Opportunities in vision and information processing

 

 

Voice-recognition for vision testing

This PhD project represents collaboration with the Life Sciences faculty in which a new method of vision testing which supports a more efficient man-machine interface compared to current methods is being developed. The objective of the PhD project is to develop a reliable speech recognition system in support of this new testing method.

Contact: Dr D W Armitage |  d.armitage@manchester.ac.uk

Position/velocity estimation using image sequences

This image processing project is concerned with methods to determine the position of an aircraftrelative to a terrain model (a digital terrain map) using an imaging sensor. The main interest is in infrared images in the 3-5 and 8-12 micron bands. Current methods for determining aircraft position (GPS/INS) are satisfactory in terms of grid co-ordinates but are rather unsatisfactory in terms of height. The aim of the project will be to implement and evaluate a new algorithm for position/velocity estimation. Our group has considerable expertise in this area. The project will be software-based using a UNIX workstation.

Contact: Dr John Oakley |  john.oakley@manchester.ac.uk

Enhancement of images in poor visibility conditions

The aim of this project is to extend a recent method for image enhancement. The interest is in improving the effective visibility in outdoor scenes in poor conditions, such as haze and fog. Significant progress has already been made by the group in this area, resulting in important publications and patents. The methods involve exploiting physical models for loss of contrast due to light scattering. This project will extend the scheme to important new applications such as night scenes and underwater inspection. The work will involve algorithm design, programming in C/C++ and possibly some real-time programming using a modern DSP device.

Contact: Dr John Oakley |  john.oakley@manchester.ac.uk

Diffractive optics

A new technique for the design of imaging optics using only diffractive components has been recently invented at UMIST. The range of applications for these new devices is very wide. This project is concerned with optimising the design of optical assemblies. This will involve the design of simulation software in MATLAB using a high performance Silicon Graphics workstation. Fabrication of devices will be done in collaboration with industry. It would be a good project for a student with an interest/background in modern optics(holograms etc)..

Contact: Dr John Oakley |  john.oakley@manchester.ac.uk

Data visualisation

Data visualisation has become important in many domains such as computer science, bioinformatics and management. It involves projection of often high dimensional data vetcors onto a visible two dimensional space. Searching for a suitable data projection method has always been an integral objective of multivariate data analysis and pattern recognition. Such as a method should enable us to observe and detect underlying data distribution, patterns and structures. Good visualisation techniques will not only provide an in-depth view of the data but can also reveal the underlying functions such as gene functionals and variations. Common data projection methods include PCA, Sammon mapping and the self-organising maps (SOMs). The project will further research in this area, in association with analysis of biological data sets.

Contact: H. Yin | hujun.yin@manchester.ac.uk

Ensemble learning schemes

Among various learning algorithms, it is hard and unfair to judge them individually on a single application. A statistical or neural model might perform well on one case but could fail on another no matter how sophisticated it is. Early mixture-of-expert and committee methods average various estimators to provide a less divergent solution. Methods such as "Bootstrapping" and "Bagging" improve the learning ability for cases of limited data. Recent ensemble methods attempt to boost "weak" learning algorithms into "strong" ones so improving the accuracy of any learning algorithms. The method involves training different estimators on dynamic selected/or learnt distributions. The project is to investigate and experiment with various ensemble schemes and the optimal theory behind them.

Contact: H. Yin | hujun.yin@manchester.ac.uk

Data mining: Automated associative search engines

Most search engines are simple text-string matching, though limited fuzziness has been brought in recently. Ideal search engines should be associative, error tolerant and feature based. Association requires each item to be related to items of the same and or similar meaning or functionality often in an ordered list or chart. Fuzzy matching provides error tolerance and imprecision for the user inputs. Search beyond simple textual characters can be based on features or descriptions or even functionality of the required object, of which an image is a typical example. Such extended search functions can be achieved using neural/fuzzy techniques and content-based indexing method. The entire search engine would be expected to be built with Java applets and applications.

Contact: H. Yin | hujun.yin@manchester.ac.uk

▲ Up to the top