Introduction

The problem of image/video database retrieval is now well known but a general solution to this problem has still not been found. This situation is not likely to be resolved in the foreseeable future. Many of the attempts reported in the literature use the low-level content of an image, such as colour, texture and shape, to construct an index for that image. The major advantage of such an approach is that little human intervention is required. However, most of these systems only allow a user to query using a complete image with multiple regions and are unable to retrieve similar looking images based on a single region. The system being developed at the University of Surrey takes this more flexible single-region based approach.

Database Population

Image to be inserted into database. Could be static image or could be key frame taken from video shot.	Segmented Image using region growing.

When an image is inserted into the database a standard set of colour and texture features are computed for every pixel. These features are then grouped together into perceptually similar regions using standard clustering techniques such as Kmeans algorithm, Kohonen networks or region growing. The mean statistics, location and size of each region is then stored as the index of that image into the database.

Database Query

User specifies rectangular region in image in which he/she is interested in finding examples of throughout the database. In this example the user is interested in finding all examples of blue sky.

A neural network classifier is trained on the colour and texture properties of this query region. This classifier is then used to search through all the indices in the database and retrieve the images which contain similar looking regions.

Example Queries

• Find all examples of blue sky.
• Find all examples of bricks similar to the cathederals.
• Find all examples of this hut.

Publications to date

Below are postscript versions of my few publications to date on this topic.

• 1998
  • Acquisition of a large database for biometric identity verification
• 1997
  • Selecting Features for Neural Networks to Aid an Iconic Search through an Image Database (Conf Paper)
  • A comparison of colour texture attributes selected by statistical feature selection and neural network methods (Conf Paper)
  • Using Feature Selection to aid an Iconic Search Through an Image Database (Conf Paper)
• 1995
  • Iconic indexing in an image database (MSc Thesis)

For further information please contact K.Messer@surrey.ac.uk