| Summary: | 碩士 === 國立臺灣科技大學 === 電機工程系 === 99 === With the increasing network resource and services, e-commerce and online advertising are more and more popular and the number of video sharing websites also shows in explosive growth. Consequently, the frequency of online shopping and the time of watching online videos increase significantly too. Many advertisers do their advertisements through video sharing websites, hoping to get the maximum revenue due to large amount of website users. On the other hand, traditional advertising mechanism is unable to meet the rapid growth of e-commerce and online advertising markets. Therefore, this thesis proposes a content-based image retrieval approach for a product recommendation system by analyzing video contents so that the recommended products can be as matching as possible to a user’s needs.
Two methodologies are proposed by this thesis to improve the recommended result based on video content. Firstly, a color region matching that uses color and region information to enhance the local feature image matching result of ASIFT. By analyzing the color similarity degree, the density of matching points and geometric consistency of regions to overcome the drawbacks of lacking object characteristics and semantics. Secondly, a mechanism that calculates the recommendation value of a product based on the content similarity, the frequency of appearance and the importance of keyframes of the product. These features exist between the video keyframes and product images.
Finally, the study uses three kinds of evaluation standards to analyze and compare the performance on the precision and rank order of experimental results. The experimental results show that the proposed approach in this thesis have effectively improved the overall recommendation result on increasing the similarity between product advertising and video content and enhancing the users’ interest on viewing the product advertising.
|
|---|
