Research Of Content-Based Large-Scale Audio And Video Data Search

This study presents my research work on Content-based large-scale audio and video retrieval. Based on the similarity between the query and reference da-ta, the research is divided in to three sub topics, namely low-level, middle-level acoustic and visual similarity based retrieval and high-level semantic retrieval. The existed difficult points in each sub-area are discussed and the possible so-lutions are described.In sub-topic of low-level multimedia retrieval, the non-supervised dupli-cate sequences detection is addressed and a video macro-segmentation system is presented in this paper.For topic of middle-level audio and video retrieval, based on the obser-vation, I refine the existent audio fingerprint approach via adding multi-scale information and using conditional entropy based feature selection algorithm to select a discriminative feature combination. Moreover, a hierarchical spatial verification scheme is presented to improve the video retrieval performance. And the works for TRECVID Content-based Copy Detection task are described in this paper.In addition, I try to solve the big problem of semantic gap by adopting Hu-man Computer Interactive scheme based on the Random Walk Algorithm.And I submitted a interactive run for TRECVID Instance Search.Finally, the experiments are conducted for the three topics respectively, showing that the test results of low-level and middle-level multimedia retrieval are quite satisfactory. And performance of interactive run is higher than auto-motive runs. Based on the absolute value of evaluation criteria, it shows that we sill need to further explore the research of how to bridge the semantic gap.