Research into trademark image retrieval has become increasingly active over the last few years (Eakins, 2001, Chap. Other SOM-based image retrieval systems include those of Breiteneder, Merkl, and Eidenberger (1999), who describe a coat-of-arms retrieval system, Barbalho, Neto, Costa, and Netto (2001), whose system uses a compressed image vector to store and retrieve images from a hierarchical SOM, and Garcia-Berro, Torres, and Isern (2003), who use a SOM to identify potential white dwarf stars. Each TS-SOM is trained on one particular feature–colour, texture, edge direction and Fourier descriptors–and outputs from multiple SOMs are combined to retrieve images. The tree structured SOM (TS-SOM) is a pyramid structured SOM that progressively gets larger as one descends the hierarchy. The Self-Organizing Map (SOM) or Kohonen network (Kohonen, 2001), however, has been used as the basis for the PicSOM system (Koskela, Laaksonen, Laakso, & Oja, 2000), an image retrieval system using multiple Tree Structured SOMs (Koikkalainen & Oja, 1990). (1994)) suggest that the problem is far from solved.ĭespite their early promise, neural networks have not been widely used for large-scale image retrieval applications (Oja, Laaksonen, Koskela, & Brandt, 1999). ![]() However, empirical tests of retrieval effectiveness (e.g. Over the last decade, researchers have proposed a rich variety of techniques, including comparison of boundary segment chains (Mehrotra & Gary, 1995), elastic deformation of templates (Pentland, Picard, & Sclaroff, 1996), Fourier descriptors (Zahn & Roskies, 1972), moment invariants (Hu, 1962), Zernike moments (Teh & Chin, 1988), edge direction histograms (Jain & Vailaya, 1996), the angular radial transformation (Sikora, 2001) and wavelets (Mallat, 1989). One such approach has been retrieval by shape similarity, but this has proved particularly challenging. Hence researchers have tried to develop a community of different models (Picard, 1996) to describe different aspects of image appearances. To date, no single technique has been developed that can accurately describe a general image. These typically describe visual characteristics of the image such as colour, texture and shape. ![]() Most CBIR techniques (see Smeulders, Worring, Santini, Gupta, and Jain (2000) for a comprehensive review) operate by computing similarity measures between stored and query images from the values of automatically extracted features. The difficulties involved in finding a desired image in a large collection has led to increasing interest in automatic techniques for content-based image retrieval (CBIR). ![]() The number and variety of image collections available in electronic form has risen rapidly over recent years, leading to both opportunities and problems for image users.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |