Supervised TV logo detection based on SVMS

In this paper a simple and practical technique for supervised TV logo detection based on support vector machines (SVMs) is shown. Specific TV logos are assumed to locate in static regions, usually four corners of one frame. Instead of sampling time averaged frames at frame level, we make use of information of shot boundary detection (SBD) to get three key-frames per shot. After extracting corners in every key-frame, we train several SVM classifiers for specific TV logos using color, edge, and key point features, and then detect logos in these regions of interest corners. At last, a two-step fusion strategy is performed to get optimum and robust performance at shot level. We tested more than 24 hours videos to detect logos of Eurosport TV station and achieved 99.98% of correct detection rate; and also more than 5 hours to detect logos of Cine Confidential, one program of Orange TV station with 99.99% of correct detection rate.