Research On Advertisement Logo Detection System In TV Media

In the modern TV media fields, with the advance ofcommercialization, a new media advertising element-advertisement logois gradually known by the public and has achieved good market effect.Advertisement logos refer to the identifications hanging in the edge ofscreen when TV programs broadcast. As a new form of TV commercials,media regulatory authorities need to obtain the information about time,length and location of advertisement logos, in order to regulate theconduct of advertisers. The operator also needs to obtain the informationthat, for example, if such a kind of advertising was put in the specifiedtime, in order to verify compliance with the contractual requirements.Therefore the research on key technologies of advertisement logodetection and its system construction is the necessary guarantee for thehealthy development of TV media industry, and has a certain theoreticalsignificance and important practical value.Advertisement logos usually accompany with TV programs whilehanging in the edge of screen. According to the requirements ofadvertisers, advertisement logos vary in shape features, such as irregularcontour, translucent, hollow and etc. On the other hand, due to the characteristics of TV programs, semantic and image content betweenadjacent video frames may exist obvious difference, so the detection ofadvertisement logos can be easily disturbed by the cluttered backgroundas well as its frequent transformations. While the method for detectingadvertisement logo can learn from the research production of imagerecognition, but due to its own particularity, targeted research is stillneeded. Based on the relevant literature and a large number ofexperiments, this thesis conducts the research from both retrievalperformance and retrieval efficiency. After put forward the advertisementlogo detection algorithm in TV media, the software for the system is alsoimplemented. The following points highlight the main work of this thesis:1. Considering that advertisement logo is generally embedded in avideo frame with complex and even cluttered background, this thesisproposes a two stage heterogeneous matching method based on globalfeature. In the first stage, the proposed algorithm establishesmatching-template that represents shape feature of image online byintroducing a self-adaptive learning method. The matching-template isupdated and established in the detection process. By ignoring score in thebackground, interference caused by cluttered background as well as itsfrequent transformations can be eliminated. In the second stage, rotationinvariant ELBP(Extended LBP), a texture descriptor is adopted so thatcandidate windows are further filtered, which makes up for the inadequacy of the first stage method. The experiments prove thatcombined with rotation invariant ELBP descriptor, the proposedalgorithm can significantly reduce the false detection in some casescaused by the similarity between the advertisement logo and column title,television logo and etc.2. While the localization method based on template matching willproduce a large amount of unnecessary computation, which results in itslimitation in object instance detection in videos. Based on thecharacteristics of correlation surface, block hill-climbing search isproposed. The algorithm locate target image with nonlinear traversalwhile improving the efficiency. In addition, SSE2instructions are usedfor accelerating the process of feature matching, and further reduce therunning time of algorithm.3. According to the theory of software engineering, theadvertisement logo detection system is designed by using the methodbased on data flow. The software system is consisted of friendly UI andreasonable module design, and with good scalability and robustness. Inaddition, detection results are trimmed by post-processing method to filterdirty data.Advertisement logo detection algorithm proposed by this thesis isconducted experimentally by datasets provided by a TV station as well asfetched by web crawler. The adaptability, accuracy and real-time processing ability of the proposed algorithm are validated by comparingthe detection results with the reference data of corresponding videos. Theconstruction of advertisement logo detection software system provides aneffective platform for advertisement logo monitoring and regulation.