In this chapter, the authors focus on the use of forests of fuzzy decision trees (FFDT) in a video mining application. They discuss how to learn from a high scale video data sets and how to use the trained FFDTs to detect concepts in a high number of video shots. Moreover, the authors study the effect of the size of the forest on the performance; and of the use of fuzzy logic during the classification process. The experiments are performed on a well-know non-video dataset and on a real TV quality video benchmark.
TopIntroduction
Nowadays, the amount of recorded video is continually increasing leading to a growing need to find a way to handle it automatically. One of the main issues is to be able to index these data with high-level semantic concepts (or features) such as “indoor/outdoor”, “people”, “maps”, “military staff”, etc.
Video indexing aims at analyzing a video, to find its seminal content, and to associate concepts to any of its part. Today effective video indexing is done manually, by a human operator, who associates concepts to parts of a video. However, due to the growth of recorded video, the introduction of automatic approaches, as data-mining-based ones, is a promising perspective.
Video mining is typically an inductive machine learning approach. It has as starting point a set of correctly labeled examples used to train or to build a model. Later, the model is used to perform an automatic classification of any of the forthcoming examples, even if they have not been met before. Video mining is becoming a very active domain today and several conferences take into account this domain in their topics (for instance, the workshop on Video Mining of last IEEE International Conference on Data Mining, or ACM Multimedia conferences, etc.). Some works related to video mining can be cited: (Pan, J.-Y., & Faloutsos, C., 2002), (Rosenfeld, A. et al.; 2003), (Zhu, X., et al. 2005), the proceedings of the TRECVid challenge organized by the US institute NIST.
Inductive machine learning is a well-known research topic with a large set of methods, one of the most commonly used approaches being the decision tree approach (DT). However, robustness and threshold problems appear when considering classical DTs to handle numerical or imprecisely defined data. The introduction of fuzzy set theory, that leads to the construction of fuzzy decision trees (FDT) able to smooth out these negative effects.
In the 2005 TRECVID competition, we studied the use of Fuzzy Decision Trees for this kind of applications (Marsala, C., & Detyniecki, M., 2005). The approach, based on single FDTs (one per concept), provided as result a set of classification rules, which were in the one hand, human understandable, thus allowing further human development; but in the other hand, this first series of tests enables us to discover that, when addressing large, unbalanced, multiclass datasets, a single classifier is not sufficient for direct automatic exploitation. Thus, based on these observations, in (Marsala, C., & Detyniecki, M., 2006) forests of FDT were introduced to cover better the whole input space. The use of forests of decision trees is well-known in classical machine learning, see for instance (Breiman, L., 2001). In fuzzy machine learning, forests of fuzzy decision trees have been introduced some years ago and are becoming more popular nowadays (Bonissone, P.P. et al., 2008), (Crockett, K., et al. (2001), (Janikow, C. Z., & Faifer, M., 2000), (Marsala, C., & Bouchon-Meunier, B., 1997). These approaches differ by the way the FDT are multiplied to grow the forest.