Design And Implementation Of An Automatic Recognization System For Interpolation Advertisements

With the rapid development of the Internet advertising industry,video advertising is playing an increasingly important role in the advertising.The automatic identification of video advertising can help the advertiser to judge whether the advertisements are released according to the agreement,and also help consulting agency to analyze the business status of advertisers according to the amount of advertising.With the increasing demand for increasing advertising conversion rate,a new type of advertisements which is interpolated in the video appears in recent years,advertisers conceive the creative advertisements based on video story background and role setting,and make the advertisement into "video clips" interwoven in the video,which is called "Interpolation advertisements".The appearance of this kind of advertisements blurs the boundary between the advertisement and the video,making the traditional methods of advertising automatic identification is no longer feasible.This paper aims to slove the problem where the traditional methods of advertising automatic identification does not adapt to the interpolation advertisements in shot segmentation,shot classify and content recognition,designing an automatic recognization system for interpolation advertisements,which applies deep learning technology and traditional computer vision processing technology,integrates image features,audio features and text features and takes information in time and space domain into consideration.The main contributions of the thesis are listed as follows:(1)A new method of shot boundary detection is proposed for video intershot gradual transition.Because the scene of advertisement and video is similar,so there are more gradual transition shot between the video and the advertising,compared with the hot abrupt transition used in traditional advertising,it is more difficult to segment the shots.Based on the observation of the interpolation advertisements,we found that black shot would appear in the process of shot gradual transition,this paper breaks the conventional idea of calculating the inter-frame distance,and proposes a simple and effective solution to detect boundary for intershot gradual transition,which uses the color change trend.(2)This paper aims to slove the problem that the high similarity between video shots and advertising shots and difficult to distinguish,presenting a combination of LSTM model and Attention network to obtain timing sequence features existing in higher dimensions of audio and video and strengthen significant features to improve the classification performance.The traditional system uses CNN network to extract deep expression of audio and video features without using the timing correlation of video frame sequence,LSTM is used to get the correlation between the current frame and the next frame in this paper.Furthermore,the Attention network is used to get the weights of features in different dimensions in the results of shot classification in this paper,and the feature vectors with significant influence on performance are enhanced.Experiments show that the final classification accuracy can reach 88%,4%higher than the traditional machine learning method.(3)This paper aims to slove the problem that the indistintion of advertising logo leads to the invalidation in identifying the advertising content in which case the logo is extremely relied on,presenting a method of text recognition combined with audio feature matching to identify the advertising content.Based on the observation of the endogenous advertisements derived from video,a large number of the text replaces the logo to identify advertising products,there are also advertisements that don't even have text but sound to identify the product Therefore,this paper adopts the method of combining OCR text recognition with audio chromaticity diagram matching to identify advertising content,and the accuracy can reach 98%.The recognition system of interpolation advertisements proposed in this paper can not only detect the existence of interpolation advertisements in video,but also identify the advertising content,which has a strong practical significance and practical value.