Data pre-processing to improve SVM video classification

In this work a pre-processing strategy to improve the performances of SVM in video clips classification is proposed. The segmentation of a video clip and the extraction of key frames, whose representation in terms of low-level features constitute the basic elements for the generation of the SVM data sets, are generally performed in an automatic way. This approach may produce several noise data, and it is therefore desirable to find a removal strategy. Noise key frames are usually detected when video includes color bars, test cards or other homogeneous frames. Duplicated key frames, generated when video is steady for a long while, also need to be removed. In this paper we propose a data clustering method that performs an automatic pre-processing of SVM data sets, to minimize the presence of noise. Our experiments show an example of classification of historical sport video clips, demonstrating that the proposed pre-processing strategy improves the overall performances of SVM.