【作者】 魏晗；

【导师】 张长江；

【作者基本信息】 浙江师范大学 ， 计算机软件与理论， 2007， 硕士

【摘要】 红外技术在21世纪的军事和民用等各个领域发挥着至关重要的作用,而且随着未来战场的需要和国民经济的不断发展,红外技术将发挥着越来越重要的作用。其中红外目标识别技术是世界各国学者研究的前沿和热点问题之一。本文根据红外目标自动识别系统(ATR)的特点,在红外车辆目标的对比度增强、自动分割、特征提取、目标识别等方面,进行了较为系统的研究。本文的主要研究内容如下:1)基于遗传算法的红外图像自适应模糊增强。由于红外车辆目标在成像过程中存在许多不确定性即模糊性,针对红外图像的模糊性将模糊理论运用到红外图像,提出了一种基于模糊理论的图像质量测度函数,把它作为遗传算法的适应度函数对非完全Beta函数的α和β参数进行自适应动态调节来拟合几种典型的灰度变换曲线,实现感兴趣区域红外车辆目标的自适应模糊增强。实验结果表明了该方法的合理性和有效性,在性能上优于传统的图像增强技术和现有的一些同类增强技术,具有较高的自适应性和智能性。2)基于遗传算法的红外车辆目标自动模糊分割。针对红外图像的特点,提出了一种基于遗传算法的自动模糊分割红外车辆目标图像的方法。首先选取图像中包含待分割的车辆目标的感兴趣区域以加快运算速度;然后对感兴趣区域图像进行模糊增强,借助于二维OTSU方法对增强后的感兴趣区域进行阈值分割。为了加快分割算法的速度,先限定一个最佳阈值取值空间,再利用遗传算法在此阈值空间内自动搜索最佳分割阈值;为了弥补单独利用二维OTSU方法分割的不足,采用缩短模糊边缘宽度的方法来提取感兴趣区域红外车辆目标图像的边缘。把二维OTSU方法分割的图像与模糊边缘提取得到的边缘图像进行或运算,最后进行填充以得到最终的车辆目标分割图像。实验结果表明:这种自动模糊分割技术比一维OTSU和二维OTSU算法能得到更准确和完整的车辆目标。3)基于径向基神经网络的红外车辆目标识别。给出一种基于径向基(RBF)神经网络的车辆目标识别算法。该方法提取平移、旋转、尺度放缩等变换下都不变的目标形状特征,其中包括8个离散余弦变换描述子,6个独立的不变矩,3个最能区别目标特征的区域描述子,把这些特征参数输入RBF神经网络进行分类识别。实验结果表明,这种RBF网络的特征融合识别方法性能稳定,比BP网络速度快,且较BP网络具有更高的识别精度。更多还原

【Abstract】 In 21 century, infrared technology is valuable in military and civilian domains. With the demand of battlefield in the future and the development of country economy, the infrared technology will be more and more important. Infrared target recognition is one of the hot problems for scholars in the world. Based on the features of infrared automatic target recognition, this paper has made some systematic studies in infrared vehicle target recognition, such as preprocessing of infrared vehicle target images, contrast enhancement, automatic segmentation, features extraction and target recognition, etc.. The main works of this paper are as follows:1) Adaptive fuzzy infrared image enhancement based on genetic algorithm. Because there are uncertainties, that is to say, fuzziness in the infrared vehicle target image, fuzzy theory is used in the infrared image processing. A new kind of image measure function is presented by fuzzy theory. We use it as the fitness function of genetic algorithm to adaptively optimize parameterαandβin in-complete Beta function. Thus an optimal gray transformation curve is obtained to enhance the region of interest in an infrared vehicle target image. Experimental results show that this method has high adaptability and intelligence. The proposed method is better than classical image enhancement methods and some existing similar methods.2) Automatic fuzzy segmentation for infrared vehicle target based on genetic algorithm. According to the characteristic of infrared images, a new auto(?)atic fuzzy segmentation method is presented based on genetic algorithm to segment vehicle target from an infrared image. Firstly, a region of interest (ROI) is selected in order to reduce computation cost. Secondly, the ROI is enhanced by fuzzy algorithm. Thirdly, 2D Maximum Between-cluster Variance algorithm is applied to segment the ROI. At the same time, the genetic algorithm is combined with 2D MBV to make the calculation faster by its capacity of searching the best answer in a threshold space. Then we detect fuzzy edge based on shortening width of fuzzy edge. The final segmentation image can be obtained by OR and filling operations for the segmented region by combining 2D MBV with fuzzy edge. Exper(?)mental results show that the new method can get higher well and truly vehicle target than 1D OTSU or 2D OTSU.3) Infrared vehicle target recognition based on RBF network. A vehicle recognition approach is proposed based on radial basis function (RBF) neural Network. It extracts the invariant features for translation, rotation, and scale change of vehicle targets: 8 discrete cosine transformation descriptors, 6 independent invariant moments and 3 region characterizations. Compared with BP network, the RBF Network is faster and has higher recognition rate.更多还原