【摘要】 如何从回声信号中抽取特征,是目标识别或无损检测的前提。实际物体,从其反射声学特性考察,大多呈多亮点结构并表现出各向异性,其回声特性主要随目标姿态角或声渡入射角的变化而呈混沌态变化。常规方法提取的回声特征,往往随入射角的变化出现严重的类内弥散与类间交迭,从而导致决策界面过分复杂、甚至完全失效。因此,在目标姿态以未知方式自然变化的情况下,如何提取适当的回波特征一直是国内外学者广泛关注的瓶颈问题。本文提供一种趋向不变识别的回声特征的神经网络自动提取方法和新的可分性测度Js与Jr。研究表明,它能直接利用回波信号自动生成具有最大可分性的紧致特征集。给出了典型的实验结果,并着重剖析了神经网络特征提取的信息蕴含及其物理意义。实验结果与理论分析的结论相一致。可望在自动目标识别、水下机器人视觉、图像处理、机械故障诊断、数据压缩、以及生物医学工程等领域获得应用。更多还原

【Abstract】 Two contributions are presented: A noval neural fasture anto-extraction scheme and its mherent principle oriented to invariant recognition using echoet, b) a pair of new data/feature separablity measures Js and Jr, whose advantage over the conventional separablity measures lies that. Js or Jr can express exactly the potential recognition ratio. Mathematically analysis and typical experiment results with underwaer targets are given. The scheme is proven to be suitable for automatic feature extraction directly based on return signals, especially for invariant recognition of orientation-dependent targets in sound fields, while the traditional feature extraction methods can hardly keep valid in these situstions.更多还原