| Video ads play a positive role in business prosperity. There is also the content of illegal, vulgar, exaggerated publicity in TV ads. The result is a bad influence in society, undermined the credibility of media in society. Meanwhile, the illegal proliferation of online video ads also showed the trend. In order to protect consumer interests, to create a positive atmosphere of the advertising industry, there is an urgent need of efficient technical means for TV and network video advertisement monitoring. Technical means of video ads monitoring must meet the requirements of accuracy and real-time. Since video fingerprint with characteristics of discriminability and robustness ,which can ensure the monitoring results to be accurate and reliable, easiness of extraction and high efficiency of fingerprint matching algorithm ,which can achieve real-time, Therefore, this paper choose to use video fingerprinting for the video ads monitoring, focusing on exploring robust video fingerprint extraction and efficient matching algorithm.Based on the comprehensive analysis of existing fingerprint methods, to against shortcomings such as weak robustness, inefficient or limited adaptability of matching algorithm , commenced exploration in the following areas:1. In the video fingerprint extraction stage, video fingerprint of multiple features is proposed which is combination of the image corner feature with temporal feature. The combination results better discriminability and robustness. We extracted corner features by improve Harris detector, and improved the temporal feature extraction method. The proposed fingerprint is proved to be strong robustness against attack like the common noise and change in brightness and imagine size. Working on compressed domain and relatively simple feature calculation ensure fingerprint extraction real-time.2. Next in the stage of fingerprint matching, the hybrid video fingerprint match scheme is designed for real-time working. In view of the similarity of sub-fingerprint sequences which is composed of element set similarity and element position similarity. In the first layer, we firstly weaken the problem of fingerprint matching as the problem of sub-fingerprint set similarity, then accelerate the sub-fingerprint matching by the modified two-layer locality sensitive hashing. Therefore, the similar sub-fingerprint will be matched as early as possible and thus massive calculation for comparison will be saved. Only if the set is similar in each other, then the typical matching method will be applied. Our approach combines the typical matching method's wide suitability advantage with the LSH's high effectiveness advantage.In summary, the proposed fingerprint system satisfies the video ad monitoring requirements of accuracy and real-time. |
PDF Full Text Request