A framework for TV logos learning using linear inverse diffusion filters for noise removal

Different logotypes represent significant cues for video annotations. A combination of temporal and spatial segmentation methods can be used for logo extraction from various video contents. To achieve this segmentation, pixels with low variation of intensity over time are detected. Static backgrounds can become spurious parts of these logos. This paper offers a new way to use several segmentations of logos to learn new logo models from which noise has been removed. First, we group segmented logos of similar appearances into different clusters. Then, a model is learned for each cluster that has a minimum number of members. This is done by applying a linear inverse diffusion filter to all logos in each cluster. Our experiments demonstrate that this filter removes most of the noise that was added to the logo during segmentation and it successfully copes with misclassified logos that have been wrongly added to a cluster.