【作者】 杜馨瑜；

【导师】 尧德中；

【作者基本信息】 电子科技大学 ， 生物医学工程， 2012， 博士

【摘要】 人类视觉系统具有非常优秀的图像处理能力。本文以目前广泛认可的人类视觉信息处理流向为主线，参照视网膜、外侧膝状体、初级视皮层、高级视皮层以及更为抽象的大脑皮质等关键视觉信息处理区域的神经机制，提出了基于上述神经机制的若干种图像处理方法，并在实际图像的处理中显示了较好的效果。主要内容包括如下方面：1．针对图像处理领域的颜色恒常问题，本文提出了两种计算模型。一个模型是基于视网膜（包括外侧膝状体）神经机制的颜色恒常计算模型。该模型模拟具有抑制亚区的视网膜神经节细胞非经典感受野特性及其颜色单拮抗机制，实现了图像处理中对色偏图像的颜色恒常。第二个模型是基于初级视皮层神经机制的颜色恒常计算模型。该模型分别从图像导数与非负稀疏编码两个角度模拟初级视皮层神经元感受野，实现色偏图像颜色恒常。通过对多个国际通行颜色恒常算法评估库的测试，上述两个模型均取得了与目前最具优势的颜色恒常技术方法相比拟的结果。在图像处理领域，上述模型表现出了潜在的应用价值；在神经科学领域，上述模型对理解皮层下神经元以及初级视皮层神经元对颜色恒常的作用提供了计算理论上的依据。2．人类视觉神经系统具有等级层次性和双向连接性，具有特征检测和学习能力等特性。在系统层次的视觉信息处理过程的启发下，本文提出一种小波域的多尺度马尔可夫随机场模型模拟视觉系统的上述特性。具体而言，该模型用小波变换实现视觉系统输入图像的稀疏表达，用多尺度马尔可夫随机场表征图像的全局拓扑特征；用金字塔结构所展示的多尺度信息处理能力模拟视觉系统的等级层次性；用自底向上和自顶向下两种信息流模拟视觉系统各层之间的双向连接性；用模型计算中的迭代过程模拟无监督学习机制；用不同的参数设置模拟不同的视觉任务，从而实现真实生物医学图像的区域分割和边缘检测功能。3．针对图像增强问题，我们以模拟大脑节律现象的Wilson-Cowan双节点耦合振子模型为基础，选取使该模型产生极限环振荡条件的参数，采用连续灰度阶图像块作为输入，兴奋性亚群节点响应作为图像增强的输出，做出刺激响应曲线，发现该曲线与图像处理领域中用于图像增强的Gamma校正曲线相似，说明Wilson-Cowan双节点耦合振子模型可以作为一种新的图像增强方法。由于此前对图像增强的视觉机制解释，是以中心外周相互作用的感受野模型为基础的，本工作表明，图像增强还可以神经元群的振荡机制来解释，或者说，本工作为传统图像处理方法中的Gamma增强方法提供了一种神经机制上的解释。将新方法与基于感受野模型的Retinex算法对比，表现出了更好的图像增强性能。更多还原

【Abstract】 Human visual system (HVS) possesses very excellent image processing abilities.Based on the extensive acceptant visual information processing flows, this papermimicks neural mechanisms of retina, lateral geniculate nucleus, visual cortex and moreabstract cerebral cortex，and propose some image processing methods equipped with theabove neural mechanisms to face some practical image processing problems. It mainlyinvolves:1. We propose two models to solve the color constancy problems. One modelnamed “single-opponent cell with non-classical receptive field (nCRF)”(SONRF), it isproposed as a potential mechanism underlying image color constancy at the level ofretinal ganglion (RG) cells and lateral geniculate nucleus (LGN) neurons. This modelsimulates the properties of inhibitory interactions among the subunits of the nCRF (i.e.,disinhibitory effects) and the inhibitory modulation of the subunits to the center with thecolor opponent mechanism of red-green, green-red and blue-yellow. Another model, bymimicking receptive fields of the primary visual cortical neurons from the views ofimage derivative and non-negative sparse coding, is a color constancy model combiningwith the image derivative framework and non-negative sparse coding. We employcommonly used color image databases to quantitatively evaluate these two models,which gave the comparable results as the state-of-the-art non-biologically inspired colorconstancy algorithms. In the view of image processing, these results demonstrate theutility and potential applications of algorithms inspired by biological mechanisms incomputer vision and other realms of image processing. In the view of neuroscience,those models provide supports for the notion of subcortical neurons and primary visualcortex’s roles on the capacity of color constancy.2. As HVS has the properties of feature detection abilities, hierarchy, bidirectionalconnection, and self-learning mechanisms, etc, we propose a multiscale Markov randomfield model in the wavelet domain by simulating some functions of HVS for imagesegmentation. Concretely, for an input scene, using wavelet transforms, our model provides its sparse representations to mimic feature detection abilities, and using thepyramid framework, our model mimics hierarchy. In the framework of our model, thereare two information flows imitating bidirectional connection. For example, a bottom-upprocedure is adopted to extract input features and a top-down procedure is used toprovide feedback controls. Moreover, iterations are the simulation of self-learningmechanisms. In addition, setted by different parameters, our model is able to excuatedifferent biomedical image segmentation tasks, such as edge detection and regionsegmentation with pixels classification.3. A model mimicking cortex rhythms is adopted to achieve image enhancement.This model is based on the coupled Wilson-Cowan oscillators with double nodes. Theinputs are images to be enhanced and the outputs are node responses of the excitedsubpopulation. As image experiments show, the method is able to be used in imageenhancement. Meanwhile, we found that if image patches with continuous gray valuesare employed as stimulus, the response curves are similar with the classical Gammacorrection curves that are used in image enhancement. This fact on one side providesevidence of the image enhancement ability of the proposed method, on the other side, itprovides the neural mechanism, oscillation of the neural population, of the classicalGamma correction method. Numerically compared with the receptive field model basedclassical center-surround Retinex algorithm, the new method shows better results.更多还原