for edge tracing of the nucleus are described. Three different features are extracted from the segmented nucleuses and used for training the ANN.
• Keywords: Pap smear, Artificial neural networks, Image analysis

Automatic Feature Detection, Tracking and Iterative Reconstruction from Image Streams

• Student: Anders Vegh, Peter Hallberg
• Advisor: Anders Heyden, Kalle Åström.
• Date finished: 97-10-17
• Abstract: A modern robot is mobile and able to perform a wide spectrum of tasks autonomously. To enable the unhindered mobility for a robot you have to equip it with some sort of camera based vision system. The depth perception can be obtained by using either two cameras, or comparing the images in a stream of images taken by one camera. To make this comparing of images possible the position of the camera has to be known at all times. Therefore it is of great importance to reconstruct the motion of the camera. We develop suitable algorithms for detecting features and tracking them through a stream of images, making this reconstruction possible. The detection of certain features in the image is based on Harris' corner detector resulting in a function we choose to call the corner response function, indicating the strength of the certain features. The detection is followed up by the tracking, which is accomplished by determining several parameters such as correlation, position and the corner response function. This rough tracking is followed up by a refined method using the epipolar constraint. Finally an iterative reconstruction method is used to mark those paths, that still may not be correct.

  upp

Studies on Computer Vision applied to Robotics

• Student: Björn Johansson
• Advisor: Gunnar Sparr, Anders Heyden, Kalle Åström.
• In cooperation with:
• Date finished: 1997-03-06.
• Abstract: For a robot to act autonomously, it could be provided with a camera based vision system which deliver video sequences. The purpose of this work is to investigate methods for reconstruction of a 3D scene using such sequences of 2D images. For this to be possible, algorithms for finding features in an image and tracking them, have been developed. The algorithms and the methods for reconstruction have been tried on a sequence of images, taken by a camera attached to a robot, with the results analyzed.

Segmentation and classification of human blood cells

• Student: Daniel Elvin, E-91
• Advisor: Anders Heyden, Gunnar Sparr
• In cooperation with: Christer Fåhraeus / Med-AI Europe AB
• Date finished: 1997-02-21.
• Abstract: Differential blood cell counting is used in clinical diagnosis because various physiological and pathological conditions result in deviations from the normal blood cell distribution. To achieve a faster and more accurate result than of the manual counting performed today, an automatic differential counting system is wanted. In this work a three step leukocyte (white blood cell) segmentation algorithm using image processing and pattern recognition techniques is proposed, consisting of: unsupervised optimal histogram thresholding level finding, locate leukocyte nucleus, and leukocyte cytoplasm detection. Also a six class differentiation method is proposed using an artificial neural network. Furthermore approximations of the erythrocyte (red blood cell) area and shape distributions are presented. This includes a simple segmentation algorithm and an area and perimeter estimation. The images in this work where all captured from blood smears by a 3-chip colour CCD-camera mounted on a transmission microscope with a 20 x magnification lens.

Fractal Image Compression using Triangulations

• Student: John Svensson, F91
• Advisor: Sven Spanne
• Date finished: Nov 1996
• Abstract: Fractal image compression has recently been advanced as an alternative to the classical transform based image compression methods. This work describes an approach to fractal image compression using a partitioning scheme based on triangles and quadrilaterals. The aim of this approach is to reduce the number of blocks by grouping neighbouring triangles into quadrilaterals while keeping the better adaptability to the image inherent in triangular decompositions. The merging of triangles into quadrilaterals allows a reduction of the number of contractive transformations composing the fractal transform, and thus a higher degree of compression without loss of visual quality.

Mahler Measure

• Student: Johan Roos, F92
• Advisor: Gert Almkvist
• Date finished: 1996-06-07.
• Abstract: The Mahler measure M(f) for a polynomial f(x) with integer coefficients is defined by D H Lehmer conjectured in 1932 tha if f is not cyclotomic then there is a e>0 such that M(f) > 1 + e. This conjecture is still open. The Mahler measure has connections to Salem and Pisot numbers, polylogarithms, constructions of large primes and entropy in ergodic theory.

