One Day Technical Meeting held on January 21 1998 at the British Institute of Radiology, 36 Portland Place, London.
Defence Evaluation & Research Agency,
This work addresses some of the difficulties that are involved in the automatic generation of image content descriptors within a specific domain of airborne reconnaissance. A number of separate avenues of research have been pursued and will be briefly overviewed. Ultimately we are interested in combining image and textual database concepts into a single system in which both can be used to reinforce the confidence in results of queries.
We initially considered the questions of combining several feature detectors to improve their performance and reliability on a wider variety of images. One of the difficulties that has existed in computer vision is that researchers have had a tendency to concentrate on developing single algorithms (eg edge and region detectors) for specific types of images. It is only recently that papers are appearing which consider the questions of combining detectors to improve their performance and reliability on a wider variety of images. This appears to us to be quite a natural process as a number of experts pooling their opinions should =91hopefully=92 provide a more accurate result. Examples given will include the detection of urban regions, roads and vehicles in airborne imagery as well as the location of driveable regions for an autonomous land vehicle.
The region descriptors once generated and represented in some suitable way could be used for the indexing, matching and searching as well as retrieval of data in an image database. Once a region is obtained then there is the issue of how to actually represent its shape. We have chosen to use a curvature scale space approach which provides a very concise representation for objects and more recently explored affine invariants.
Moving on from the idea of actually combining individual detectors we are now experimenting with the use of genetic algorithms for optimising firstly the choice of parameters for an image processing algorithm and secondly the choice of algorithms themselves.
The feature detectors used in this work currently mainly cover the detection of points, edges, ridges and regions. Region detection consists of Belief Networks, Fractal and Wavelet.
The main applications of this work will be in the area of image databases. The amount of imagery that is currently being collected is going to require some form of automatic (or at least semi-automatic) processing prior to insertion into an image database.
Some of this work has resulted from collaboration with the University of Oxford and the University of Surrey. In particular we wish to thank Zi-Yan Xie at Oxford for the Fractal and Wavelet research and Farzin Mokhtarian at Surrey for a copy of the curvature scale space software.
(c) British Crown Copyright 1998/DERA, Published with the permission of the Controller of Her Britannic Majesty's Stationary Office.
Earth Observation Sciences
Since January of 1997, EOS has been working with partners from leading arctic institutions on a project to demonstrate the feasibility of an ice routing tool for safer and more efficient ship transport in ice-covered waters. This paper reports results from investigations into the use of artificial intelligence technologies to generate high-resolution, operational-quality ice information for ship routing. The data sources include Radarsat synthetic apperture radar imagery, SSM/I passive microwave data, meteorological information, and contextual data such as basins and currents. The character of the problem is unusual, in that the primary data sources - Radarsat imagery - reveal that the distribution of ice is highly fractal. Also, object boundaries are often difficult to define, even by eye.
The techniques explored include fuzzy expert systems, growing neural gas, and others developed specifically for the problem. Preliminary results are presented.
G. Herries and T. Selige
GSF - National Research Centre for Environment and Health,
Institute of Biomathematics and Biometry
Layered feed forward Neural Networks (NN) have been found to have good generalisation properties and their use is becoming increasingly prevalent in the field of remote sensing. The traditional classification approach when using neural networks is for all land-cover classes to be classified with a single neural network. The authors have found this to be an inefficient method.
A new method is proposed, a modular approach, which involves training a neural network for each individual land-cover type to be classified. This increases the number of neural networks required, but reduces the complexity of each individual network and reduces the training time significantly. The individual training times/epochs can be reduced by several orders of magnitude, with the sum of the training times for all the modular networks being much lower than the single network approach. The accuracies produced by this modular approach tend to be higher than the traditional single neural network approach. Another significant benefit of the modular approach is that poorly classified classes (which have an individual neural network) can be easily re-trained, without effecting the other neural networks or performing a lengthy retraining exercise.
This paper applies these two approaches to optical airborne data at varying resolution. The advantages and disadvantages of each approach are presented. The area used for this work is a research farm in Bavaria, Germany, which comprises of a highly dynamic terrain with small field units. High resolution land-use maps and yield data have been produced for the research farm, using GPS equipment attached to crop harvesters. These maps are used to validate the results produced by the two approaches.
