Year: 2011
Type: dissertation
Abstract: This M.Sc.Thesis presents a comparative study of visual descriptors with an exhaustive evaluation for near-duplicate 1 scene detection within video content.The concept of near-duplicate images refers to a different geometric and photometric transformations of the images that belong to the same scene among several types of video scene.As a content set for the study we have created a proprietary database of real images made up from different sets of videos such as the BBC rushes videos from TRECVID Video Retrieval Evaluation 2 or Time-Slice® Film's videos from Vimeo 3 among others.We first start by describing certain features, techniques and methods of duplicate image detection based on different algorithms and image features proposed in the literature.After an introduction to near-duplicate scene detection and an exposition of different applications related to this issue, a review of the state of the art of the selected visual descriptors and content-based image retrieval techniques is presented.Below a comparative analysis of the selected methods is performed, including the proposal of improvements for the existing techniques by their combination.In the experimental section, we have carried out different improvements of every basic algorithm which are included in a optimization stage showing the advantages and disadvantages of each visual descriptor by a comparative evaluation.Then individual and combined refinements have been performed in order to demonstrate the improvements sought respect to the original results.Besides, a computational cost evaluation of the improved techniques has been performed in order to obtain more rigorous and useful conclusions for the inclusion of the techniques to in real systems.As a result we have obtained the combination of techniques that perform better in terms of precision and computational performance.Finally, the conclusions and feature work directions are presented.
Author: Óscar Boullosa García
Language: Spanish
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2342546817