Κατηγοριοποίηση Εικόνων Με Τεχνικές Χωρικού Ταιριάσματος Και Δεικτοδότησης

Abstract

In the framework of the subject thesis, we introduce a new images categorizationmethod, which combines support vector machines (SVM) and Hough spatial matching(HPM), while implemented within the indexing procedure of image retrieval based oninverted file. After stating the HPM kernel proof, we train single SVM classifiers for eachcategory and combine them in a multiclass fashion, on the approach one versus the rest.The method is summarized as follows: we extract SURF descriptors from the queryimage, attribute visual words to them, compare query descriptor vectors with the respectivedatabase images’ vectors, perform HPM spatial matching of the query image and thesupport vectors (SV) of all subject category classifiers and finally we perform multiclassclassification, choosing the category with the maximum decision value in each case.For the first time, we introduce in SVM a kernel, other than the linear BoW, combiningspatial information and translation, scale and rotation invariance. We exploit the sparseand compact representations SVM provide, in combination with the spatial informationHPM efficiently uses. Moreover, in conjunction with indexing included in inverted file,the new method promises more efficient image categorization machines, as far as timecomplexity and space requirements are concerned, when refering to rapidly growing imagecollections.The method is evaluated on experiments, performed on our own dataset and is comparedwith the existing kerneled SVM with BoW. The dataset is constructed from the Worldcities dataset of Image, Video and Multimedia Systems Laboratory in NTUA based on alandmark recognition method combining visual and geographical clustering.