The present invention relates to a method for measuring the dissimilarity between images, a method for ranking images from the most similar to the less similar to a query image, a method for categorizing a query image into at least two categories and a method for measuring the dissimilarity between video sequences.

Fig 1. Examples of images ranking (from COREL Database)

TECHNICAL DESCRIPTION

The invention proposes a new global description based on Sparse Multiscale Patches. The method comprises the following steps : 

• multiresolution decomposition of the first and the second images

Laplacian pyramid decomposition ; classical, complex and/or redundant wavelet transforms;  steerable pyramid, bandlets, curvelets, etc.

• constituting vectors (patches)

multidimensional feature vectors (patches) that capture interscale and intrascale dependencies among subband coefficients. These are better adapted to the description of local image structures and texture.

• evaluation of the dissimilarity between the probability density function of patches

Kullback-Leibler divergence, the Battacharya measure, mutual information, the Hellinger distance, or more generally a Bregman divergence (possibly symmetrized).

BENEFITS

• More effective measure of the dissimilarity
• Dissimilarity ranking
• No prior text annotation of the images

INDUSTRIAL APPLICATIONS

Image processing for:

• searching for digital images in large databases
• categorization of images,
• ranking images from the most similar to the less similar to a query image
• detection of video copies

Useful for Art collections, Photograph archives, Retail catalogs, Medical diagnosis, Intellectual property, Architectural and engineering design, Geographical information, …

EXAMPLES OF CONTENT-BASED RETRIEVAL IMAGES

Retrieval results for 5 images of the COREL database. For each row, left to right: query image; first 4
ranked images of the database (excluding the query image).
For each retrieved image, the SMP similarity measure to the query is also shown.

For further information, please contact us (Ref 02268-01)