| With the increasing demand for monitoring, it is urgent to introduce rich media (i.e., audio, image and video) into environment monitoring activities on the basis of sensor networks, thus perform a complete and accurate environment monitoring. In recent years, multimedia sensor networks (MSNs) have been paid wide attention. Audio semantic analysis plays an important role in detecting environmental characteristics and improving the detection accuracy. As a result, audio semantic extraction in monitor environment is one of the most important research fields in MSNs.This thesis aims at the problem of audio semantic analysis in MSNs, and proposes a series of novel methods from three different aspects:audio feature selection, basic audio element detection, and high level audio semantic analysis. The main contributions of this thesis are as follows:(1) Discriminating principal components analysis based feature extraction method. Audio feature extraction and selection are the basic problems for audio semantic analysis in MSNs. On the one hand, the selected features must represent the important classification characters of the audio informantion in both time domain and frequency domain. On the other hand, we should limit the dimension of the feature vector and reduce the redundancy of the features to save the posterior computing energy. In this thesis, we propose a discriminating principal components analysis based feature extraction method. This method is the combination of principal component analysis and linear discriminant analysis, which extracts the best representative audio features and enhances the discriminating ability through analyzing the categories of training samples. By using this method, we can extract the pivotal and independent features.(2) Weihgted association graph based element detection method. Audio element detection is the bridge to connect the low-level audio feature and high-level semantic content. In this thesis, we present a weihgted association graph based element detection method to detect the audio elements in MSNs. In the proposed method, we train the basic audio elements separately based on Hidden Markov Model, and combine them together by some prior knowledge in specific domains. Moreover, considering that the detection errors for different audio elements have different decision risk, we introduce the minimum risk Bayesian decision method to solve this problem. By using the proposed method, we can detect the basic audio elements accurately in MSNs with complex background and multi-audio events occurrence.(3) Neural network based method for high-level audio semantic analysis. High-level audio semantic analysis is important for the audio content understanding. The tranditional audio analysis method usually use statistical or rule-based approaches to extracting high-level semantic. However, in MSNs, the computation ability of sensor node is limited and the background noise is complicated, these methods designed for TV programs can not be deployed in the MSNs directly. In this thesis, we propose a neural network based method for high-level audio semantic analysis. With the neural network based approach, human knowledge and machine learning are effectively combined together in the semantic inference. By using this method, we can detect the high-level audio semantic content in MSNs correctly.(4) Verification system. Based on the MSN propotype system designed by our laboratory, we design and implement the verification system for our proposed algorithms. We deploy the discriminating principal components analysis based feature extraction method, the weighted graph based audio element detection method and the neural network based high-level audio semantic analysis method on this verification system, and then, we utilize the verification system to analysis the audio semantic information in the traffic and the ceremony environments. The experimental results of verification system demonstrate that the proposed methods can analyze the semantic informantion in the MSNs accurately.
|
PDF Full Text Request