INVESTIGATORS

Mr. Kent K. T. Cheung, Dr. Horace H. S. Ip

BRIEF DESCRIPTION

CLR is a CBIR system for logos based on multiple image features. We aim at retrieving images based on three different aspects (color, shape and spatial relationships) of images. It is hoped by considering more different aspects of images, we can retrieve relevant images based on any one of these aspects to suit the different requirements of different users. Moreover, the retrieval rate of relevant images is also enhanced by employing relevant feedback of these image features.

CLR is also an application of CBIRFrame - an object-oriented framework for content-based image retrieval.

FUNDING AGENCY

N/A

PUBLICATIONS

1. Horace H. S. Ip Dinggang Shen and Kent K. T. Cheung, "Affine Invariant Retrieval of Binary Patterns Using Generalized Complex Moments", Proc. 2nd Intl. Conf. on Visual Information Systems, 15-17 Dec. 1997, San Diego, USA, pp. 301-308.

2. Kent K. T. Cheung and Horace H. S. Ip "Image Retrieval in Digital Library Based on Symmetry Detection", Proc. Computer Graphics International 1998, 22-26 June. 1998, Hannover Germany, pp. 366-372.

3. Kent K. T. Cheung and Horace H. S. Ip (invited), "Semantic Image Retrieval by Color Keywords", Proc. 7th Intl. Conf. on Distributed Multimedia Systems, 26-28 Sep., 2001, Taipei, Taiwan, pp. 324-331.

Multiple Image Features

CLR rely on multple image features to allow the uses to search the database by different image features as required. Unlike other CBIR systems of trademarks, our retrieval features have semantic information so semantic retrieval can be achieved.

• Color

We group similar colors together and name them with color keywords. Thus, the user is allowed to retrieve by naming the colors directly without any knowledge about the color space. It is also shown that retrieval by color keywords histogram improves retrieval rate significantly when compared with using color histogram.

• Shape

Shape retrieval is based on GC moments and symmetry detection in a related project. Symmetry and fold number is shape features readily observed by a common user so semantic retrieval based on these shape features are useful.

• Spatial Relationships

Spatial relationships of image components are extracted and represented as spatial strings. Unlike older approaches of retrieval by spatial strings are simple and easy to extract. We have also implemented an interface to allow the user to speicfy the spatial relationships by sketching the components and their relationships.

• Image Retrieval

The user can search the database by one of the above image features by specifying the query with a specially designed interface. In addition, CBIR by image example is also supported. After specify the image example, the user can also specify the type of image features used for the query.

• Composite

By treating image similarity as fuzzy membership of the fuzzy set of images that are similar to a certain query, we can easily combine results of different queries.

Relevant Feedback

The user is allowed to specify relevant images and then select a kind of image features to refine the query.