Multimedia Content Analysis Based On Unsupervised Feature Learning

As one of the most important part of the Artifcial Intelligence, Com-puterVisionhasreceivedmuchattentioninrecentyears. Inthemeantime,multimedia is becoming indispensable in people’s daily life as the devel-opment of Internet. The amount of data is growing rapidly and it is almostimpossible to understand the content of all the data only by human forces.Basedonbothofthereasons, howtoletthecomputersunderstandthemul-timedia content automatically has become the focus of research. In orderto understand the multimedia content, learning representative features isvery important. In this work, we do research on how to extract features ofhigh quality from multimedia data.Diferentwithtraditionalfeatureextraction,wefocusonunsupervisedfeature learning algorithms which is to fnd and extract feature from multi-mediadataautomatically. Thispaperfocusesontwokindsofunsupervisedfeaturelearningmethod: sparsecodinganddeeplearning. Bothimageandvideo feature extraction have been studied and three unsupervised featurelearning algorithms are proposed.First, sparse coding is used to further explore the common structurein feature vectors extracted from images. Two important phases in Bag-of-Words model are improved and images can be represented more precisely.Second, deep learning algorithm are used for unsupervised feature learning in images. We improve the structure of current neural networkand incorporate Markov transition probability model.Third, deep learning algorithm is applied to multimodal learning invideos. Here, we use deep learning algorithm to fuse the image, motionandaudioinformationandprovideaprecisemodelforvideocontents. Thisalgorithm is applied to violence detection in videos and yields good per-formance.