Workpackage 4 Image Analysis Algorithms Progress Update Sept

Workpackage 4 Image Analysis Algorithms Progress Update Sept. 2001 Kirk Martinez, Paul Lewis, Fazly Abbas, Faizal Fauzi, Mike Westmacott, Marc Chiaverini Intelligence, Agents and Multimedia Research Group Department of Electronics and Computer Science University of Southampton UK PMT meeting, Sept 22, 2001

Overview • Progress on Texture-Segmentation and Classification • Query by Low Quality Images • MNS • Query by Sketch • colour clustering • craquelure detection PMT meeting, Sept 22, 2001

Progress on Texture Segmentation and Classification • • Texture in image processing is concerned with repeating patterns Work on texture is currently concentrating on wavelets Wavelet transforms analyse the image according to scale and frequency Transforms can use different decomposition strategies and different base wavelet functions (cf Fourier which uses sines and cosines only) PMT meeting, Sept 22, 2001

Segmentation for Texture Indexing • • Idea is to divide the image into major regions of homogeneous texture Then store representation of each significant texture so that images containing similar textures can be retrieved eg we have an image of a textile. We may wish to ask, “are there other images containing a similar textile pattern? ” Texture may also be a useful contributing key for style classification PMT meeting, Sept 22, 2001

Query by Low Quality Images eg Faxes • • Modified the standard wavelet retrieval to use all but the lowest frequency coefficient Using a set of 19 faxes we evaluated retrieval by fax using a database of 150 images including the originals for the 19 fax images. PMT meeting, Sept 22, 2001

Using Daubechies Wavelets Ranking PWT Modified PWT Top 5 3 9 6 -10 5 3 11 -20 4 2 21 -30 1 3 31 -40 1 0 Other 5 2 PMT meeting, Sept 22, 2001

Fax Queries and Database Image PMT meeting, Sept 22, 2001

PMT meeting, Sept 22, 2001

PMT meeting, Sept 22, 2001

MNS- Multi-Nodal Signature • • • Uses colour pair patches as key for matching Original version only used presence of a colour pairs and no real scope for indexing Now exploring use of quantised colour pairs, an indexing strategy and use of frequency of occurrence within an image and inverse of document frequency as weightings. PMT meeting, Sept 22, 2001

Query By Sketch • No work yet but could use paint package to create sketch and feed into M-CCV or MNS algorithms PMT meeting, Sept 22, 2001

Colour Space Custering PMT meeting, Sept 22, 2001

Identifying a cluster PMT meeting, Sept 22, 2001

Labelling an image with pigment PMT meeting, Sept 22, 2001

Crack Detection Vertical + horizontal detection Original image diagonal detection PMT meeting, Sept 22, 2001 Detected cracks

cracks: another example • Next stage is to classify them! PMT meeting, Sept 22, 2001