基于形态学理论的目标检测技术

【作者】 庾农；

【导师】 王润生；

【作者基本信息】 中国人民解放军国防科学技术大学 ， 信息与通信工程， 2000， 博士

【摘要】 图像目标的自动检测在现代高技术战争中具有十分重要的意义，研究与发展可靠性高、适应性广的检测算法与处理系统的需求越来越迫切，从而也推动了计算机视觉与自动目标识别理论的发展。由于视觉过程本身的高度非线性，采用非线性的信号处理技术就自然成为图像领域的一种重要发展趋势。数学形态学是一种颇具特色的非线性理论，本论文具体研究了利用这一理论解决目标检测所面临的关键技术问题。 利用形态学理论实现目标检测的核心内容，是基于形态学理论建立目标检测模型，并形成相应的自动处理过程。该模型由形态学统一表示定理构成的多结构基元滤波器和结构参数优化学习两个部分组成。在多结构基元滤波器设计中，通过学习人-机交互选定的目标样本，自动确定形态变换的组合规则及其结构元素，最终以神经网络形式构成滤波器。在结构参数的优化学习中，利用应用领域的先验知识，分别设计了自适应BP学习、启发式遗传学习和引导式模拟退火学习等三种最优化计算方法。 将本文提出的研究方法用于实景图像分析，获得了满意的实验结果，既可自动检测出红外运动图像目标，也能提取多种光学静态图像目标。实验结果表明，本文设计的检测算法对复杂变化的图像环境具有良好的滤波性能和稳健的适应能力。更多还原

【Abstract】 Automatic target detection in images is of vital importance to modern high-tech warfare. Demand of researching and developing detection algorithms and processing systems with high reliability and high adaptability have been increasing more and more these years, and computer vision and automatic target recognition have been advanced as well. Because vision procedure is actually nonlinear, using nonlinear techniques become one kind of important research tend in image fields. Mathematic morphology is one of nonlinear theories, and has distinguishing features. In this paper key issues of applying the theory to target detection are studied.The core content in realizing target detection with morphological theory is to construct a target detection model and to form a corresponding automatic process. The model consists of two parts: the multi-structuring elements filter formed according to the generalized morphological denotation theorem, and an learning procedure to obtain optimal structural parameters. In designing a multi-structuring elements filter, combination rules and structuring elements of the morphological transform are determined automatically, and one kind of neural networks is taken for the filter, In optimzing structural parameters of the filter, three computation methods are designed respectively, by adopting some priori information in application fields to guide optimal structural parameter learning procedure, which are the BP adaptive learning algorithm, the heuristic genetic learning algorithm and the inductive simulated annealing learning algorithm.The approch developed in this paper is applied to some real image data, and satisfied results are obtained. Both the moving targets in a set of infrared images and the static targets in optical images can be detected automatically. Experimental results indicate that object detection models and relative algorithms have better and robust performance.更多还原