【作者】 李建锋；

【导师】 邹北骥；

【作者基本信息】 中南大学 ， 计算机应用技术， 2013， 博士

【摘要】 摘要：结合人类视觉机理的图像处理方法与应用是目前数字图像处理技术的热点领域,研究的思路一般为根据人类视觉机理、现象构建数学模型,并将数学模型用于相应的图像处理任务,目前研究大致分为三类：一是人类视觉感知信息的表示与建模；二是人类视觉神经元机理及其运行机制建模；三是视觉皮层功能机理及其信息处理机制建模。脉冲耦合神经网络(Pulse Coupled Neural Network:PCINN)作为视觉皮层神经元模型的典型,由于接近人类视觉神经元机理及其运行机制,成为研究基于人类视觉神经元机理及其运行机制的图像处理方法的重要手段。但是,PCNN无法通过数学方法清晰的描述内在特性,使其在图像处理中无法充分发挥模型的优势,如何解决该问题对于提升PCNN图像处理的性能,进一步完善基于人类视觉机理的图像处理新方法意义重大。本论文针对PCNN存在的问题,在分析视觉皮层神经元工作机理的基础上,以PCNN为研究对象,从解决PCNN存在的问题,提升PCNN图像处理性能的角度,进行以下三个层次的研究：1)PCNN关键参数的特性及其在图像处理中的应用研究,2)结合任务多种特征的PCNN图像处理方法研究与应用,3)自适应PCNN及其在图像处理中的应用。主要的研究内容和贡献有以下三个方面：1.针对标准PCNN在图像处理中参数设置没有规律以及参数多、计算复杂等问题,提出了一种改进PCNN及其关键参数设置的图像边缘检测方法。该方法改进并简化了标准PCNN模型使其更符合边缘检测,同时在边缘检测中针对改进PCNN的关键参数提出了不同的设置方法。该方法首先将标准PCNN的参数由9个简化至4个,然后采用以下方法设置改进PCNN的参数：(1)通过图像灰度值的局部特征设置连接系数β,(2)考虑神经元之间的灰度值差异设置权值矩阵,(3)结合局部梯度确定放大系数VE以及时间衰减常数αE,(4)结合最大方差比确定最佳迭代次数Ⅳ等。在实验中与同类型视觉皮层神经元模型标准PCNN算法、交叉视觉皮层算法进行比较,实验结果表明本章提出的改进模型以及关键参数的设置方法能够精确的提取图像边缘,提出的方法优于同类型的标准PCNN、ICM以及其他对比算法,提取结果符合人类视觉的感受。2.针对标准PCNN与图像处理任务之间的关系不明确,影响PCNN图像处理性能的问题,提出结合图像离散系数特征以及侧抑制特性的PCNN阴影检测方法,结合人类视觉亮度特性和图像亮度对比度特征的PCNN图像融合方法。提出的方法根据图像处理任务特征的数学描述建立与PCNN之间的联系,从而指导PCNN在图像处理中的运行状态,达到提升PCNN图像处理性能的目的。(1)在结合PCNN的阴影检测中针对PCNN进行以下改进,一是针对PCNN对灰度值相近但分属不同区域的像素区分能力弱的问题,引入人类视觉侧抑制特性改进PCNN模型；二是引入图像阴影离散系数特征指导PCNN的阴影检测过程；(2)针对结合PCNN的图像融合提出以下改进PCNN的方法：一是提出适合PCNN图像融合的Main-Auxiliary PCNN模型；二是通过图像融合常用的对比度和亮度特征建立PCNN模型与图像融合任务之间的联系,指导图像融合过程。3.针对标准PCNN数学方法无法清晰的描述,参数之间的协作机制不明,使得PCNN在图像处理中无法根据图像处理任务动态调整PCNN运行状态的问题,提出结合优化算法的PCNN性能提升方法并将提出的方法用于图像分割。具体应用中,通过免疫克隆算法优化PCNN,该方法利用免疫克隆算法理论要求弱的优势,将PCNN在图像处理中动态调整运行状态的问题转化为基于免疫克隆算法的优化问题,实现PCNN在图像处理中动态调整运行状态的目的。该方法首先在标准免疫克隆算法的基础上加入自适应操作和梯度操作,提高免疫克隆算法收敛速度和全局收敛性,然后在标准PCNN的基础上采用简化PCNN模型,将简化PCNN的参数定义为抗原,将图像分割结果的熵定义为抗体,通过一系列克隆变异机制动态调整PCNN的运行完成图像分割任务。实验中与标准PCNN分割方法、ICM分割方法、PSO-ICM分割方法、PSO-PCNN分割方法、ISCA-PCNN等同类型算法以及其他多种分割方法进行比较,实验结果表明提出的方法达到了PCNN在图像分割中动态自适应调整运行状态的目的,图像分割的性能优于同类型算法和其他对比算法。更多还原

