Automatic TV logo detection and classification

Abstract

Television (TV) logos are the only semantic objects that appear commonly in all TV broadcast videos. And they can be utilized in the development of many useful applications such as TV commercial detection, and audience measurement. In this study, we have developed an automatic TV logo identification system. The proposed TV logo identification system consists of two parts, namely, TV logo detection and TV logo classification. In the TV logo detection part, we utilized from the idea that ‘the broadcast video content is changing over time except the TV logos’ and we used time averaged edges method to obtain static regions (TV logos) in broadcast videos. In the TV logo classification part, we have used Support Vector Machine(SVM) as classifier. We have compared some well known subspace analysis methods such as Principle Component Analysis (PCA), Non-negative Matrix Factorization (NMF), Independent Component Analysis (ICA), and Discrete Cosine Transform (DCT) to find best feature to describe TV logos. Before applying the subspace analysis methods, all logo images are converted into a fixed size representation by using Grid Descriptor(GD) method. For classification experiments, a TV logo DB of 3040 images is constructed from 152 different TV channels. The best classification performance is obtained by ICA2 with an accuracy rate of 99.21%. For the logo detection and identification experiments, we have collected 240 videos from the 12 most popular TV channels of Turkey. The proposed system achieves to 99.17% logo detection rate and 96.03% average accuracy rate for logo identification. Results of the experiments show that the proposed TV logo identification system works with high accuracy rates and can be utilized in an audience measurement process.