导　　师： 严国萍

学科专业： H10

授予学位： 博士

作　　者： ;

机构地区： 华中科技大学

摘　　要： 自动目标识别是当今世界军事技术研究中最具攻关性的课题之一，众多计算机视觉技术均应用于该课题的研究领域中。许多诸如物体边缘检测、空间位置估计和运动跟踪等对于生物视觉来说是轻而易举视觉任务，然而至今在计算机视觉上仍然不能得到很好的解决，由此成为制约自动目标识别技术发展的一个瓶颈。本论文的完成基于2003 航天支撑基金项目“基于生物视觉的自动目标识别算法研究及仿真”的研究工作，提出了一种生物视觉感知系统中的感受野改进模型，并将这种模型应用于自动目标识别中的物体边缘检测和运动估计及跟踪等关键技术的研究中。主要内容包括：
　　 ⑴在广泛查阅国内外文献的基础上，讨论了几种具有代表性的经典和非经典感受野模型以及对它们的改进；实际上，对感受野模型的每一次改进均对应着某种视觉生理机制的融入。发现对于一种有趣的视觉生理机制：当人眼长时间注视某一斜方向变化的光强时，会对该方向的光强变化变得敏感；传统的生物视觉各向同性感受野模型及其改进型均不能对它进行解释。于是，提出了一种具有任意方向的各向异性弹性感受野模型。
　　 ⑵利用所提出的具有任意方向的各向异性弹性感受野模型成功地解释了生物视觉同时对比现象，建立了生物视觉同时对比机制的数学模型。并对将生物视觉同时对比机制的数学模型应用到边缘检测的高通滤波技术中的可行性进行了研究，从而提出了“一种新的基于区域角度搜索的各向异性高通滤波算法”和“各向异性高通滤波中的一种改进型边缘方向估计算法”。
　　 ⑶利用所提出的具有任意方向的各向异性弹性感受野模型成功地解释了生物视觉的注视现象，基于生物视觉突触短时可塑性导致视觉感受野在接受外界持续刺激时将发生感受野形变这一生理现象，提出了对感受野形变的假设，并建立了能描述生物视觉注视机制的数学模型。而“基于人眼视觉注视机制下突触短时可塑性的图像边缘检测算法”的提出也刚好验证了假设的合理性。进一步的研究使得这种基于视觉注视机制的自适应线型滤波算法得以建立和完善。
　　 ⑷将所提出得具有任意方向的各向异性弹性感受野模型应用于运动估计中。提出了物体运动产生“虚拟边缘”这一新的概念，并对“虚拟边缘”给出了严格的数学定义，采用该感受野模型作为能检测“虚拟边缘”的探测函数，将该函数引入小目标运动事件中，分别建立了理想的帧间差探测函数模型和实际的帧间差探测函数模型。在对这两种模型的分析比较中，明确了参数选择策略，从而提出了一套基于帧间差探测函数模型的运动估计算法。
　　 ⑸设计了几种新的模块用于解决目标跟踪中的方向纠错、可实现性和实时性问题，并将这些模块与前文所提出的运动估计算法相结合，从而构成了一套基于帧间差探测函数模型的运动目标跟踪算法。实验表明，该目标跟踪算法能够有效地对在简单和复杂背景下运动的小目标进行跟踪。随后对这种目标跟踪算法的复杂度进行了分析和比较。本文的最后对主要研究成果进行了概括，并指出以后需要进一步研究的问题和探索方向。 ATR /(Automatic Target Recognition/) is one of the tough challenges in the research of the world military technology nowadays, in which the computer vision technology has been applied. However, many visual missions such as edges detection, spatial location estimation and target tracking, which seem to be very simple in biological vision, are unable to be solved well now in computer vision. It leads to the bottleneck for the improvement on ATR. This doctoral dissertation is based on the project named Research and Simulation on the Algorithm of Automatic Target Recognition Based on the Model of Biological Vision. And the project is supported by the Foundation of Astronomic Technology in 2003. In this paper the improvement on the model of the receptive field in the system of biological vision is presented. With the improved model being used, several key technologies such as edges detection and target tracking in ATR are focused in this paper. The main parts are as follow: /(1/) Based on the widely learning of references, several representative types of classical and non-classical model of receptive field even the improvement on the models are discussed. Each improvement corresponds to a mechanism in biological visual. And the mechanism that when we keep our eyes on the optical intensity changes in given direction for a long time, the eyes are sensitive to the optical intensity change in the direction is interesting. But we can find that neither the classical receptive field which has the construction of concentric circles nor its improvement can explain the mechanism. Then an anisotropic stretch receptive field model which has arbitrary direction is put forward. /(2/) The new model of receptive field is applied to explain the phenomenon on the simultaneous contrast in biological vision successfully. Then the mathematical model that can describe the mechanism of the simultaneous contrast is constructed. After the analysis of the feasibility that applying the mathematical model to high-pass filtering on theory and in practice, two algorithms named A New Anisotropic High-pass Filtering Algorithm Based on Region Angle Searching and An Improved Edge’s Orientation Estimation Algorithm are proposed respectively. /(3/) The new model of receptive field is applied to explain the phenomenon on the gazing in biological visual system successfully. According to the mechanism that the Short-Term Synaptic Plasticity caused by repeating stimulation can lead to distortion in receptive fields of neurons in human vision system, a hypothesis on the manner of distortion in receptive fields is proposed in this paper. Based on the hypothesis, the mathematical model that can describe the mechanism of the gazing is constructed. A new algorithm named An Edge Detection Algorithm Based on Short-term Synaptic Plasticity in Gazing Mechanism in Human Visual System is advanced in order to verify the rationality of the hypothesis. Finally an adaptive linear filtering algorithm is established. /(4/) The new model of receptive field is also applied to motion estimation in this paper. The new concept called virtual edge and its mathematical definition are presented. The new receptive field model can be used as a function to detect the virtual edge with target’s moving. With the small target being inducted to the events of motion, the ideal model of temporal difference detection function and the actual one are brought forth. With the analysis of the characteristics of the two models, a rational strategy about parameters chosen is determined. Then a motion estimation algorithm based on the temporal difference detection function is proposed. /(5/) Several new modules are designed to solve the problems such as error correction on directions, the feasibility and the real-time performance in target tracking. Combined the proposed motion estimation algorithm with the several modules above, a new algorithm based on the temporal difference detection function is put forward. The experiments show that the small target which is moving at random on simple and complex backgrounds can be tracked effectively. The analysis and the comparison of the algorithm’s complexity are also completed at last. Finally, the main contributions in this dissertation are summarized and some suggestions and directions for the future work in this field are given.

关 键 词： 目标跟踪 图像识别 生物视觉 图像处理

分 类 号： [TP391.41 TP317.4]

领　　域： [自动化与计算机技术] [自动化与计算机技术] [自动化与计算机技术] [自动化与计算机技术]