Vision, Graphics and Visualisation

A study on the relationship between the world, images of it and their representation and comparison

One often has to represent a large number of views of an object, be it, e.g., the environment of a robot [1], an archaeological artifact or a database of faces [2]. An approach is to build such representations without extracting high-level information about the objects from images of them, such as geometry, but rather to use sub-symbolic methods: the appearance-based approach. In particular image manifolds in the image space can be built [3]. This project will explore problems such as understanding the relationship between the object, image formation (in particular colour spaces) and how this potentially large amount of data can be represented in an efficient way.

Proposed Supervisors: Dr Fred Labrosse

References:

[1] M. Neal and F. Labrosse, Rotation-invariant appearance based maps for robot navigation using an artificial immune network algorithm, Proceedings of the Congress on Evolutionary Computation, pp. 863-870, 2004.

[2] M. Bichsel and A.P. Pentland, Human face recognition and the face image set's topology, CVGIP: Image Understanding, 59 (2), pp. 254-261, 1994.

[3] J.B. Tenenbaum, Mapping a manifold of perceptual observations, Advances in Neural Information Processing Systems, 10, 1998.

Contrast based texture segmentation

This is a collaborative project between the Department of Computer Science, Aberystwyth University (Dr Reyer Zwiggelaar and PhD student), the School of Computing Sciences, University of East Anglia (Dr Mark Fisher), Institute of Geography and Environmental Science, Aberystwyth University (Dr Richard Lucas), and the Breast Imaging Department, Norfolk and Norwich University Hospital (Dr Erika Denton).

Texture is one of the least understood areas in computer vision. Although no generic texture model has emerged so far a number of problem specific approaches have been developed successfully [1,2]. More recently, approaches have been investigated which aim to automatically determine a feature vector to be used for segmentation purposes [3,4] or provide a more fundamental approach to texture segmentation [5-7].

The main aim of medical image segmentation is to divide an image into a set homogenous regions (see [8] for standard techniques) or generic structures [9]. In general to provide robust segmentation results the feature that drives the region growing process should contain more than just grey-level information.

The main research aims within the proposed project are: a) to develop segmentation approaches using a combination of grey-level and texture information, b) to develop appropriate distance metrics to compare two (or higher) dimensional information, c) to investigate how the segmentation techniques can be used to discriminate between vegetation types in satellite images, and d) to investigate how the segmentation techniques can be used to estimate dense mammographic tissue distribution.

Proposed Supervisors: Dr Reyer Zwiggelaar.

References

[1] M.W. Haralick. Statistical and structural approaches to texture. Proceedings of the IEEE, 67(5):786-804, 1979.

[2] T.R. Reed and J.M.H. Dubuf. A review of recent texture segmentation and feature-extraction techniques. Computer Vision, Graphics and Image Processing, 57(3):359-372, 1993.

[3] C.C. Reyes-Aldasoro and A. Bhalerao. Volumetric texture description and discrimant feature selection for MRI. Lecture Notes in Computer Science, 2732:282-293, 2003.

[4] R. Zwiggelaar. Texture based segmentation: automatic selection of co-occurrence matrices. In 17th IEEE International Conference onPattern Recognition, pages 588-591, 2004.

[5] M. Varma and A. Zisserman. Classifying images of materials: achieving viewpoint and illumination independence. Lecture Notes in Computer Science, 2352:255-271, 2002.

[6] M. Chantler and L. Van Gool. Special issue on "texture analysis and synthesis". International Journal of Computer Vision, 62:5, 2005.

[7] M. Pietikainen. Image analysis with local binary patterns. Lecture Notes in Computer Science, 3540:115-118, 2005.

[8] M. Sonka, V. Hlavac, and R. Boyle. Image Processing, Analysis and Machine Vision. Chapman and Hall Publishing, 1993.

[9] R. Zwiggelaar, S.M. Astley, C.J. Taylor and C.R.M. Boggis, Linear structures in mammographic images: detection and classification, IEEE Transaction on Medical Imaging 23 (9), pp. 1077-1086 (2004)

Detection and staging of prostate cancer from MRI

This is a collaborative project between the Aberystwyth University (Dr Reyer Zwiggelaar and PhD student, Department of Computer Science) and the Norfolk and Norwich University Hospital (Dr Stuart Williams, Radiology Department).

The main research aim within the proposed project is to develop software approaches to assist medical interpretation of MRI which is known to be difficult and prone to inter-observer variation. To achieve this we want to develop an anatomical tissue/structure segmentation approach. In combination with the previously developed volumetric prostate boundary and extra-capsular extension estimation techniques this will provide the potential for a complete prostate cancer detection and staging tool. The techniques that will be developed for this software tool will be able to solve the complexity of the prostate anatomy (in terms of orientation and distribution) and as such are directly translational to develop tools to assist radiotherapy planning (e.g. brachytherapy distribution within the prostate gland).

This work will be building upon the automatic segmentation and extra-capsular extension (ECE) staging of prostate cancer from standard MRI data [1,2]. To date we have developed an automatic 3D segmentation approach which provides robust volumetric surface estimation in line with expert radiologists annotations. In addition, we have provided an initial investigation into the staging of prostate cancer through the evaluation of ECE [4,5]. The final end-point of our research in the area of prostate cancer would be a staging/treatment tool, which shows the capability to: a) detect the various stages of cancer development, with the emphasis on stages II and III, b) visualisation of the prostate, c) visualisation of abnormal regions, which can also be used for treatment planning, and d) investigate the development of prostate cancer over time. It should be noted that this project covers aspects a) and c), and as such form major components towards our end-goals. Such a software based approach will assist medical interpretation and is expected to reduce inter-observer variation.

