Multimedia Retrieval Based On The Statistical Analysis Of Heterogeneous Characteristics

From early last century, content-based multimedia retrieval has become a research focus in the area of information retrieval. The traditional content-based multimedia retrieval mainly focused on the retieval of a single type of multimedia objects, like content-based image retrieval, content-based Audio retrieval and content-based video retrieval.But the retrieval above has largely failed to support the mutual retrival between different types of multimedia objects. Psychology researches find that human brains need to process different senses of information synchronously to know the world similarly,to form a overall understanding of multimedia semantics, multimedia retrieval must can cross different types of multimedia data freely. This is so-called cross-media retrieval metioned in this article.Cross-media retrieval is a new retreval technology which is originated from the content-based multimedia retrieval.content-based multimedia retrieval has been proposed since the early nineties, and it is always a very active area of computer vision research direction, the research integrated the applications of statistical analysis, machine learning, pattern recognition, human-computer interaction, multimedia databases and other fields to solve the problem of traditional text-based multimedia retrieval which faced the shortcoming of wasting time and effort ,and large defects in the subjective differences. content-based image retrieval was proposed firstly, we extract the image color, texture, shape and other visual features from the image to constitute a feature vector used as the image index. This technique was also applied to content-based video retrieval and audio retrieval.Most of these methods achieved good retrieval results, but they did not support the retrieval which can cross different types of multimedia objects flexiblely, cross-media retrieval is just based on the idea above. The research goal of cross-media retrieval is to enable computer to integrated process and retrieve images, audio and other digital visual and auditory information, and finally build an overall understanding of multimedia semantics for us, break through the traditional method's restrictions in multimedia types in our retrieval process.Intrinsically, the fundamental challenge in cross-media retrieval lies in the heterogeneity of different low-level features.First, canonical correlation between meida objects of different modalities is explored.An isomorphic subspace is constructed based on the analysis of bost visual features and auditory features, and the mapping process maximally preserves initial correlation unchanged. Also polar corordiantes are used to judge the general distance of media objects with different modalities in the subspace. when you submit one type of multimedia data sample, you can retrieve different types of multimedia objects which have similar semantic.