This thesis is taking part of the ANR project SoLSTiCe (Similarity of locally structured data in computer vision). The goal of this fundamental research project is to develop new models and tools to represent and analyse images and videos. The target applications are object recognition in images, object tracking and activity recognition in video. To tackle these applications, we propose to represent spatial and temporal information as grid. The edge and vertices of these grids are labelled with classical attributes use in
computer vision to describe local information such as geometry, texture, color or visualwords. The next step is to develop new algorithms to extract knowledge from this new description.
The first goal of this thesis is to study the representation of an image as a grid of
visual word. From this, we will try to find frequent pattern through these grids to define relevant patterns in order to be used in image classification. This part could be achieved with the proposition of a new grid mining algorithm.
A second goal will be to use previous research to apply them in video classification.
Special attention will be paid to the representation of the temporal dimension in video, the impact of this dimension to the grid representation of the videos and the impact on the frequent pattern mining.