Research On Semi-supervised Cross-media Feature Learning Methods Based On L_2,p Norm

Multimedia data with different types is rapidly growing on Internet.The retrieval of multimedia data helps to make full use of massive data on Internet.We propose two improved cross-media feature learning methods using l2,p norm based on existing cross-media retrieval methods in this thesis.The work is significant for the development of cross-media retrieval and multimedia data processing.The existing cross-media retrieval methods mainly use the mapping matrix to map multimedia data from initial feature space to a joint space,and then measure the similarity between media data with different types for cross-media retrieval.The main challenges of existing cross-media retrieval methods are as follow:The low sparsity of mapping matrix leads to unsatisfactory joint spatial representation of multimedia data,which affects the performance of cross-media retrieval;the multimedia data is manually labeled,and the labeling error has influence on multimedia retrieval models.We propose two improved cross-media feature learning methods based on existing cross-media retrieval methods in this thesis,and discuss efficient extraction methods of keyframes from video.The main contents includes:(1)We use/2,p norm to improve the overall sparsity of mapping matrix,in consider of the influence of mapping matrix of multimedia data from initial feature space to joint space on retrieval performance,to improve the efficiency and accuracy of retrieval.(2)We use l2,p norm to construct a joint space to suppress the influence of edge points and noises in datasets,which improves the stability and robustness of cross-media retrieval.(3)We propose an improved method for extracting keyframes from video using dynamic color histogram and fast wavelet transform histogram.This method uses mutual information to filter keyframes,which is helpful for retrieval of video data.The proposed two cross-media feature learning methods are tested on the public datasets XMedia and Wikipedia,and are compared with four classical cross-media retrieval methods.Extensive experiments show that the proposed cross-media feature learning methods improve the performance of cross-media retrieval to some extent.The proposed cross-media feature learning method can be used for applications such as network content monitoring,cross-media intelligence engine and network public opinion trend analysis.The work helps to extend the reach of cross-media retrieval and improve the processing of multimedia data.