Segmentation von Szenen in historischen Dokumentarfilmen

The goal of scene segmentation is to split a movie into separate units of action. Current algorithms use features that are not available in historical artistic documentaries either due to film composition or technical aspects. This master thesis introduces a new method that is specifically developed for historical artistic documentaries. The first step of scene segmentation is the detection of shot boundaries. For each shot that is found a keyframe is selected. The keyframes are compared by image features. Two shots that lie in-between a certain time span and exceed a limit of similarity belong, as all the shots that lie between these two shots, to the same scene. The features the algorithm uses to compare the keyframes are SIFT Keypoints, the Edge Change Ratio and block-based histograms.
Core scenes are the result of the previous step - the identification of similar shots. Between the core scenes a few shots exist that cannot be assigned. These loose areas are classified by recursively decreasing the similarity thresholds for the respective features. Another part of the thesis is devoted to the evaluation of the features with differing thresholds for a quality assessment of the features. The quality of a feature is defined by a weighted ratio between correctly and falsely classified similar shots. Furthermore, the influence of the keyframe selection strategy is investigated.
The implementation of the method is tested with historical artistic documentaries and modern movies. This allows for evaluating the implementation's ability to solve the given problem of scene segmentation in historical artistic documentaries as well as in modern movies. Finally, the proposed method is compared with other state-of-the-art techniques.