超声波技术处理含油污泥研究进展

以Indian Pine数据集为研究对象,利用等角特征映射对其进行特征提取,然后选取BP网络传统的梯度下降训练方法和正切拟牛顿法、Polak-Ribiere共轭梯度法、Levenberg-Marquart法3种数值优化的训练方法对其进行分类. 对分类结果进行对比分析,结果表明:基于数值分析的训练方法训练网络的耗时均比梯度下降法长,但收敛效果更好;总体分类精度均比梯度下降法至少提高6%;各类别制图精度都较高,且较为稳定,而梯度下降法只对易分的类别精度高;3种数值分析训练方法中,正切拟牛顿法和Polak-Ribiere共轭梯度法的收敛效率和分类精度比 Levenberg-Marquart法高.

Comparative Analysis of Several Training Methods of BP Network on Hyperspectral Image Classification

This paper study on hyperspectral image classification using BP network trained by numerical analysis methods such as OSS,CGP,LM and traditional gradient descent methods ,which based on Indian Pine dataset taking fea-ture extraction using isogonism characteristic mapping methods .Comparing the classification results ,it shows:the classifi-cation result using numerical analysis methods is better than GD methods .Firstly,the training time using numerical anal-ysis methods is longer than GD ,but the convergence effect is better .And then ,the total accuracy improves 6%at least a-gainst GD.In the end,the mapping accuracy of each category is better and more stable than GD ,which is only sensitive on easily classified category .At last ,Comparing the three numerical analysis methods ,the OSS method and CGP method is better than LM method .