Dynamic Scene Classification Based On Topic Models

Scene classification is defined as automatic annotation according to the semantics of agiven image for image data set. And it is currently a hot research direction in computervision. In this paper, we take the dynamic scene as the research object and adapt the bag ofvisual words-semantic topic modeling-semantic classification of dynamic scenes as themain line. We study the dynamic scene classification method based on topic models. Thestudy includes building dynamic scene visual dictionary, topic modeling by beliefpropagation based on word prior knowledge and the implement of semanticclassification for dynamic scene. The main work is as follows:1) Due to existing methods of topic semantic scene classification are mostly limited tostatic image sets, visual word generation method can not be directly applied to dynamicscenes. In full consideration of the spatial and temporal dynamic stochastic situation, wepresent the method of visual word generation for dynamic scenes based on SIFT (ScaleInvariant Feature Transform) flow. This method is based on the SIFT feature pointextraction, so it is robust to the scale, rotation, affine transformation. It has strongerrobustness for the local information of static image frames. In the process of flow fieldcalculation, this method can avoid the pixel gray value unchanged assumption intraditional optical flow. We use dense SIFT descriptors between adjacent image frames asflow field basis which can overcome the noise interference in flow field calculations. AndFor problem of losing position information of visual words in the quantization stage of thevisual feature, we adopt the thought of uniform blocking to quantitate visual wordsaccording to the position of SIFT flow. Experiments show that dynamic visual dictionarygeneration method proposed leads to good classification performance. The result shows10percent improvement in average accuracy rate in contrast with the method based ondifferential figure gray feature scenes descriptor.2) In dynamic scene classification, traditional topic model (PLSA, LDA) takes long training time and classification accuracy is not high. In this paper, we propose an improvedTMBP model named Knowledge-TMBP topic model by adding the word priori knowledgeto the TMBP model. This model use inverse document frequency as word priori knowledgeto rewrite the message passing in TMBP. Priori knowledge ensures that the importantvisual words are assigned with greater determination in topic inference.3) Application of semantic classification based on dynamic scenes. In this paper, weimplement dynamic scene classification based on dynamic visual words, TMBP model andKnowledge-TMBP model. Related experiments on the dynamic scene database which has14categories are performed. We analyze the performance comparison of PLSA, LDA,TMBP and Knowledge-TMBP four kinds of the topic models as well as low-level temporalfeature classification. The experimental results show that Knowledge-TMBP modeltraining time is shorter than PLSA, LDA and a little longer than TMBP. But it outperformsthree other methods in classification accuracy. Compared to the lower temporal featureSOE (Spatiotemporal Oriented Energy) method, the classification recall rate is increasedby2percent.