Related Software Tool: POINCaRe Related Publications (selected):
Understanding the content of images or similarity between them is an extremely difficult problem for computer vision algorithms. This research follows a human centred approach to define quantitative measures of similarity between images based on their local visual descriptions. ![]() Although the exact mechanism of the human visual perception is unknown, we used several conjectures about various aspects of the human perception and utilized mathematical models to represent these aspects. The main conjecture is that visual perception is approximate in nature with limited resolution. For example, objects that are close enough in terms of their visual descriptions are seen as almost similar. Also visual perception is formed by grouping similar parts of the image(s) together (second principle of Gestalt theory of visual perception). ![]() ![]() we used tolerance neighborhoods to
represent these groups of almost similar elements. we define the similarity
measures based on how these neighborhoods cover both images. Another conjecture
is that the transition between the perception of similarity and dissimilarity is
gradual rather than abrupt. This lack of a sharp threshold in visual perception
of similarity was my motivation for defining novel fuzzy tolerance neighborhoods
and fuzzy similarity measures. ![]() ![]() The resulting measures can be used for
content-based image retrieval (CBIR), i.e. searching for images based on
content. |
Research >