Image Processing and Pattern Recognition


Generalised Hough Transforms, Pseudo-Inverse methods, Shape from Shading

J R Barker, S. Hill, J. Slaewaegen, M . H. Wahab

Novel image processing and pattern recognition techniques are being developed to automatically identify terrestrial and planetary impact craters. The techniques involve 3D reconstruction algorithms and statistical parametric shape identification tuned to recognise buried and overlapping craters.

Pattern recognition techniques for remote detection of impact crater morphology on planetary surfaces

Computer Vision, Scene Analysis and Adaptive Pattern Recognition

J R Barker, S. Sens, F. Smoes and F Young

Scene analysis, for both single frame and sequences of frames, is an important area of artificial intelligence which is making computer vision a reality. It is important for applications which range from surveillance, computer-assisted tomography, microscopy to industrial product monitoring and robotics. Our interest is specifically in the analysis of images obtained from scanning tunnelling microscopy and related techniques. It is particularly important to recognise a particular molecule against a particular background and to discriminate it from “dirt” or artefacts of the imaging process. It is also important to measure the detected target and to find its distortion from the perfect form. We have therefore developed algorithms which can detect one or more molecular targets within a scene of adsorbed molecules plus substrate. At this level we use a self-organising or synergetic algorithm which represents an unknown scene as a potential surface in an abstract feature space. the input scene is filtered against known images using a non-linear overdamped technique which locates the most dominant recognisable sub-image in the scene and identifies it. By shutting off the attention parameters for the recognised component(s) the program them seeks the next identifiable objecxt and proceeding recursively finally obtains a hierarchical description of the scene in terms of known objects. The unknown features are filtered out and may be “learned” if they represent a new object or further analysed. A second algorithm is used to determine the orientation, position and scaling of any of the recognised objects with repect to a training image (the affine transformation parameters). The novel elements of this project involve the ability to perform scene analysis, invariant pattern recognition up to an affine transformation and for adaptive learning. The techniques are proving particularly useful in molecular electronics applications where the “scenes” involve “semi-transparent” distorted molecules covering layers of other molecular and inorganic materials. Our long term interest invoves the applications of the above ideas to general problems in computer vision. Funded by the EU TOPFIT project (see grants)

Algorithms for  scanning tunnelling microscopy

Computer vision

Olfactory sensing (electronic nose)

Algorithms for image processing in biosciences (see bioelectronics)


Comments are closed.