Full text
| |
| Author(s): |
Filipe de Oliveira Costa
Total Authors: 1
|
| Document type: | Doctoral Thesis |
| Press: | Campinas, SP. |
| Institution: | Universidade Estadual de Campinas (UNICAMP). Instituto de Computação |
| Defense date: | 2016-06-14 |
| Examining board members: |
Anderson de Rezende Rocha;
Esther Luna Colombini;
Sandra Eliza Fontes de Avila;
Agma Juci Machado Traina;
Roberto Hirata Junior
|
| Advisor: | Anderson de Rezende Rocha; Zanoni Dias |
| Abstract | |
Digital documents (e.g., images and videos) have become powerful tools of communication with the advent of social networks. Within this new reality, it is very common these documents to be published, shared, modified and often republished by multiple users on different web channels. Additionally, with the popularization of image editing software and online editor tools, in most of the cases, not only their exact duplicates will be available, but also manipulated versions of the original source (near duplicates). Nevertheless, this document sharing facilitates the spread of abusive content (e.g., child pornography), copyright infringement and, in some cases, defamatory content, adversely affecting the public image of people or corporations (e.g., defamatory images of politicians and celebrities, people in embarrassing situations, etc.). Several researchers have successfully developed approaches for the detection and recognition of near-duplicate documents, aiming at identifying similar copies of a given multimedia document (e.g., image, video, etc.) published on the Internet. Notwithstanding, only recently some researches have developed approaches that go beyond the near-duplicate detection task and aim at finding the ancestral relationship between the near duplicates and the original source of a document. For this, the development of approaches for calculating the dissimilarity between near duplicates and correctly reconstruct structures that represent the relationship between them automatically is required. This problem is referred to in the literature as Multimedia Phylogeny. Solutions for multimedia phylogeny can help researchers to solve problems in forensics, content-based document retrieval and illegal-content document tracking, for instance. In this thesis, we designed and developed approaches to solve the phylogeny reconstruction problem for digital images and videos. Considering images, we proposed approaches to deal with the phylogeny problem considering two main points: (i) the forest reconstruction, an important task when we consider scenarios in which there is a set of semantically similar images, but generated by different sources or at different times; and (ii) new measures for dissimilarity calculation between near-duplicates, given that the dissimilarity calculation directly impacts the quality of the phylogeny reconstruction. The results obtained with our approaches for image phylogeny showed effective, identifying the root of the forests (original images of an evolution sequence) with accuracy up to 95%. For video phylogeny, we developed a new approach for temporal alignment in the video sequences before calculating the dissimilarity between them, once that, in real-world conditions, a pair of videos can be temporally misaligned, one video can have some frames removed and video compression can be applied, for example. For such problem, the proposed methods yield up to 87% correct of accuracy for finding the roots of the trees (AU) | |
| FAPESP's process: | 13/05815-2 - Image and Video Phylogeny Reconstruction |
| Grantee: | Filipe de Oliveira Costa |
| Support Opportunities: | Scholarships in Brazil - Doctorate |