University Information Service: Contents | Index | Search

GIP Deep Multispectral Image Processing

Multispectral imaging entails acquiring several images of the same scene using different spectral bands. For instance, a digital colour camera detects three separate images for the red, green and blue components of light. Collecting several spectral bands generally provides more information than would be obtained from a single monochrome image.

This idea has been applied in the field of remote sensing for nearly 20 years. LANDSAT satellites are capable of acquiring 57 spectral bands spanning visible and non-visible wavelengths such as infra- red. The full set can be processed to identify different kinds of land use automatically. We describe this as shallow multispectral image processing because the number of spectral bands is small compared to the number of spatial sample points in any direction. Our reseach concerns deep multispectral image data, where the number of spectral bands is comparible to the number of spatial sample points in any direction. Certain kinds of advanced scientific instrument, such as analytical electron, X-ray and ion microscopes, are theoretically capable of acquiring hundreds or even thousands of spectral bands for a single scene. For instance, we have recently upgraded the multispectral analytical electron microscope (MULSAM) at the University of York to permit as many as 8192 energy analysed electron (Auger) images plus 1024 energy analysed X-ray images to be collected simultaneously. Processing and interpreting this kind of data is not straightforward. Existing multispectral image processing techniques are not easy to extend to such large data sets.

The Bradford Graphics and Image Processing Group is investigating new ways to visualize, analyse and compress deep multispectral image data. Although the project has been running for only one year, significant progress has been made on all three fronts.

Compression

Deep multispectral image data requires an enormous amount of storage. For example, a 512x512x8192 16-bit deep image set would occupy 4Gbytes.

Conventional compression algorithms are very ineffective for this kind of data because of the large amount of noise present. A new hybrid lossless compression technique that can handle noisy data has been developed. It involves a spectral decorrelation stage followed by optimal variable bit length (Huffman) coding. Unlike conventional Huffman coding, no assignment table need be passed. Instead the code assignments are reconstructed using a statistical model derived from a small sample of the data.. This makes it possible to encode large integer data values, such as 32-bit integers, without needing to record the code assigned to each. Highly efficient compression methods suitable for archiving are also being investigated. One promising approach is to use the Karhunen Loeve transform to eliminate redundancy in the spectral domain in conjunction with an efficient spatial method, such as wavelet compression. Early results suggest compression ratios of up to 500:1 could be achieved.

Visualization

Human vision is geared to looking at surfaces in 3D, not full 3D arrays of data values. State of the art visualization packages such as Khoros or Explorer are of very limited use for deep multispectral image data. A new volume rendering algorithm has been developed that can present a deep multispectral image set as a single picture. It involves an empirical peak quantification stage followed by a modified form of ray-casting that uses interpolation. Significant spectral features are automatically rendered as opaque regions with a distinct colour according to the energy. Spectral bands containing no significant features are made to appear transparent.

Example of 3D spectrum-image visualization

Automatic Analysis

Automatic analysis techniques produce higher-level representations of the raw image data. For instance, a single image where each of the different types of region appears as a distinct colour would be easier to interpret than a 3D volume. The ultimate aim of automatic analysis is to identify and label every region with its chemical composition. We are currently investigating segmentation (a.k.a. partitioning) techniques for multispectral data. We have extended standard k-means (cluster) segmentation to handle variable numbers of region types in a robust manner and to handle many dimensions efficiently. An interactive segmentation method using 3D scatter diagrams has also been implemented and successfully evaluated by the Surface Science group at the University of York, who are collaborating with this project.

Industrial or Academic Collaboration

We currently collaborate with a small number of academic and industrial research institutions, mostly on an informal basis. Enquiries from other workers interested in the development, implementation or application of deep multispectral image processing would be most welcome. We are particularly keen to obtain a wide variety of test data in order to evaluate our new techniques.

Research Student Opportunities

We would welcome enquiries from prospective MPhil and PhD students interested in pursuing research in the area of deep multispectral image processing. A suitable student would have, or would soon have, a good honours degree in a numerate/physical science such as computer science, mathematics, physics, electronics or chemistry.

Further Information

For further information about Deep Multispectral Image Processing please contact:

Dr. Peter Kenny
Computing Department
University of Bradford
Bradford, West Yorkshire
BD7 1DP

Email: p.g.kenny@comp.brad.ac.uk
Phone: 01274 383928
Fax: 01274 383920

Publications

Author: P.G.Kenny
Last update: 16/2/96

Up to GIP group homepage