《红外成像目标检测与识别方法研究》由会员分享，可在线阅读，更多相关《红外成像目标检测与识别方法研究（116页珍藏版）》请在金锄头文库上搜索。
1、西安电子科技大学博士学位论文红外成像目标检测与识别方法研究姓名：刘靳申请学位级别：博士专业：模式识别与智能系统指导教师：姬红兵20100701摘 要 I 摘 要 利用红外成像实现自动目标的检测与识别是现代军事武器装备的主要技术发展方向。红外图像的目标检测与识别问题是武器系统中的关键技术之一，同时也是光学和图像领域的研究热点。随着现代高技术条件下战场对抗的日趋激烈，要求武器系统能够在复杂的自然环境和人为干扰条件下对目标进行检测和识别。 本文依托红外成像制导中自动目标检测与识别技术研究的课题背景并结合具体的应用需求，对红外图像中目标的预处理、检测以及识别进行了深入的研究，取得的主要研究成果如下：
2、1、 研究了红外图像的预处理方法， 包括红外图像的噪声平滑、 背景杂波抑制、抖动补偿和双模图像的配准方法。针对可见光与红外图像的融合问题，研究了基于小波多分辨分析的多源图像融合方法和基于 Contourlet 多尺度几何分析的多源图像融合方法，提出了一种基于能量与区域相关性的融合检测算法，分析了图像融合算法的性能评价指标。 2、 针对复杂背景下单帧图像的红外小目标检测的问题， 提出了一种基于当前残差改进 M 估计的红外背景抑制算法。 该算法利用 M 估计的基本模型预测背景，将目标像素和观测噪声视为背景估计的混合干扰，提出与背景图像残差相关的校正函数自适应地调整估计增益，从而减小异常样本对背景估
3、计的影响，提高了估计的准确性。同时，引入遗忘因子使得算法能够适用于非均匀背景的估计，提高了算法的鲁棒性。 3、 针对常规管道滤波中噪声对管道中心坐标位置的干扰导致序列检测算法性能下降的问题，提出了一种基于移动式加权管道滤波的弱小目标序列检测方法。该方法在管道滤波检测目标时根据目标位置实时地修正管道中心坐标位置，有效地抑制了边缘噪声对目标检测的干扰，改善了算法的检测性能。 4、 研究了基于均值漂移的红外图像分割方法。 针对灰度分布不均匀以及背景复杂的情况下，基于均值漂移的分割算法在区域合并后存在的过分割问题，提出了一种基于标准割与均值漂移相结合的图像目标分割方法，实现了红外图像由“粗”到“细”的
4、分割，能够有效地消除过分割现象，保证了目标分割的准确性。 5、 针对快速独立分量分析法中迭代初始值选取敏感和独立分量特征提取无序的问题，提出了一种基于距离函数准则的快速独立分量分析算法。该算法通过一维搜索策略使其收敛性不依赖于初始值的选取。同时，基于类内类间距离函数设计新的特征优选准则，根据评估因子最小化对独立分量特征进行有序排列并保留对目标分类贡献大的独立分量特征，实现了用少量特征对目标的有效描述，提高II 红外成像目标检测与识别方法研究 了分类的准确率。克服了在高维特征子空间下随着训练图像样本数的增多红外目标识别率和稳定性下降的问题。 在红外多目标分类方面， 针对K-近邻分类器在多分类中误
5、分率较高的问题，提出了一种基于 Hadamard 码与K-近邻分类器相结合的红外多目标分类算法。该算法利用 Hadamard 码优越的纠错性能提高了K-近邻分类器的分类准确率。 关键词：红外图像预处理 能量与区域相关性融合 M 估计 加权管道滤波 均值漂移 标准割 独立分量分析 Hadamard 纠错码 K-近邻分类器 Abstract III ABSTRACT The utilization of infrared imaging in automatic target detection and recognition is one of the main technological de
6、velopment directions of modern military weapon equipments and a key technique for military weapon systems, and has been the subject of intense investigation in recent years. The intensified battlefield rivalry under modern high technique conditions requires weapon systems to be capable of detecting
7、and recognizing various targets in complex natural background and artificial interference. Sponsored by the research project of automatic target detection and recognition in infrared imaging guidance, this thesis conducts a thorough study on image preprocessing, target detection and recognition in l
8、ight of specific application requirements. The main research results are as following: 1. On the basis of investigation of infrared image preprocessing (such as noise smooth and cluster suppression), jitter compensation and registration of dual-mode image, multimodality image fusion based on wavelet
9、 multi-resolution analysis and multimodality image fusion based on Contourlet multi-scale geometry analysis against the problem of fusion of visual and infrared images, this paper presents a novel fusion detection algorithm based on energy and region correlation and analyses the performance evaluati
10、on criteria of image fusion algorithm. 2. Directed against the problem of lower detection probability of small IR targets in complex backgrounds, an improved M-estimation filtering algorithm for suppressing background clutters based on residual improvement is proposed. This algorithm introduces a ba
11、sic model of M-estimation to predict background, and treats target pixels and observed noises as the mixed interference of background estimation. It uses the correction function related to residual to adaptively adjust gain to reduce influence of abnormal samples on background estimation so as to in
12、crease the accuracy of estimation. Meanwhile, the proposed algorithm introduces a forget factor to make the algorithm adaptive to non-homogeneous background prediction to improve the robustness of the algorithm. 3. Directed against the problem of lower detection probability of traditional pipeline f
13、ilter algorithm due to marginal noise interference on pipeline center coordinates, a variable weighted pipeline filter algorithm is presented for detecting small targets in IR image sequences. An adaptive learning scheme is employed to IV 红外成像目标检测与识别方法研究 modify pipeline center coordinates in real ti
14、me according to targets positions. This method can effectively restrain marginal noise interference. Experiments show that detection performance is significantly increased by using the proposed algorithm. 4. Segmentation of infrared images based on mean shift is investigated. Directed against over-s
15、egmentation of the mean shift based segmentation algorithm under the condition of inhomogeneous gray distribution of objects and complex background, an infrared image segmentation approach based on mean shift and normalized cut is proposed. This approach segments an image from coarse to fine and can
16、 effectively eliminate over-segmentation so as to ensure accurateness of segmentation. 5. Directed against the problem that the selection of the initial values for Newton iteration in the Fast ICA algorithm is very sensitive and of the disordered extraction of independent elements, the author proposes a rapid independent element analysis algorithm based on distance function criteria. A one dimension search strategy is imposed on the direction of Newton iterative to ensure converge
