【作者】 唐奇伶；

【导师】 桑农；

【作者基本信息】 华中科技大学 ， 模式识别与智能系统， 2007， 博士

【摘要】 轮廓与边界定义了目标的外表形状,确定了区域之间的分界线,它们是人类与计算机进行目标识别的重要特征。然而从纷乱的自然场景中提取目标的轮廓与边界是一件非常困难的任务。为了做到这点,三个主要的问题需要解决:1.排除大量由背景纹理所产生的局部边缘成分;2.根据场景中上下文的信息将局部成分组织成有意义的全局特征;3.有些重要的结构缺乏明确的物理定义(例如纹理的边界),并且有的部分目标可能与背景具有相同的强度,使得边界的响应非常弱甚至缺乏局部的有效证据。针对这些困难,本文根据初级视皮层感知机制建立了各种轮廓与边界的检测模型,并通过合成图像与自然图像检验了算法的性能。为了减少背景中的纹理边缘成分并突出区域的边界,本文利用非经典感受野循环抑制的动态属性提出了一种纹理抑制方法。这个方法对纹理与边界采取不同的处理方式,从而很大程度地减少了背景中无意义的干扰成分,并有选择性地保存了孤立的轮廓和区域的边界。如何从复杂的场景中将具一致空间结构的成分组织成显著的轮廓是研究的另一个主要方面。本文结合共圆规则与视觉对低曲率路径的偏好定义了一个轮廓结合局部聚集函数,这个函数将同排列与同方位两种属性巧妙的联系起来。通过上下文的相互作用将局部成分整合成一个有意义的全局特征并从背景中突出。通过空间分离的兴奋与抑制作用区域,本文将环境抑制与空间增强统一在一个结合模型中,从而允许两种对立的感知行为同时存在。基于这个模型,本文强调了空间增强与环境抑制在轮廓与边界检测中起的不同作用,抑制主要表现在表面与纹理的分割,而增强主要是用于轮廓的结合和图形背景的分离。彩色图像比灰度图像携带更多的信息,能有助于图像的分析并产生更好的结果。因此将灰度模型进一步扩展到彩色图像的处理。彩色模型将涉及更多属性的同质性抑制,能更有效的去除纹理边缘;另一方面,颜色为轮廓聚集提供了更多的信息,更有利于相同属性的整合。最后通过两个应用项目——血管造影图像的增强与合成孔径雷达图像的道路检测,表明了该模型的广泛用途。更多还原

【Abstract】 Contours and boundaries that define object shape and indicate outer limits for regions. They are critical for human or computer recognition of objects. However, it is extremely difficult to extract contours from cluttered scenes automatically. To do that, three major problems need to be solved: (1) eliminating non-meaningful edges engendering from texture fields rather than object boundaries; (2) grouping local elements into meaningful global features according to context information; (3) many important structures are only implicitly determined, such as by texture boundaries, or are entirely physically absent, such as where a background is by chance the same color as a foreground object. To address these problems, we construct different models according to perceptual mechanisms of primary visual cortex and verify the performances of the models by synthetic and natural images.To order to reduce cluttered and textured elements, we present a method for suppressing texture edges via dynamic properties of recurrent inhibition in non-classical receptive field. The method deals with texture and boundary in different ways, and thus dramatically reduces non-meaningful distractor elements, while selectively retains region boundaries and isolated structures.How to organize coherent spatial configurations into salient contours is another important issue of our study. We combine a co-circularity rule with visual preference for low curvature to define a local grouping function of contour integration, which relates the axial specificity with the modular specificity. Local elements are grouped into a meaning global feature according to contextual information, thus allowing them emerge from their backgrounds.We unify the dual processes of spatial facilitation and surround inhibition in an integrated model by spatially segregated regions of excitatory and inhibitory inputs, thus allowing the model to implement multiple perceptual tasks that require opposing interactions. Here, we put some emphases on the two different roles– spatial facilitation and surround inhibition - played in the contour extraction. Inhibitory interactions are supposed to play a more important role in the segmentation of surfaces and textures, while excitatory contextual interactions are deemed to be more significant in contour integration and figure-ground segregation.Color of an image can carry much more information than gray level, which can help the image analysis process and yield better results than approaches using only gray scale information. Thus we extend the gray-level version to color image processing. The color version involves homogenous inhibition of more properties, which can more effectively remove textured edges; on the other hand, color information provides more cues for contour grouping, which can help to organize the same property.Finally, we apply the scheme to two aspects of vessel enhancement in DSA images and road detection in SAR images.更多还原