Research On The Detecting System Of Advertisement Logo In Videos Based On Convolutional Neural Network

As a new form of video ads,both the media operator and media regulator need to know the occurrence time,location,and duration of such advertisements in the video programs.The former needs to confirm whether the ads are served in line with the provisions of the contract,the latter needs to regulate the behavior of advertisers.Therefore,the detection technology of video advertisement logo and the research of detection system are conducive to the healthy development of the video media advertising industry.Due to the characteristics of different shapes,different sizes and complicated background,it is difficult to select an appropriate feature to represent different categories of advertising logos in videos at the same time.In order to avoid explicit feature extraction and improve the accuracy of classification and generalization ability of the algorithm,this thesis proposes a video advertising logo detection method based on convolutional neural network and constructs the corresponding detection software.Through the actual detection of the advertisement logos with different deformations in different videos,and compared with the commonly used detection algorithm,the experimental results show that the proposed method has a high generalization efficiency and a high detection precision under certain detection efficiency.And it has a certain degree of robustness to hollow,block,shift,rotate,zoom and other deformation.Furthermore,it can also maintain a very low miss rate and false detection rate when the video is in poor quality,or the image is not clear enough.The main contents are as follows:1.For the hierarchical structure,small size and other characteristics of the video advertisement logo,this thesis proposes an efficient graph-based image segmentation algorithm combined with the selective search strategy to obtain the candidate windows for the target advertisement.Compared with the sliding window method,the number of candidate windows is greatly reduced,thereby improving the detection efficiency.2.For the video logo ads exist in the hollow,block,zoom,rotation and other characteristics,we propose an implicit feature extraction and class judgment method for the target candidate window by convolution neural network.Compared with the explicit feature extraction method,the convolution neural network can automatically obtain the essential characteristics of the target advertisement from the sample,thus improving the detection accuracy.3.According to the theory of software engineering,the video advertisement logo detection system is designed based on the data flow method.The system has a reasonable module design,as well as friendly human-computer interaction interface.It can help advertising operators and media regulators on the placement of advertising logo to conduct a comprehensive monitoring.