The research project will: a) develop a generic texture based segmentation methodology, b) investigate texture based tissue/structure segmentation within the prostate, c) investigate the detection of normal and abnormal regions within the prostate, and d) develop a software tool to assist medical interpretation, based on all current and developed detection/classification aspects.

Proposed Supervisors: Dr Reyer Zwiggelaar.

References

[1] R. Zwiggelaar, Y. Zhu, and S.Williams. Semi-automatic segmentation of the prostate. Lecture Notes in Computer Science, 2652:1108-1116, 2003.

[2] Y. Zhu, S. Williams, and R. Zwiggelaar. A hybrid ASM approach for sparse volumetric data segmentation. Pattern Recognition and Image Analysis, 15(2):346-349, 2005.

[3] R. Zwiggelaar, Y. Zhu, and S. Williams. Towards classification of prostate MRI. 8th Conference on Medical Image Understanding and Analysis, London, UK:204-207, 2004.

[4] Y. Zhu, S. Williams, M. Fisher, and R. Zwiggelaar. The use of grey-level profiles for detection of extracapsular extension of prostate cancer from MRI. 9th Conference on Medical Image Understanding and Analysis, Bristol, UK:215-218, 2005.

Laser Scanning Technologies and Their Applications

While a CCD digital camera simulates only a single human eye, outputting projective images, the latest laser scanning system (range camera) simulates with a point laser and a CCD digital camera two human eyes. Hence, it inherently creates a stereo vision system. Through processing some measures of interest, it outputs range images that depict 3D information of an object surface. As a result, it has an advantage over single digital cameras since it can directly obtain the depth information from the object of interest. Also, the recovery of depth from projective images is often sensitive to noise due to point sampling on the object surface, various illumination conditions, surface orientations, and different reflective properties. Thus, the laser scanning techniques provide with easy data acquisition a great impetus to research on 3D shape modelling, processing and understanding.

Using 3D (structured) point clouds captured, this project will investigate how to accurately and automatically register different point clouds [2,4,5], how to estimate the camera motion parameters between different viewpoints [2,4,5] at which the data were captured, how to integrate the registered data [1] for the reconstruction of a full model without redundant data of the object and/or the environment of interest, how to efficiently render the 3D data through more compact representation of the geometry of the object and/or environment of interest, and how to recognise objects and compare the different free form surfaces in the 3D data [3].

Such research tasks have numerous applications: data registration is pre-requisite of 3D model reconstruction for computer aided geometric design, object modelling, computer graphics, computer animation and game, medical diagnosis and therapy planning, and robot path planning and obstacle avoidance; camera motion estimation is useful for robot localization and path planning and medical operation planning, efficient rendering of 3D data is useful for data visualization and understanding, computer animation and game, and special effects in film and advertisements, and object recognition and comparison is useful for object identification and segregation, 3D image search over internet, bioinformatics, and protein segment alignment and search.

To this end, this project will investigate a wealth of different techniques: mean field annealing, doubly stochastic matrix, SoftAssign, generalised entropy, entropy maximization, uncertain reasoning, Bayesian inference, replicator dynamics, Evolutionarily Stable Strategies, Nash Equilibrium, thin plate spline, radial basis functions, approximation theory, machine learning, clustering, relaxation labelling, Markov chain, Lyapunov function, Fast Fourier transform, wavelet transform, etc. Ultimately, the project is attempting to develop novel accurate and efficient techniques for the analysis and understanding of 3D world depicted in the form of range images captured using the latest laser scanning systems. It is believed that modelling iterative image understanding process as an evolutionary system will open a novel avenue to the research and development of 3D laser scanning technologies.

This overall project will be divided into a number of small scale projects suitable for PhD study within a period of three year/s.

Proposed Supervisor: Dr Yonghuai Liu.

References:

[1] Hong Zhou, Yonghuai Liu, and Longzhuang Li. Incremental Mesh-based Integration of Registered Range Images: Robust to Registration Error and Scanning Noise. Proceedings of the Seventh Asian Conference on Computer Vision, Hyderabad, India during January 14-16, 2006, LNCS, vol. 3851, pp. 958-968.

[2] Yonghuai Liu. Automatic 3d free form shape matching using the graduated assignment algorithm. Pattern Recognition, vol. 38, no. 10, pp. 1615-1631, 2005.

[3] Yonghuai Liu, Guoqiang Fei, Baogang Wei, Longzhuang Li. 3D Free Form Surface Matching Based on Orientation Difference Length Distribution. 2005 IEEE International Conference on Acoustics, Speech, and Signal Processing, March 18-23, 2005, Philadelphia, PA, USA, vol. II, pp. 773-776.

[4] Yonghuai Liu. Improving ICP with Easy Implementation for Free Form Surface Matching. Pattern Recognition, vol. 37, no. 2, pp. 211-226, 2004.

[5] Marcos A. Rodrigues and Yonghuai Liu. On the Representation of Rigid Body Transformations for Accurate Registration of Free Form Shapes. Robotics and Autonomous Systems, vol. 39, no. 1, pp. 37-52, 2002.