Abstract

In the dissertation a new method for improving the quality of classifications of images in video streams has been proposed and analyzed. In multiple fields concerning such a classification, the proposed algorithms focus on the analysis of single frames. This class of algorithms has been named OFA (One Frame Analyzed).In the dissertation, small segments of the video are considered and each image is analyzed in the context of its closest neighborhood, which is defined by a shifting time window. The class of algorithms representing such an approach has been named FSA (Frame Sequence Analyzed).Experiments on a number of video streams of different types have confirmed that the FSA method improves the classification results by reducing the level of error on average by 20%. Two variants of FSA algorithms have been analyzed: iFSA - which considers only OFA decisions, and fFSA - which considers OFA decisions as well as the similarity between the analyzed frames. Furthermore, the variants differ in terms of their computational complexity. The analysis of the proposed FSA algorithms included different configurations of decision functions, multiple similarity measures, as well as method parameters such as: the window size, the significance weight distribution parameter or the decision acceptance threshold. The FSA algorithms have been evaluated in terms of those attributes, which has proven their applicability in terms of the type and intensity of distortions in the video stream. Furthermore, the performed tests have confirmed the effectiveness and versatility of the FSA method.

Author (1)