《弱信号检测随机共振机制的网络模型及应用研究》由会员分享，可在线阅读，更多相关《弱信号检测随机共振机制的网络模型及应用研究（59页珍藏版）》请在金锄头文库上搜索。
1、杭州电子科技大学 硕士学位论文 弱信号检测随机共振机制的网络模型及应用研究 姓名：耿丽硕 申请学位级别：硕士 专业：模式识别与智能系统 指导教师：范影乐 20091201 杭州电子科技大学硕士学位论文 I 摘 要 强噪声背景下的弱信号检测研究，是测量技术的热点和难点之一，在生物医学、测控、 以及军事等领域有着重要的应用前景。近年来，随机共振理论和实验研究的开展，为弱信号 检测提供了新的思路和方法。 目前基于随机共振的弱信号检测研究中，更多关注于单层或开环结构的模型。但在实际 应用中，此类模型容易受到背景噪声强度和信号幅值的影响。因此本文提出将多层和反馈结 构应用于随机共振模型中，对单层、双层以
2、及反馈结构的周期和非周期信号响应分别进行了 研究、仿真和比较；以图像复原为例，进行了随机共振弱信号检测的实际应用。本文主要工 作和研究成果如下： (1). 研究了 FitzHugh- Nagumo (FHN)神经元模型和双稳态模型的随机共振现象， 分析了周期信 号和非周期信号作用下的响应，验证了噪声对随机共振的作用，为后续实验奠定了基础； (2). 模拟神经系统中神经元之间的会聚方式，构建了双层 FHN 神经元网络。并采用互信息率 等评价方法，对单个神经元和双层神经元模型的随机共振性能进行了定量描述和比较， 分析了该网络在噪声环境中的信号检测能力。实验结果表明，与单个神经元模型相比较， 其检测
3、性能受噪声强度和信号幅值的影响较小，更适合于动态环境下的弱信号检测； (3). 为了降低开环网络在噪声强度多变的环境中对弱信号检测的不稳定性，本文提出将反馈 环节引入双层 FHN 神经元网络模型，以改善可变噪声背景下的弱信号检测性能。研究结 果表明，闭环神经元网络模型的随机共振现象要优于开环双层网络和单个神经元，能够 在更宽的噪声范围内反映输入信号的规律，提高了稳定性能； (4). 本文将随机共振机制应用于低信噪比图像复原中。在充分考虑图像像素空间相关性的基 础上，采用 0 和 180 Hilbert扫描法，对二维图像进行独立降维；利用双稳态系统的非线 性特性，通过添加特定强度的噪声，实现污染
4、图像目标信息的增强；最后对降维信号的 双稳态响应进行了决策和重构，实现了低信噪比图像的复原任务。实验结果表明，该方 法抑制噪声的能力较好，对细节的重现效果清晰。尤其对于被强噪声污染(噪声强度=300) 的图像，在主观视觉效果和信噪比评价上，与传统复原方法相比，具有较佳的性能。 本文研究成果表明在双层闭环网络结构中，基于随机共振的含噪信号检测具有更佳的稳 定性能。而在图像复原中的具体实践，显示了随机共振在弱信号检测领域具有良好的实际应 用前景。 关键词：随机共振，微弱信号检测，FitzHugh- Nagumo 神经元模型，双稳态模型，图像复原 杭州电子科技大学硕士学位论文 II Abstract
5、 Weak signal detection in the background of strong noise is one of the hot and difficult measure technology. It has an important application prospect in the bio- medical, measurement and control, as well as military and other fields. In recent years, stochastic resonance theory and experimental rese
6、arch provided new ideas and methods for weak signal detection. Currently, in the study of weak signal detection based on stochastic resonance, much has been concerned with single or open- loop model. However, in practical applications, these models are vulnerable to the effect of background noise in
7、tensity and signal amplitude. Therefore, this essay proposes multi- layer and feedback structure to apply to stochastic resonance model. The response of periodic and non- periodic signal with single, double and feedback structure were studied, stimulated and compared. Take image restoration for exam
8、ple, practical application of weak signal detection based on stochastic resonance is conducted. In this paper, the work and research results are as follows: (1) It is a base of follow- up experiments that the FitzHugh- Nagumo (FHN) neuron model and the bistable model of stochastic resonance phenomen
9、on are studied. The responses of periodic signals and non- periodic signals under the actions are analyzed and the role of noise on stochastic resonance is verified; (2) To simulate the assemble pattern of neurons system, a double- layer FHN neural network is built. Using mutual information rate and
10、 other evaluation methods, the stochastic resonance performance of single FHN nerve cell and double- layer FHN neural model is compared by quantitative description in the noise environment detection capability. Experimental results show that the impaction of its detection performance due to the nois
11、e intensity and signal amplitude is smaller, compared with a single neuron model, more suitable for dynamic environments weak signal detection; (3) In order to reduce the instability of open- loop network for signal detection in the environment where noise intensity changing frequently, this paper a
12、dded a feedback link for the double- layer neural network model of FHN. The research result shows that the stochastic resonance phenomenon of the closed- loop neural network model is superior to the double- decker open- loop network and a single neuron. It reflects the law of the input signal within
13、 the wider range of noise, and also improves the stability; (4) This article applies stochastic resonance mechanism to low SNR image restoration. Taking full account of the image pixel space correlation, based on the use of 0 and 180 Hilbert scanning 杭州电子科技大学硕士学位论文 III method, the dimensionality of
14、two- dimensional image is reduced independently; the polluted image target information is enhanced by adding a specific intensity of the noise with non- linear characteristics of bistable systems; Finally, two one- dimensional sequence of bistable response signals are reconstructed by decision- maki
15、ng. It realized the restoration of the low SNR image. The experimental result indicates that the ability of inhibiting the noise of this method is better, and the reappearance of details is clearer. It still has a better performance in the background of strong noise images (noise intensity = 300) re
16、covery in the subjective evaluation of visual effects and signal to noise ratio, as compared with traditional recovery methods. The research result indicates that signal with noise detection based on stochastic resonance has much better stability in the case of multi- layer and feedback network structure. What s more, the image restoration of the concrete practice shows good prospect of stochastic resonance in weak signal detection area for practical applications. Keywords: stochastic re