【Abstract】 Abstract:Image processing approaches and their applications inspired by human vision mechanism have become one of most active topics in digital image processing field recently. The general framework is to develop a mathematic model for human vision mechanism and apply it to a specific image processing task. Generally, there are three principle categories in recent research:the representation and modeling for human vision perceptual information, modeling for mechanism of human visual neurons and its working mechanism, and modeling for the function mechanism of visual cortex and its information processing mechanism. As one of the most successful computational models, PCNN has become one of the most important accesses to studying image processing based on mechanism of human visual neurons and its working mechanism. However, PCNN’s intrinsic characteristics cannot be demonstrated with elegant mathematic methods, which laminate its application in image processing. Therefore, solving the above issue is of great significance to improve the performance of PCNN and motivate new image processing methods based on human vision mechanism.In order to solve the aforementioned issue, we analyze the working mechanism of visual cortex neurons. In this dissertation, the study object is PCNN, and our goal is to solve the aforementioned issue in PCNN and improve its performance in image processing. Our research mainly focuses on the following three aspects. First, we discuss the characteristics of the key parameters in PCNN and their applications in image processing. Second, we study the task dependent PCNN and apply it to image processing. Finally, we introduce the adaptive PCNN and its application in image processing. The key contributions of this thesis lie in three-fold:1. Aiming at the issues that the computation complexity of the standard PCNN is high and numerous parameters have to be set without any regulation when applying to image processing, we propose a novel method for image detection via the updated PCNN, mainly focusing on the key parameters. The proposed method improves the performance of the standard PCNN and simplifies it so that it meets the needs of edge detection in a more rational way. Meanwhile, in edge detection, difference parameter setting strategies are introduced according to the updated key parameters in PCNN. A four step method is proposed to decrease the parameter number of the standard PCNN from9to4:(1) setting the connection coefficient β according to the local gray level,(2) setting the weight matrix according to the dissimilarity between two neurons,(3) computing the amplification coefficient VE and time attenuation constants aE based on the local gradient,(4) deciding the optima number of iteration N based on the maximum variance ratio. Experimental results show that our method outperforms the standard PCNN and ICM with higher edge detection precision and our detection results meet the needs of human visual perception better.2. Focusing on the ambiguous relationship between the PCNN and image processing tasks, which adversely affects PCNN’s application, we present a novel shadow detection method, which combines the discrete coefficients of the image and the lateral inhibition of the PCNN. Meanwhile, a new image fusion method is proposed, where brightness features of human vision and image intensity contrast are considered. The proposed method focuses on the relationship between the mathematic formulation of the specific image processing task and PCNN to guide the running state when applying PCNN into image processing and improve the performance further. In terms of shadow detection based on PCNN, first we introduce the lateral inhibition to improve the discrimination ability to pixels with similar intensity from different regions. Then we introduce the shadow coefficient feature to guide the shadow detection. In terms of the image fusion based on PCNN, we first present Main-Auxiliary PCNN model, and then guide the image fusion via establishing the relationship between PCNN and image fusion by combining the contrast and intensity features.3. We propose an adaptive immune clone PCNN based image segmentation method to address the problem that the standard PCNN cannot be described definitely in mathematic language, which results that the running state of the PCNN cannot be adjusted to the specific image processing task. We formulate the image segmentation problem as an optimization problem of immune clone algorithm to change the PCNN running state dynamically. First, the adaptive operation and gradient are added into the standard immune clone algorithm to accelerate the convergence. Then the standard PCNN is simplified, and its corresponding parameters are viewed as antigen of biological immune system while the entropy of the segmentation result is the antibody. After a series of dynamic clonal variation, we finally obtain the segmentation result.We compare our methods with PSO-PCNN, standard immune clone based PCNN, PCNN, ICM, PSO-ICM and other methods, the experimental results indicate that our method can dynamically adapt to the segmentation task and outperforms the state-of-the arts.更多还原