Environmental Science Department,
Natural Resources Institute
Martin Brown, Hugh Lewis, John Manslow, Mark Nixon, Adrian
Tatnall, Alberto Aguado
University of Southampton
FLIERS is a fourth framework EU project which is being investigated at the Universities of Southampon, Leicester and Thessaloniki (Greece) as well as the Joint Research Centre (Italy) and VTT (Finland). The aim of the project is to represent and quantify the uncertainty associated with developing automated land classification systems. The sample areas from which training and validation data is being collected include urban, semi-natural and large field and small field agriculture, and a key part of this process is being able to estimate the errors associated with the data collection and assimilation process, as this has a strong influence. The project is also developing appropriate visualisation tools for exploring the data and classification results, which includes extending traditional boolean-type displays to represent mixed class data and advanced VR systems which can display 3D scatter plots etc. in a more natural enviroment.
However, a central premise of the project is that a better understanding and representation of the different forms of uncertainty which are involved in the data collection and classifier design processes will improve the performance and reliability of such systems. This can be stated succinctly as: maximising the representation of uncertainty during the design process, in order to minimise the effect of the uncertainty during the operational phase. Initially, the project is considering mixed-pixel class mixture models which explicitly represents the proportional land membership associated with each discrete pixel. This can be regarded as a vector of fuzzy set memberships where instead of assigning a single class label to each pixel, thus ignoring the underlying class mixing, a vector of numbers is produced, each of which represents the proportional land coverage. Uncertainty will also be represented in the empirical classifier design period by adopting a Bayesian learning framework for modelling parametric uncertainty.
The performance of many empirical classification systems is often determined by the suitability of the prior assumptions made about the data. In Artificial Neural Networks (ANN), this typically corresponds to the type of network used and its size. Recent work in the ANN community has focussed on making these algorithms more robust (either using regularisation or model selection approaches) and some of the research into data selection via Support Vector Machines has been shown to improve the accuracy and robustness of various algorithms for large, classification datasets. While the FLIERS project will not compare all possible classification approaches, the fundamental ideas about regularisation, feature selection and data selection will be compared and evaluated.
Currently, the project is 15 months old and this paper will discuss some of the initial software, data analysis, techniques and results which have been achieved so far.
R C Wilson and E R Hancock
A statistical active contour model (snake) is developed for segmenting synthetic aperture radar (SAR) imagery, aiming to identify image regions of homogeneous speckle statistics. The technique is based on the measurement of both the local mean (tone) and variance (texture) of the image and an assessement of the probability that the measured values result from pixels of a desired class. The algorithm measures statistics along the contour so that no smoothing across segment boundaries occurs. The model is cast as an energy functional to be minimised, with constraints being added in the form of curvature and tension energies, to promote a smooth contour of evenly spaced nodes. The weights given to these energies in the model formulation are determined analytically by considering the energy balance of the contour segmenting features of different length scales. The algorithm allows the spawning of smaller snakes that lets the model represent multiply connected regions, allowing the formation of holes and islands. The algorithm allows the segmentation of noisy images to single pixel accuracy.
P Hou, A Hojiatoleslami, C Underwood, and M Petrou
University of Surrey
Microsatellites have the capacity to obtain and store significant number of images. However, they remain above the horizon of a receiving station usually for too sort periods to allow the downloading of all the obtained images. It is necessary, therefore, to compress and perhaps edit the images onboard to improve the downlink capacity of the satellite. A systematic study of image compression techniques has shown that by far the most expensive in terms of required bits for coding, are regions near thick cloud boundaries. Those regions exhipid rapid transitions from very bright to dark that require high frequency components to be encoded for the accurate reproduction. In this paper we use a completely automatic algorithm for the accuarte detection of the rims of these clouds. Further, the algorithm is also used for the detection of large sea areas. Both cloud and ses areas are encoded separately from the land areas with improved compression ratio.
Note: Missing abstracts were submitted in hard copy
Please feel free to comment
on this page.
Creator: Paul Rosin <
Keywords: BMVA meeting
Date: 5 xii 1997