基于SVM的多方位声散射数据协作融合鱼分类与识别

Multi-azimuth Acoustic Scattering Data Cooperative Fusion Using SVM for Fish Classification and Identification

为解决基于声散射数据的鱼分类与识别问题,提出了一种基于SVM的多方位声散射数据协作融合鱼分类方法。首先,提取多方位声散射数据的小波包系数奇异值、时域质心及离散余弦变换系数特征,并进行特征融合;然后,采用支持向量机(SVM)分类器对每个方位提取的特征做出决策,并将决策结果表示成后验概率的形式,同时利用每个方位的决策概率对其他方位的决策进行加权;最后输出分类结果。采用3类鱼作为研究对象,得到不同方位数量条件下基于协作融合方法的分类正确率最终达到92%以上。试验数据处理结果表明,随着方位数量的增加,总体分类正确率呈升高的趋势,基于SVM的协作融合方法可以有效提高分类正确率。

In order to solve fish classification and identification problems based on acoustic scattering data, a data fusion method based on SVM posterior probability was deduced, and a multi-azimuth acoustic scattering data cooperative fusion fish classification method based on support vector machine （SVM） was proposed. Firstly, the wavelet packets coefficients singular value feature, temporal centroid feature and discrete cosine transform coefficients feature using multi-azimuth acoustic scattering data were extracted, which reflected acoustic scattering characteristics of fish from different aspects. Secondly, the SVM classifiers made the decisions for features of each azimuth and the results were expressed in the form of posterior probabillty, each azimuth decision probability was used to weight the decisions of other azimuth simultaneously. Finally, the classification results were the ultimate output. Three kinds of fish were selected as the research objects and the classification accuracy （ more than 92% ） was presented based on the cooperative fusion method under the conditions of different numbers of azimuth. The processing results of experimental data indicated that the overall classification accuracy showed an increasing trend with the increase of number of azimuth. To examine the performance of classification further, large carp samples and small carp samples were used as training and testing samples mutually. The classification accuracy showed a increasing trend with the increase of number of azimuth in both cases, which reached more than 90% ultimately. The multi-azimuth acoustic scattering data cooperative tusion method based on SVM can improve the correct classification ratios effectively.