TV logo classification based on convolutional neural network

This paper presents region-based, convolutional neural network for accurate and efficient TV logo classification. Although many previous methods have been applied to TV logo classification, most of them only process the video images. For images captured by smartphone, these methods often can't achieve satisfactory performance. Three contributions are given to the problem of TV logo detection and classification. Firstly, maximally stable extremal region (MSER) is used as a method of generating candidate boxes per image. Geometry constraint is introduced to filter certain boxes which are obviously different from TV logo shape in geometry space. Secondly, we design an efficient Convolutional Neural Network (CNN) for classifying TV logo. Thirdly, we create a large TV logo dataset, all of which are collected from video streaming. Some augmentation methods have been taken to improve the diversity of TV logo images. Experimental results show that the proposed method outperformed previous methods.