Dynamic texture classification using deterministic partially self-avoiding walks on networks

Abstract

This paper presents a new approach to dynamic texture classification based on deterministic partially self-avoiding (DPS) walks on complex networks (or graphs). In this approach, for each pixel is assigned a vertex and two vertices are connected according to a given distance. In order to analyze appearance and motion, we propose two graph modeling: a spatial graph and a temporal graph. The DPS walks are agents that can obtain rich characteristics of the environment in which they were performed. Thus, the DPS walks are performed in the two modeled graphs (spatial and temporal) and the feature vector is obtained by calculating the statistical measures from the trajectories of the DPS walks. The results in two well-known databases have demonstrated the effectiveness of the proposed approach using a small feature vector. The proposed approach also improved the performance when compared to the previous DPS walks based method and the graph-based method.