The Compression and Fusion Technique Research of Underwater Target Feature
|School||Harbin Engineering University|
|Keywords||underwater target recognition feature extraction feature compression feature fusion|
The analysis of echoes is an effective method to detect targets in long range. In the field of active detection, the echoes received consist of target signals and reverberation. These two kinds of signals have different properties in time-frequency domain, respectively, so they can be classified by time-frequency analysis methods. Original features are not suitable for classification directly because of higher dimension. And each time-frequency analysis method focuses on one part of the properties of signal, so it is often necessary to utilize feature compression and fusion so as to increase the accuracy of classification.In this paper, the criterion of separability between target signals and reverberation in the feature space is discussed. Based on the highlight model, target signals have regular distribution on the time-frequency plane while reverberation doesn’t. After extracted time-frequency feature, the samples of target signals and reverberation have different regions in the feature space. The smaller the overlap is, the easier the classification is. In this paper, the ratio of between-class scatter distance to within-class scatter distance of feature samples is used as the criterion of separability.From the principle of maximizing the separability, this paper studied the nonlinear compression of time-frequency feature. Linear Discriminate Analysis (LDA) is a common method in feature compression, which transforms the original feature samples into several variables by linear combinations in order to reduce the dimension of features. This paper researches a nonlinear extend of LDA, which called Kernel Foley-Sammon Discriminate Analysis (KFSDA). This method can compress time-frequency features with the nonlinear information which included in feature samples and cannot be extracted by LDA.This paper studied the time-frequency feature fusion method based on the improved serial fusion. The serial feature fusion is a method with the most universal meaning, while it is not used widely in practice as the correlation between original features is weak. This paper deals with the feature samples before fusion by Canonical Correlation Analysis (CCA), which integrates two kinds of feature samples into several pairs of variables with strong correlation between them. Therefore, the fusion features created by the method in this paper will get more identifying information than traditional method.All the methods adopted in this paper are validated by the experiment data from sea trial. The processing result shows that the feature samples which have processed by feature compression and fusion have stronger separability than original feature samples. In this paper, Support Vector Machine (SVM) is used to classify target signals and reverberation. Compared with original feature samples, the classification accuracy of fusion features is increased obviously.