In this thesis a novel approach to image classification is presented. The thesis explores the use of color feature vectors and CBIR retrieval methods in combination with Genetic Programming to achieve a classification system able to build classes based on training sets, and determine if an image is a part of a specific class or not. A test bench has been built, with methods for extracting color features, both segmented and whole, from images. CBIR distance-algorithms have been implemented, and the algorithms used are histogram Euclidian distance, histogram intersection distance and histogram quadratic distance. The genetic program consists of a function set for adjusting weights which corresponds to the extracted feature vectors. Fitness of the individual genomes is measured by using the CBIR distance algorithms, seeking to minimize the distance between the individual images in the training set. A classification routine is proposed, utilizing the feature vectors from the image in question, and weights generated in the genetic program in order to determine if the image belongs to the trained class. A testset of images is used to determine the accuracy of the method. The results shows that it is possible to classify images using this method, but that it requires further exploration to make it capable of good results.
| Identifer | oai:union.ndltd.org:UPSALLA1/oai:DiVA.org:ntnu-9327 |
| Date | January 2006 |
| Creators | Edvardsen, Stian |
| Publisher | Norges teknisk-naturvitenskapelige universitet, Institutt for datateknikk og informasjonsvitenskap, Institutt for datateknikk og informasjonsvitenskap |
| Source Sets | DiVA Archive at Upsalla University |
| Language | English |
| Detected Language | English |
| Type | Student thesis, info:eu-repo/semantics/bachelorThesis, text |
| Format | application/pdf |
| Rights | info:eu-repo/semantics/openAccess |
Page generated in 0.0021 seconds