PhD Project Proposals in Computer Vision and Digital Systems
David Pycock
School of Electronic and Electrical Engineering
Tel: +(0)121 414 4330
Fax.: +(0)121 414 4291
Email: D.Pycock@bham.ac.uk
 

1. Automation of Colposcopy Interpretation

Colposcopy is the first line of investigation in the early detection of Cervical Cancer. It involves staining the cervix and viewing the resulting appearance to identify atypical patterns of vascularisation, and distribution of endometrial cells. Normally colposcopy is performed in a hospital clinic. Experiments are in hand to evaluate the use of video images that could be collected by a trained nurse and assessed by a clinician on a separate occasion. Automated procedures for evaluating these images would mean that only the potentially abnormal cases would need to be referred back to a consultant. The automated measurements of suitable criteria requires innovation in the development of image interpretation procedures. The particular challenge will be to produce quantitative descriptors for features on a surface in 3-D to identify significant changes. Surface geometry can greatly alter appearance and the human visual system is very adept at interpreting structures on complex surfaces. In this case it is not practical to use projected light patterns or stereoscopic images. The proposal is to investigate the use of colour and particular shading to infer shape. Shape from shading is a well established technique but it has not been investigated using colour images to define shading. Colour should improve the performance of shape from shading by enabling the effects of changes in surface orientation and illumination to be more effectively isolated. Whilst it is not practical top use a pair of “calibrated” stereoscopic images an alternative line of investigation would be to combine data from images taken at positions with a small, ill-defined relative displacement of the view point. A third strategy would be to either use thermal images on their own or combined with an optical image. Both these strategies have the potential to make it easier to identify patterns of vascularisation. This research will be conducted in collaboration with Professor David Luesley at the Birmingham Women’s Hospital.
 

2. Stereoscopic Active Colour Vision

Stereoscopic vision relies on the verging of binocular views. Until now this has only been done using monochrome images. The use of colour should make it possible to more clearly and rapidly identify correspondence as the feature set provided by a colour scene is inherently richer and produces less ambiguity than do grey-level images. Whilst dark regions in a colour scene cannot provide any additional information they are at least as rich as a monochrome image. In addition there is evidence to show that use of colour should reduce dependence on ambient illumination. It is, in principle at least, possible to define invariant measures based on colour change, independent of intensity change. The greatest potential weakness in using colour for stereoscopic imaging is in the lower spatial resolution of colour cameras. However it may be possible to overcome this potential weakness by taking advantage of the improved robustness and density of colour features. This project will focus on defining invariant colour features or image mappings starting with an investigation of colour spaces and lead on to an investigation of the robustness and accuracy of colour-based stereoscopic imaging.
 

3. Heterogeneous Architectures for Integrating Numeric and Symbolic Processing

This project continues research already under way. To date effort has focused on an investigation of performance estimation models and developing a programmers model for a heterogeneous system based on the Contract-Net protocol. The proposed architecture differs significantly from the many “Network of Workstation Systems” that are currently under development in the use of an heterogeneous set of processing modules. The style of task allocation in which the distribution of tasks to processors depends on the outcome of earlier computations is an important, novel feature of this strategy. Most computational procedures that involve a combination of numeric and symbolic data analysis are non-deterministic in nature and this approach is designed to manage such computation which is not well handled in traditional schemes of data analysis. Our experience in attempting to model this system has confirmed our suspicion that existing modelling strategies are not well suited to analysing the behaviour of this system. Software simulations have, however, shown that this strategy has considerable potential to accelerate the speed of computation. We have demonstrated better than linear speed up in particular contexts. These experiments have been performed using a homogeneous set of processing modules. Some of the processing modules were handicapped to simulate heterogeneity. We are now seeking to develop the hardware elements for this computational architecture and build a prototype system. The major issue that will need to be addressed in this next phase is the development of a message interpretation module. This will be a specialised processor for performing pattern matching but also able to evaluate the complexity of algorithms. The present programmers model implements the major facets of the Contract Net protocol in Glenda, the GNU version of Linda. This programmers model uses the shared memory management provided by Glenda to implement task announcement, bidding, the negotiation of “contracts”, the transfer of data and the transfer of computational procedures. A number of issues in this task allocation or scheduling process still need to be addressed. These include acquiring and using knowledge of the capabilities of processing modules, the control of forward planning; when to stop task planning and to start execution.
 

