【作者】 刘建立；

【导师】 左保齐； Xianyi Zeng； P.Vroman； B.Rabenasolo；

【作者基本信息】 苏州大学 ， 纺织工程， 2010， 博士

【摘要】 本文首次提出采用小波纹理分析和鲁棒贝叶斯神经网络方法对不同梳理工艺下生产的,以点轧热熔加固的5个不同等级的非织造材料外观质量等级进行识别,为基于智能技术的非织造材料外观质量评价和识别进行了探索性研究。本文以实现非织造材料外观质量的客观、准确识别为最终目标,主要研究了Besov空间多小波域消噪算法、小波纹理分析和鲁棒贝叶斯神经网络在非织造材料图像消噪、纹理特征提取和外观质量等级识别中的应用。?根据非织造材料不同区域纹理光滑特性具有显著差异的特点,本文提出了Besov空间多小波域非织造材料图像消噪方法,并从理论与实验上深入研究了小波域个数与消噪效果、耗时之间的关系。创新地提出以凸集映射前后高频小波系数差的平方和差分小于等于零为Besov空间多小波域消噪算法收敛准则,该准则有效地控制了算法的复杂性,提高了运算速度。相比于小波通用阈值消噪方法,Besov空间多小波域消噪方法不仅可以有效消除噪声,峰值信噪比大于等于40dB,还可以保持非织造材料图像纹理的光滑特性,有效抑制了过消噪现象。以VisTex数据库中的6幅典型纹理图像为参照,采用Tamura纹理特征中的对比度参数和Fourier功率谱图来度量和确定非织造材料图像纹理的整体对比度和方向性,研究了非织造材料图像的纹理特性。以点轧热熔加固5等级非织造材料为研究对象,建立了非织造材料图像小波纹理分析模型,系统讨论了高频小波系数1范数L1、2范数L2能量基特征和广义高斯分布控制参数特征κζ在非织造材料纹理特征表达中的应用,两类纹理特征可以在不同尺度和方向上对非织造材料图像的纹理进行刻画和描述。首创性地提出了基于1-近邻分类器的小波纹理特征质量评价方法,以此衡量不同小波基在非织造材料纹理特征提取中的性能差异。最后,以非织造材料的小波纹理特征为基础,采用鲁棒贝叶斯神经网络对5等级625幅试样进行外观质量等级识别。深入研究了鲁棒贝叶斯神经结构设计、权值优化、孤立点概率估计和网络模型选择等问题,详细探讨了基于UCMINF(An Algorithm forUnconstrained, Nonlinear Optimization)算法的鲁棒贝叶斯神经网络权值优化,基于小样本集推理原则和结构风险最小化原则的神经网络结构设计,以及根据网络识别精度和证据框架理论的最佳网络模型选择。在网络结构设计与模型选择方面,本文创新地提出了鲁棒贝叶斯神经网络结构设计极限风险准则和基于网络识别精度和证据框架理论的最佳网络模型选择方法,并将它们成功应用于鲁棒贝叶斯神经网络的结构设计与网络模型的评价和选择。其中,采用coif3小波基对非织造材料图像进行4层分解,提取12个L1能量值作为纹理特征对隐含层神经元个数为3的鲁棒贝叶斯神经网络进行训练和测试,该网络对125个测试样本的正确识别率为99.2%。研究表明,从Besov空间双小波域消噪后的非织造材料图像提取两类小波纹理特征对鲁棒贝叶斯神经网络进行训练和测试,以识别5等级625幅试样的质量外观等级的方法具有较高的正确识别率和较强的鲁棒性,该方法是有效可行的。更多还原

【Abstract】 To explore the evaluation and inspection of visual quality of nonwovens using intelligent methods, an algorithm is proposed originally that integrates wavelet texture analysis and robust Bayesian neural network to identify the visual quality of nonwovens coming from five grades that are formed with different combing and thermal bonding processes. The final target of this work is to recognize the visual quality grade of nonwovens objectively and precisely, which is realized through three stages, nonwoven image denoising, feature extraction and reorganization by implementing the multiple wavelet basis image denoising using Besov projections algorithm, wavelet texture analysis and robust Bayesian neural network.According the characteristic that the smoothness of nonwoven texture in different domain, the multiple wavelet basis image denoising using Besov projections algorithm is used to eliminate the noise within nonwoven image, and the relationship between the number of wavelet bases and PSNR (peak signal to noise ratio), computation time is also researched deeply from the theoretical and experimental vision. A convergence criterion for multiple wavelet basis image denoising using Besov projections algorithm is proposed that takes the difference of the square sum of the high frequency of wavelet coefficients to be no larger than zero. Compared with denoising method using common thresholding, the proposed method in this paper not only eliminates the noise effectively, e. g. the PSNR is no less than 40 dB, but preserves the character of various smoothness and prevents from over-denoising phenomenon obviously.One of the Tamura textural parameters and the power spectrum of Fourier transform are used to evaluate the contrast and direction of the nonwoven texture, and the results are compared with the ones of the six images of VisTex database. In this research, the texture analysis model of nonwovens is established, and two feature extraction methods are explored deeply based on energy features L1 and L2 of wavelet coefficients and the two control parameters of generalized Gaussian density model, i.e. the scale and shape parameter,κandζ, which can describe and portray the texture of nonwoven image at different scales and from various directions. Additionally, the recognition accuracy R of 1-neighhour classifier is considered as an evaluation parameter of the discrimination of the texture features extracted with the 2 schemes.Finally, with the two types of texture features, the robust Bayesian neural network is implemented to recognize the visual quality of 625 nonwoven samples belonging to 5 different grades. In this part, the structure and design of robust Bayesian neural network, weight optimization, outlier probability estimation and model selection are researched in system, especially the optimization of weights with UCMINF (An Algorithm for Unconstrained, Nonlinear Optimization) algorithm, the structure design method based on inference principle of small samples and structural risk minimization principle, and the optimal model selection with recognition accuracy and the evidence framework theory are also introduced in details, respectively. In the aspects of structure design and model selection of robust Bayesian neural network, the criterion of maximum risk and the standard that combines the reorganization accuracy and log-evidence are explored initiatively. For example, when decomposed at 4 levels with coif3 wavelet base and twelve L1 energy features are used as inputs of the optimal robust Bayesian neural network with 3 hidden neurons, the recognition accuracy of 125 test samples is 99.2%.The research results from the 625 samples from 5 different indicate that the method for visual quality reorganization of nonwovens based on wavelet texture and robust Bayesian neural network has high accuracy and robustness, which is feasible and valid.更多还原