【作者】 吴田富；

【导师】 高隽；

【作者基本信息】 合肥工业大学 ， 信号与信息处理， 2005， 硕士

【摘要】 目前,主动视觉是机器视觉研究的热点和发展方向。主动视觉的核心内容是:为了完成给定的视觉任务,如何主动、智能、有选择地来获取视觉信息;从计算的观点来说,就是要建立视觉选择性注意机制的计算模型,对图像数据进行显著性度量。 本文围绕建立视觉注意计算模型展开,主要研究内容包括: ● 将尺度、显著性和物体识别放在一个框架下研究。针对在传统机器视觉研究中,尺度、显著性和物体识别多数被分开研究的现状,本文首先分析了三者之间的内在联系和相互关系,得出应该在一个框架中来研究它们的结论;讨论了视觉中的尺度问题和尺度空间表示方法、显著性度量的计算本质以及基于局部不变性特征的识别方法。 ● 结合尺度与特征引导的视觉注意自下而上的计算模型研究。本文针对隐式注意,建立一个自下而上的结合尺度与特征引导的计算模型。根据神经科学的研究,本文选取强度、颜色和方向三种特征以及尺度引导注意。首先基于视觉感受野和整合野机制建立一种具有竞争和协作双重特性的滤波器对三种特征图进行迭代,形成特征空间显著性度量;然后对图像中主尺度进行估计,建立特征图的尺度空间表示,结合尺度与特征度量显著性,并对注视点进行最佳尺度选择。 ● 结合视觉注意计算模型和基于局部不变性特征识别的研究。针对复杂背景下的物体识别,讨论了视觉注意和基于局部不变性特征识别的结合,从而使得视觉注意过程中进行的各种早期特征提取和尺度空间表示不像在传统算法中那样被“废弃”,同时通过视觉注意可以改进基于局部不变性特征识别使其获得背景不变性。更多还原

【Abstract】 At the present, active vision is the hot field and the developmental direction of machine vision, in which the key problem is how to acquire visual information actively, intelligently and selectively under a given visual task. From the viewpoint of computation, it is to implement a computational model of visual selective attention mechanism to compute the saliency of image data.This thesis focused on the computational model of visual selective attention mechanism, including the following research contents:Firstly, a unified computational model of scale, saliency and object recognition is studied. In traditional studies of machine vision, the study on scale, saliency and object recognition are performed respectively. Contrary to the traditional idea, they are unified in this thesis in which the problem of scale in vision and the scale-space representation, the essence of the computation of saliency and the method of recognition based on local invariant feature are discussed.Secondly, a computational model of visual attention deployed by scale and features is studied. The model presented in the thesis aimed at the bottom-up aspect of covert attention which deployed by scale and features. Based on the research of neuroscience of visual attention, Intensity, color and orientation are used as the features attracted attention in this thesis. At first, a filter which has dual characteristics of competition and cooperation is created based on the mechanism of the visual receptive field and the integrated field, and the three primary feature maps are iterated by the filter to compute the feature-space saliency. Then the primary scale of the input image is estimated followed by the scale-space representation of the input image being generated. So far, the saliency of scale and features can be computed, and at the same time the optimal scale of the fixation is selected.Finally, a model which unified the visual attention and visual recognition is studied. Focusing on the problem of recognition in complex background, the thesis discussed how to combine the visual attention model presented in the thesis and the model of recognition based on local invariant features. The implementation of the unified model made it possible to reuse the features extracted in the stage of visualattention in the stage of recognition, and could add the background invariance into the traditional local invariant feature-based method.更多还原