4. Model-Based Multi-Modal Image Registration

This research is concerned with automating the registration of multi-modal medical images. Present registration techniques assume that the corresponding features in multi-modal images have a similar appearance. This is a manifestly unsound assumption. Studies have shown that significant errors of registration occur when such assumptions are made in non-rigid schemes of registration. The development of new functional imaging modalities means that it is rapidly becoming more important to be able to register anatomic and functional images. The approach that we propose is to use automatically generated models to define the correct registration of images from differing modalities. The key stages in this process are to identify equivalent topological structures and landmarks in each modality. The basis of our strategy is the tenet that topology will define equivalent features in each modality more reliably than procedures based on the direct detection of grey-level ridges. When such topological landmarks have been identified we will be able to compute measures to describe key structures in the image around each landmark that will be used to define a model of correspondence. As a first step towards the extraction of topological structures we have investigated algorithms for computing multi-scale medial axes or cores of an image. This research has led to an improved algorithm which eliminates many of the anomalies generated by previous methods and produces an axis that is localised with greater precision and improved noise immunity. There are four key steps in the process described above that need further study. These are: 1. Building a topological description. 2. The interpretation of what is likely to be a complex topology. 3. Building models that describe corresponding landmarks. 4. Using these concepts for multi-modal medical image registration. It is anticipated that a research student would concentrate on at the most two of these topics.
 

5. Reasoning by Analogy for Image Interpretation

For a number of years there has been considerable research into model-based schemes for interpreting complex images that has resulted in a significant improvement in the ability to analyse selected types of image. However the effort needed to develop these techniques for a particular class of application remains a major obstacle to the widespread adoption of image processing. In recent EPSRC funded research under my supervision robust and adaptable strategies for interpreting complex scenes have been demonstrated that achieve a level of performance comparable with that of a human observer2. The conceptual objective is to create an image analysis system that can be trained, using examples, how to select an appropriate algorithm. Techniques to recognise both when and how algorithms can be adapted from one domain to another will be required. This will involve the development of algorithms, for example, to recognise both when the difference between images in one domain and another involve a simple transformation and to determine an appropriate transformation. Other algorithms will be needed to identify which of the many alternative algorithms for each stage of image processing are appropriate. It will involve the development of measures to compare computational procedures and the development of ways to describe the relationship between a hierarchical description of the procedures. This is needed so that appropriate alternative image interpretation strategies can be selected. The proposed research extends the concept of image processing systems based on hierarchical definitions of procedures. A key issue will be to develop strategies for reasoning about action (i.e. image processing actions). This strategy extends the concepts developed by Owen1 for Reasoning by Analogy. The intention is that the analysis system should be able to reason at both the knowledge and the meta-knowledge levels.

1. Owen S, Analogy for Automated Reasoning, Academic Press, 1990.
2. Zhou, P and Pycock D, Robust Statistical Model-Based Cell Image Interpretation, in Proc. BMVC’95, pp 117-127.
 

6. Model-Based Interpretation of Stereoscopic Images

Recent research in the School has led to the development of a technique for identifying the 3-D structure of textured surfaces using stereoscopic images and projected light patterns. A hierarchical active contour method is used to identify the path of the projected stripe in each image and a 3-D reconstruction formed. The further research proposed will extend this work by using an active contour to identify the disparity between the stereoscopic image pair.

Grace AE, Pycock D, Multi-resolution active contour models in textured stereo images, in Proc. BMVC’96, pp 575-584.