Full text
| |
| Author(s): |
Mateus Roder
Total Authors: 1
|
| Document type: | Master's Dissertation |
| Press: | Bauru. 2021-03-23. |
| Institution: | Universidade Estadual Paulista (Unesp). Faculdade de Ciências. Bauru |
| Defense date: | 2021-02-25 |
| Advisor: | João Paulo Papa |
| Abstract | |
In the past decade, the exponential growth of data has supported the development of a wide range of algorithms based on machine learning, enabling its uses in daily basis activities. Besides, such improvement is partially explained due to the advent of deep learning techniques, i.e., the composition of simple architectures that generate complex and robust models. Although both factors produce excellent results, they also have disadvantages concerning the learning process, since training complex models in large data sets are computationally expensive and time-consuming. This problem becomes evident when it comes to the video analysis and processing, as recognition of actions or events, since sequences of images (frames) are considered and usually generate a single output. Another relevant problem concerns the low number of high-level events classification databases, making it difficult to develop some conceptual aspects. Some studies consider transferring learning or a domain adapting, that is, approaches that map knowledge from one domain to another, to lighten the training load as most of them operate in individual blocks or small blocks of frames. Therefore, this work proposes a new approach to map knowledge between domains, from action recognition to event recognition/classification using energy-based models as a mapping function. Also, it is proposed a modification in the video processing for the employed models, capable of processing all frames simultaneously by carrying spatial and temporal information during the learning process, denoted as Somatório processing. The experimental results conducted over two public video data sets, the UCF-101 and the HMDB-51, portrait the effectiveness of the domain adaptation approach and the proposed Somatório models, reducing the computational load when compared to the standard energy-based models, such as Restricted Boltzmann Machines, Deep Belief Networks, and Deep Boltzmann Machines. (AU) | |
| FAPESP's process: | 19/07825-1 - Deep Boltzmann machines for event recognition in videos |
| Grantee: | Mateus Roder |
| Support Opportunities: | Scholarships in Brazil - Master |