A lecture entitled “Image Reconstruction and Retrieval with Parallel Singular Value Decomposition” will take place on Jun. 4 at 10:00. Prof. Marian Vajtersic will show a vector space approach to image reconstruction and retrieval, its parallelization, and a viable alternative to this approach that is based on the Singular Value Decomposition (SVD). The lecture will be held in the room TH:A-1435.

For search and retrieval of images in parallel distributed environments, mostly a vector space approach is adopted. In this approach, objects (images, documents) and their features are represented by a corpus matrix where columns characterize objects and rows their features. Specifically, for image retrieval, we have relatively few features, but each of the images is characterized by a full feature set. Hence, the corpus matrix is dense, ias to its shape, it is usually thin and long (i.e., rows are substantially larger than columns).

Our goal is to find some specific requirements of image retrieval and to translate them into a suitable model that will be amenable to parallel processing. In our model, we distinguish three types of queries depending on its type, source entity, and target. For each type of queries, we compute similarity vectors. In order to evaluate these vectors in parallel, we consider three basic partitionings of the corpus matrix. Fundamental operations in these computations are accumulations and sortings across the partitions.

The system configuration under consideration is a 2D mesh with a special tree-like communication structure enabling to perform the query process efficiently. The retrieval efficiency has been tested for three different mesh configurations. It is shown that our model is promisingly good for queries returning objects and the speedup of delivery increases with the size of image collections.

In the second part of the talk, we present a viable alternative to the vector space approach that is based on the Singular Value Decomposition (SVD). It is known that the basic space model fails in some cases but the SVD does not. This routine is so robust that even a low-rank SVD factorization of the corpus matrix can lead to good approximations in the query process.

On the other hand, SVD is costly to be computed. Therefore, in order to exploit SVD for retrieval problems on very large data collections, its parallelization is inevitable. At the end of the talk, we mention some of our newest algorithmic developments for parallel computation of SVD on parallel computer platforms with distributed memory.

The event is free of charge and you do not have to register anywhere. The lecture is intended for academicians, students or researchers interested in parallel and distributed computing, numerical methods, and image processing.

- Event type
- Lecture
- Lecturer
- Prof. Marian Vajteršic
- Department of Computer Sciences, Paris Lodron University of Salzburg, Austria
- Institute of Mathematics, Slovak Academy of Sciences, Bratislava, Slovakia
- Date
- June 4, 2018, 10:00
- Place
- Conference room TH:A-1435, Building A
- Thákurova 7, Prague 6
- Language
- English
- Video
- will be not recorded

Professor Marian Vajtersic received his Ph.D. from Slovak Academy of Sciences in Bratislava in 1984. In 1994 he received the DrSc degree from the Comenius University in Bratislava. He got a habilitation in Parallel Computing from the University of Salzburg in 1996. From 2002 he is Full Professor for Computer Architecture and High-Performance Computing at this university. He is author of four monographs and more than 130 scientific papers in the area of parallel algorithms and scientific computing.

- People
- Lecturer: Marian Vajteršic
- Places
- TH:A-1435

- Person responsible for the content of this page
- prof. Ing. Pavel Tvrdík, CSc., pavel.tvrdik@fit.cvut.czHead of the Department of Computer Systems

Last modified: 30.5.2018, 12:52