Real-time TV logo detection based on color and HOG features
Fei Ye, Chongyang Zhang, Ya Zhang, Chao Ma · 2013
This paper proposes a real-time TV logo detection algorithm that can detect logos embedded in TV videos or in the real-world videos/images. Unlike most existing TV logo detection methods, the proposed algorithm makes no assumption on temporal motion, spatial location, or any other visual view constrains on TV logos. The detection process consists of three stages: in the first stage, a color based region segmentation and candidate selection strategy is developed, which can narrow down the candidate search space and reduce computation cost significantly; at the second stage, SVM based classifier is trained, where geometric correction based on minimum rectangle bounding is used to improve the accuracy of the classifier, and affine transformation is adopted to construct a robust sample database; finally, the candidates are recognized by the trained SVM Classifiers using their HOG features. Experiments on several video sequences and logotypes have been carried out to verify the robustness and effectiveness of the proposed method.