基于人耳主观反应的听觉特征量及其在目标识别中的应用

【作者】 王娜；

【导师】 陈克安；

【作者基本信息】 西北工业大学 ， 声学， 2006， 硕士

【摘要】 水下目标自动识别是现代海战的重要环节，特征提取是目标识别中的关键。为了得到稳定有效的识别特征，传统的特征提取方法主要是通过信号处理与变换技术来提取信号内蕴含的有效特征，虽然这些技术的发展很快，但复杂环境下的目标识别率始终不理想。为此，本论文通过对人耳听觉机理的研究，针对水下目标分类识别，从心理声学的角度寻求基于人耳主观反应的新的目标识别特征，为水下目标识别中的特征提取开辟新的途径。 论文通过对人类听觉系统的深入了解，研究了人类听声辨物的过程与机理。人类听觉系统的特殊结构决定了人耳对声音的频率分辨、音调识别、强度分辨、时间延迟等听觉特性，而人对声音的主观感受主要由响度、音调、音色等特征来描述，同时又受到掩蔽效应、优先效应等特性的影响。因此，心理声学研究者提出用上述心理声学参数客观描述不同声音造成的主观感受，用它们定量地反映听觉感受的差别。为此，本论文深入研究了响度类特征和音调特征的物理含义和计算方法，并针对声学目标，尤其是水下目标识别进行了改进。 考虑到人类听觉系统识别声音时，人脑神经系统起到关键性的作用。因此，本论文建立的目标识别系统选择近似于人类大脑的人工神经网络分类器进行分类识别。针对本论文的实验研究，设计了前馈BP三层神经网络分类器。 对实测的水下目标辐射噪声、汽车噪声和无线电通信噪声进行了目标识别的实验研究，寻找有效识别特征，得出结论：人耳听觉特征量中的特性响度、特性尖锐度、总响度和总尖锐度可作为水下目标识别的有效特征，反映了目标声音的强度特性，在一定程度上能提高水下目标识别率；音调作为相关特征，补充了响度类特征中不包含的声学目标的频率特征，与响度类特征结合使用可以提高识别率。 在不同类型目标分类识别研究中，在以特性响度为识别特征的前提下，利用遗传算法进行特征选择，既降低特征维数又提高了目标识别率。深入研究了样本时间长度和不同频率分析尺度等因素对识别结果的影响，发现用临界频带和1／3倍频程频带分析得到的特性响度特征最能有效识别目标，而临界频带具有普遍的通用性。同时对训练样本和测试样本的组合方式加以改进，使识别过程更接近实际情况。更多还原

【Abstract】 Automatic underwater targets recognition has long been an important research topic in the field of acoustic signal processing. Feature extraction is the key step in the underwater targets recognition. Unfortunately, traditional approaches using signal processing and signal transformatinon technique applied to target recognition have drawbacks. A feature extraction approach based on auditory properties and psychoacoustic model is proposed to enhance underwater target recognition ability in this thesis and the research work focuses on the applications of auditory theory into acoustic targets recognition.Firstly, the features of the human auditory system are investigated, expecially hearing process and hearing perception mechanism. Human auditory system’s special configuration effect the characteristics of sound in ear, such as frequency distinguish, pitch recognition, strength distinguish, time delay and so on. Human’s subjective perceptions can mainly be described by loudness, pitch and timber, which can be calculated based on Zwicker’s theory, and pitch period is estimated by the dyadic wavelet transform. Meanwhile, human brain plays a very important role in human auditory recogniton system. Therefore three-layer back-propogation neural network classifier is used to recognize the acoustic target in the thesis.Secondly, in order to find effective features, aimed at ship-radiated noise, automobile noise and radio noise, target recognition is examined experimentally. The conclusions are as follow: 1. Specific loudness, specific sharpness, loudness and sharpness, the human hearing features based on psychoacoustics are effective features in the targets recognition. 2. Pitch as a correlation feature, reflects the acoustic targets’ frequency characteristic. 3. Loudness feature combine with pitch feature is used to improve the target recognition.Finally, specific loudness feature is selected by genetic algorithm, which can reduce the features’ dimension and enhance the accuracy rate. And from more experiments, it is concluded that samples’ time length and frequency analysis bandwidth affect the recognition results. It is shown that from such experiments that the specific loudness based on critical-band rate and 1/3 octave-band rate is most effective for recognizing acoustic targets. Furthermore, the critical-band rate analysis is applicable for the ordinary targets recognition.更多还原