Ridgelet-based signatures



This work deals with knowledge extraction from visual data for content-based image retrieval of natural scenes. Images are analysed using a ridgelet transform that enhances information at different scales, orientations and spatial localizations. The ridgelet transform of an image corresponds to the activity of a mother ridgelet at different orientations, scales and spatial localizations. At a given orientation, there are 2n localizations at the highest scale, 2n-1 at the next lowest scale, and so on. For an image of size 2n x 2n, this results in a response of size 2n+1 x 2n+1. The challenge is therefore to create a signature for the image from these responses, that leads to a reduction of the size of the feature whilst preserving relevant information useful for discrimination.
We proposed a method that reduces the size and the redundancy of the ridgelet representation, by defining both global and local signatures that are specifically designed for semantic classification and content-based retrieval. These visual features have been proven efficient for natural image classifications and retrieval when they are used in conjunction with a support vector machine classifier (see the refrences at teh bottom of this page).

Algorithm and source code

To compute the ridgelet signature of a 2n x 2n image you can use Beamlab. It requires to install Wavelab.

Simple signature (global)

1 - Extract a square of maximal size such that its side is 2n x 2n. You can use a resampling method to adjust the size: F_extractLargestSquare.m
2 - Compute the ridgelet transform of this central square of the image.
3 - Let use the following program to compute the signature: F_GlobalRidgSig.m

Generalized signature (global and local)

1 - Extract a square of maximal size such that its side is 2n x 2n. You can use a resampling method to adjust the size: F_extractLargestSquare.m
2 - (skip to compute the global signature) Let divide this square into 16 smaller squares according to a 4x4 grid.
3 - Let design a template of activity. Here are some examples:
  • global templates to compute signatures with 256x256 ridgelet transform (i.e on 128x128 central squares). protoFull.mat computes the "simple signature (global)" described in the above algorithm.
  • local templates to compute signatures with 64x64 ridgelet transform (128x128 central squares givin 16 small squares of size 32x32)
4 - Let use the following program to compute the signature according to this template: F_RidgSig.m


This work is reported in the following papers:

Hervé Le Borgne, Noel O'Connor
Natural scene classification and retrieval using Ridgelet-based Image Signatures
Proc. of the Advanced Concepts for Intelligent Vision Systems (ACIVS 2005), pages 116-122, Antwerp, Belgium, 20-23 september 2005.
[pdf file |link to editor]

Hervé Le Borgne, Noel O'Connor
Ridgelet-based signatures for natural image classification
Proc of the 2ième conférence en Recherche d'Informations et Application (CORIA 2005), Grenoble, France, 9-11 mars 2005.
[pdf file]

Postal adress

Dr. Herve Le Borgne
Centre de Fontenay-aux-Roses - Bat. 38/2 - DTSI/SRCI
BP 6 - 92265 Fontenay-aux-Roses

Mail and phone:

Avoid SPAM

Websites : Hervé LE BORGNE Noel O'CONNOR,