الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 76 |  |  |
| | | from | 76 | | |
Abstract
This thesis tackles the problem of visual object retrieval in very large collections of images.
An instance of the object is selected by a user in an image. The aim is to return other
OCCUITences of the selected object from the image collection database quickly and accurately,
despite possible changes in imaging conditions such as scale, viewpoint, lighting and partial
occlusions.
Standard techniques in particular object retrieval use a visual word representation for
fast search. Regions are detected in each image invariant to affine viewpoint and lighting
changes and a descriptor is used to represent local image statistics. These descriptors are then
quantized, typically using k-means. Unfortunately, k-means is not scalable for large datasets.
Hierarchical k-means(HKM) [30] and approximate k-11leans(AKM) [34] quantization were
presented in the literature to solve the scalability issues of k-mean.
We propose another approach of quantization usingfast soft (semi-fuzzy) k-means (FSKM).
Using this approach provides a more good visual vocabulary with a reasonable running-time
that allows scalability for more than 1M database images. Evaluation process showed the
superiority of our new approach over the existing ones.