A Method for Image Reconstruction in AFM Using a Tip Characterizer

• Student: Tobias Melin, F91
• Advisor: Gunter Grossmann, Lars Montelius, Gunnar Sparr
• In cooperation with: FTF
• Date finished: 1996-05-06.
• Abstract: AFM images are generated by scanning a tip across a sample. The image is not equal to the sample as the tip, because of its finite extension, cannot follow all the sample features. The image is thus rather a mixture of tip and sample features. A method based on mathematical morphology is used that eliminates features due to the tip shape in order to get an image that is more similar to the actual sample. Equality is only possible to achieve for points of the sample that were in contact with the tip during the image generation. Other parts of the restored sample are assigned values that are only an upper bound to the true sample surface. A binary image is computed that indicates these regions. The morphological method used requires knowledge of the tip geometry, which is measured with the AFM itself. This is possible due to the symmetry between tip and sample; their roles are interchangeable. Having a known sample, a tip characterizer, we can extract the tip shape from an AFM image using the same morphological method. In particular micro-spheres are shown to be suitable tip characterizers.

Optical Character Recognition using Neural Nets

• Student: Mats Emilsson, D91
• Advisor: Gunnar Sparr
• Date finished: 1996
• Abstract: Methods for Optical Character Recognition, OCR, using feedforward neu\-ral nets with back-propagation, are developed. OCR refers to the transform of machine-printed document pixel images to ASCII codes. A character database, with data from multiple fonts in different angles and typefaces, implementing a large character set, has been constructed. A big neural net consisting of 16 x 16=256 inputs, 150 hidden nodes and 116 outputs, has succesfully been trained with noisy images, using the back-propagation algorithm. For each document image to be decoded, the characters are isolated using a segmentation algorithm, and then preprocessed and classified by the neural net.

Computerized Analysis of Myocardial Perfusion Images

• Student: Hannes Rahn, E90, och Peter Ridell, E89
• Advisor: Gunnar Sparr och Rickard Berthilsson
• In cooperation with: inst för Klinisk Fysiologi, LU
• Date finished: 1996-03-29
• Abstract: Coronary heart diseases affect millions of people and it is one of the leading causes of death in the western world today. This paper is about myocardial perfusion images and how to interpret them by the use of Artificial Neural Networks (ANNs). The main problem to overcome is to reduce the amount of data presented to the ANN without losing any important information. In this study, this is done by using image processing transforms such as the wavelet transform and the Hotelling transform. We have also developed a geometric method to shrink the dimensionality of the ANN's input space. The work is done in cooperation with the department of Clinical Physiology, Lars Edenbrandt and John Palmer.

Correction by means of Digital Image Processing of Distorted Depth Perception due to Hyperstereopsis

• Student: Jonas Kjellström, E91, och Petter Skerfving, D91
• Advisor: Gunnar Sparr och Anders Heyden
• In cooperation with: FFV, Arboga
• Date finished: 1995-10-23
• Abstract:

Interpretation of Polyhedral Images

• Student: Daniel Svedberg
• Advisor: Gunnar Sparr and Anders Heyden
• Date finished: 1995-09-28
• Abstract: In this paper an interpreting system using an image of a polyhedral scene as input is developed and described. Polyhedral scenes are built up by piecewise planar objects and constitute a common sort of scenes encountered in robot vision and reconstruction. Hence, there is a need to automatically interprete images of obtained from projections of polyhedral scenes. Two projective invariants, shape and depth, will be introduced. Using the line-drawing obtained from edge-detection of the image and the notions of shape and depth, the image may be interpreted and classified. Criteria for the image to be classified correctly and a method for adjusting for the influence of noise will be described. Following interpretation, a correct reconstruction of the three dimensional scene is possible.

Två industriella bildbehandlingsproblem: Realtidsföljning av Hörn i Bildsekvenser och Virkessortering med Bildbehandlingsmetoder.

• Student: Fredrik Kahl
• Advisor: Gunnar Sparr och Kalle Åström
• In cooperation with: Innovativ Vision AB, Linköping

Real time methods in automatic wood inspection

• Student: Henrik Leijonhufvud, E90
• Advisor: Gunnar Sparr
• In cooperation with: Innovativ Vision AB, Linköping
• Date finished: 1995-03-25
• Abstract: The use of neural networks on an industrial problem is studied. The industrial application involves parquet sorting by automatic inspection. Firstly, the neural networks are used to detect the structure differences between classes of blocks. Several different layouts are tested before a conclusion is made. Secondly, neural networks are tried on discoloration detection. The results are not satisfying, and an algorithmic approach is suggested. The approach involves a transformation of the colour space combined with local measurements. The algorithm uses relative colour parameters, and is shown to be robust when trials are made on a larger data set.

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
Examensarbeten -1998