OpenVT: aplicação escalável de código aberto para busca de imagens semelhante utilizando árvore de vocabulário

Abstract

Content-based image retrieval (CBIR) is the application of computer vision techniques to the image recovery problem. ”Content-based” means that information contained in the images - features - is used to perform the searches. This type of search system can be used for numerous purposes, such as verifying that images are being used without authorization, locating people and objects in movies and videos, and searching for landmarks such as real estate and sights. Commonly encountered CBIR applications use patented algorithms or have their codes closed. With this in mind, the objectives of this work are: the construction of an image search engine, which is scalable to a lot of images and is also robust to the different conditions that an image can find, such as scale, rotation and lighting; besides verify the possibility of using only open source methods, resulting in an open source application with satisfactory performance, seeking to foster and help future developments. Thinking about efficiency and scalability as main features, the method proposed by Nister e Stewenius (2006) - vocabulary tree - was defined as the basis for this work. In order to reach the proposed objectives, a survey was made of related works in order to discover the main gains and challenges of using vocabulary trees as a base structure. After investigating and specifying the functionalities, the proposed application was developed, aiming for efficiency, agility and flexibility. The three main tasks that this work focused on were: image feature extraction, building of the vocabulary tree and score system. To evaluate the efficiency of the application resulting from this work, this was compared with patented algorithms and the state-of-the-art in image recognition. Finally, it was possible to conclude that quality and efficient applications can be created using only open source methods and that vocabulary trees are adaptable and flexible structures with diverse applications and excellent search performance.