Image interpretation via material specific spectral characterisation models

Ela Claridge, School of Computer Science, The University of Birmingham

Supported by the Leverhulme Trust grant number F/00 094/M


Computer image interpretation is concerned with applying computational and mathematical techniques to digital images in order to extract specific information from image data. Most methods have been developed to interpret images of the three-dimensional world. As surfaces were considered to be the most ecologically important structures of the world, images were assumed to represent primarily light reflected from surfaces [1-2]. The originality of our method stems from the insight that the "appearance of a surface [...] is a result of [...] processes at and within this surface"[3]. This insight is at the heart of a novel and generic method of image interpretation which complements the existing techniques.

The image is considered to be the product of the interaction of light with the constituents of the scene. The physics of light transport is utilised to construct a spectral characterisation model - a mathematical model capable of predicting the range of colouration expected from materials known to occur within the scene. By comparing image data to these predictions material characteristics can be deduced.

Our group has successfully applied this approach to develop a novel skin imaging system [4-5]. From two images of the skin, one colour and one acquired through a near-infrared filter, informative parametric maps are computed, containing detailed information about the concentration of melanin and blood and collagen thickness across the imaged skin [5]. Current medical trials are showing the maps to be of great value in the diagnosis of melanoma [6].


The ultimate goal of the project is to formulate a generic approach to image interpretation based on the spectral characterisation models. This should enable the structure and composition of the materials and tissues to be deduced from their images acquired through a small number of optical filters using a standard digital camera.

The specific objectives of this work are to:

Significance and originality

The proposed project is interdisciplinary, it draws ideas from physics, chemistry, applied mathematics, computer science and visual perception. Together, these ideas expand the boundaries of what is currently possible through image interpretation.

The key novelty of our approach is in its development and exploitation of the spectral characterisation models of specific materials. The structure of the model allows us to create a simple cross-reference between the object composition and its colour (or some richer spectral characterisation). As the model is constructed using physics (rather than, for example, statistics) the physical origin of colours seen in images can be explained with reference to the actual physical phenomena. The mathematical form of the model allows us to take advantage of the well developed spectrophotometric quantitative methodologies whilst presenting the results as images, which convey rich spatial information.

The practical significance of the proposed method is that for suitable classes of materials, quantitative estimates of their composition can be obtained from their image data. The method is instant and non-invasive. Once a set of optical filters is defined for a given class of material, the filtering during the image acquisition phase followed by nearly real time post-processing enables appropriate image sets to be presented for interpretation. The underlying scientific technique is fundamentally generic and is the potential basis of a unique non-destructive testing technology which does not require complex equipment and could be implemented using only a digital camera with filters. The applicability of the method to a given class of materials can be assessed due to the underlying predictive theory.


To construct spectral characterisation models we shall use the well known `Monte Carlo' computational method. By supplying parameters characterising a given material (the number and thickness of layers, optical coefficients of the layer components, etc), this method can be used to computationally simulate interactions of light photons with the material and generate the spectrum(a) characterising the remitted light. A collection of such spectra, linked to the input parameters, is a spectral characterisation model.

A spectrum specifies the light levels associated with each of several hundred discrete spectral wavelengths. Although it is possible to acquire digital images corresponding to many wavelengths, this is costly in terms of storage space and processing time. Standard colour video cameras "compress" the whole light spectrum to just three components: red, green and blue. This is achieved by filtering the incoming light through red, green and blue filters. Most people find that this represents images quite adequately, because these three filters are tuned to the colours that the human visual system perceives best. These filters are, however, unlikely to be the best choice for all the materials. One of the most important parts of the project is to develop general methods for defining filters which would be most appropriate for specific materials. An ideal filter will produce an image showing even small variations in a property of a single component of a material. There exist statistical and mathematical methods for finding such "optimal" filters.

Finally, to evaluate the methodology, the developed techniques will be practically applied to images from two different domains. Possible candidate domains include agriculture, ophthalmology and remote sensing.


  1. Marr D (1982) Vision, a Computational Investigation into Human Perception and Processing of Visual Information: Freeman.
  2. Watt R (1991) Understanding Vision: Academic Press.
  3. Egan WG, Hilgeman TW (1979) Optical Properties of Inhomogeneous Materials: Academic Press.
  4. Cotton SD, Claridge E (1996) Developing a predictive model of human skin colouring, Vanmetter RL, Beutel J Eds., Proceedings of the SPIE Medical Imaging 1996 vol. 2708, 814-825.
  5. Cotton SD, Claridge E, Hall PN (1997) Noninvasive skin imaging, Information Processing in Medical Imaging (Springer-Verlag, LNCS 1230), 501-507.
  6. Moncrieff, M, Cotton SD, Claridge E, Hall PN, (2000) Spectrophotometric Intracutaneous Analysis assists in the identification of dermatoscopic features. J.Assoc Derm Venereol,220.

Page maintained by Ela Claridge