Authors
Cruchkova E. N.
|
Degree
|
PhD in Physics & Mathematics, Professor, State Technical University |
|---|---|
|
E-mail
|
kruchkova_elena@mail.ru |
|
Location
|
Barnaul |
|
Articles
|
Classification of complex images based on semantic gImage classification is a complex problem due to classes’ natural variability, possibly visual intersections
and due to lacking of sufficient information in visual representation only. Classification methods
bases on representing images as a set of visual words and then transforming them into appearance
frequency histogram has proved it’s stability in last years. However, such approach is based
on usage of a set of separated classifiers trained on some learning sets, and lacking any information
about relations between them. Such information can be useful when image being analyzed contains
some form of classes’ intersection or instances of more than one class. In that case whole-image
classification can become unreliable and some king of post-processing is required. Current work
explores the possibility to use information contained in semantic graphs for described problem. With
a set of words in natural language as vertices in semantic graphs it is possible to gather connected
learning images with usage of search systems like Google. Edges in semantic graph can be used
a metric base for verification and correction algorithms which runs after a separated whole-image
classification process. It is possible in some cases to combine semantically-close classes when analyzing
a complex image and when separated whole-image classification becomes unreliable. Such
method is given, with used formulas and results in table form.
Read more...
|
