Details of Grant 

EPSRC Reference:	GR/J36662/01
Title:	PSYCHOPHYSICAL MODELS FOR COMPUTATIONAL IMAGE SEGMENTATION
Principal Investigator:	Thomas, Professor B
Other Investigators:	MacKeown, Dr W Troscianko, Professor T
Researcher Co-Investigators:
Project Partners:
Department:	Computer Science
Organisation:	University of Bristol
Scheme:	Standard Research (Pre-FEC)
Starts:	01 April 1994	Ends:	31 May 1996	Value (£):	74,896
EPSRC Research Topic Classifications:	Image & Vision Computing
EPSRC Industrial Sector Classifications:
Related Grants:
Panel History:
Summary on Grant Application Form
1. To develop a psychophysically plausible computational model of segmentation which includes the use of colour and texture information.2. To optimise the model by comparison with human segmentations from a high-quality database of labelled colour images of outdoor scenes.3. To extend the model by using a neural network to incorporate higher-level knowledge extracted from the database.Progress:In order to allow various segmentation models to be compared, it was necessary to investigate methods for quantitatively assessing the performance of individual segmentations. A metric was devised to allow comparison of any two segmentations giving a measure of their similarity. An important case here involves comparison of a machine segmentation with the ideal human segmentations from our database of natural scenes. This allows us to quantify the performance of any segmentation method. The metric is based on area overlap and perimeter distance measures. The metric allows discrimination between over-segmentation and under-segmentation since this is important for further work. A straightforward monochrome region-based segmentation method is currently being used to allow initial assessment of the contribution of colour and texture. This is shortly to be replaced by a method incorporating edge information and perceptual grouping techniques which we are currently developing. A neural network has been trained to make merge decisions for adjacent regions in over-segmented images by comparison with the ideal segmentations. We have shown that the resulting segmentation is significantly better when colour and texture information is provided than with monochrome information alone. The current network achieves 83% accuracy on merge decisions. The input features used are based on a 32-channel Gabor image representation for texture together with 36-bit colour information.
Key Findings
This information can now be found on Gateway to Research (GtR) http://gtr.rcuk.ac.uk
Potential use in non-academic contexts
This information can now be found on Gateway to Research (GtR) http://gtr.rcuk.ac.uk
Impacts
Description This information can now be found on Gateway to Research (GtR) http://gtr.rcuk.ac.uk Summary Date Materialised
Sectors submitted by the Researcher
This information can now be found on Gateway to Research (GtR) http://gtr.rcuk.ac.uk
Project URL:
Further Information:
Organisation Website:	http://www.bris.ac.uk