Content Based Image Retrieval by Twofold Segmentation
- Author(s):Michael Scharrer
The research field of CBIR deals with the extraction of visual and semantic information form images in a database, in order to facilitate the searching among those images based on such data. Generally, these queries use images as keys as well. The query image has information extracted equally, leading to the same type of data that has been generated for each database image. Similar images can be found, by comparing the extracted data. While the definition of similarity differs by use case, this thesis is only concerned with visual similarity.
A very common method in CBIR works by first calculating local visual information for each pixel position and then extracting the occurrence distribution of the resulting local values. For Performance reasons, the set of local values is segmented into a relatively small number of classes in order to compactly describe the resulting distribution of local features. The similarity of two images is estimated by comparing these distributions.
|extracting local features||segmenting into classes|
In this thesis an extension of this method is proposed that aims to improve accuracy, by first segmenting the images into slices, and using the aforementioned algorithm on each slice. The slice direction is chosen specifically, to retain near rotational invariance in the final method. Performance is only affected slightly, since the majority of work is only done once per image analyzed.
|segmenting an images into sensible slices|