【作者】 李玲玲；

【导师】 丁明跃；

【作者基本信息】 华中科技大学 ， 模式识别与智能系统， 2005， 博士

【摘要】 随着图像传感器技术的发展,多传感器图像融合已成为图像理解、计算机视觉以及遥感领域中的一个研究热点,并广泛应用于自动目标识别、智能机器人、遥感、医学图像处理和制造业等领域。像素级多传感器图像融合获取的原始信息量最多、检测性能最好、应用范围最广,是各级图像融合的基础。本论文结合有关国家自然科学基金、航天技术创新基金等课题要求,针对像素级多传感器图像融合方法和应用进行了深入研究。主要工作可总结为以下几个方面: 多传感器图像配准是进行像素级多传感器图像融合的前提,其误差大小直接影响融合结果的有效性。对图像配准方法进行了综述,分析了各种方法的适用性、优点和不足之处。在此基础上,针对目前研究中存在的问题,提出了一种基于Harris 角点特征的图像自动配准方法。该方法首先提取参考图像和待配准图像的Harris 角点特征点集,然后通过角点邻域相关匹配和马氏距离仿射变换不变性实现角点的匹配,从而完成图像的配准。实验结果表明该算法在保持高配准精度的同时有较高的执行速度,能够实现有大旋转角度、平移以及灰度差条件下不同传感器或波段的图像自动配准。传统的基于小波变换的图像融合方法存在移变问题。针对此问题,提出了一种具有平移不变性基于离散小波框架的多传感器图像融合方法。提出了低频基于改进的邻域熵、高频基于跨尺度的邻域空间频率的融合策略,有效地抽取了变换域各尺度、方向上的显著特征,并将它们融合在一起。大量的实验验证了该方法的有效性。为了进一步满足图像融合对连续方向的要求,在研究了可操纵方向金字塔变换的原理和性质的基础上,提出了另一种具有平移不变性基于可操纵方向金字塔变换的多传感器图像融合方法。该方法利用可操纵方向金字塔变换良好的方向控制能力和平移不变性,有效地捕获了源图像的方向信息,提高了融合图像的直觉可视性,获得较基于离散小波框架方法质量更高的融合图像。提出了一种基于双树复小波变换的多传感器图像融合方法。该方法在保持近似的平移不变性和良好的方向分析能力的同时,只引入有限的数据冗余,在获得较高质量融合图像的同时,进一步降低了算法复杂度,减少了计算量和对更多还原

【Abstract】 With the development of image sensor technology, multi-sensor image fusion has attracted many attentions in image analysis, computer vision and remote sensing, which is widely applied in a variety of fields such as automatic target recognition, intelligent robots, remote sensing, medical image analysis and manufacturing. Pixel level multi-sensor image fusion can obtain more original information, and has a better detection performance and application. Therefore, it is the basis of the image fusion on other levels. The researches in this thesis were partly supported by three national research grants and focused on the techniques and applications of pixel level multi-sensor image fusion. The main contents of the thesis include: Multi-sensor image registration is the prerequisite of pixel level multi-sensor image fusion. The error of image registration significantly affects the result of image fusion. First, image registration techniques are reviewed, and their advantages and drawbacks are analyzed. To solve these problems, a multi-sensor image registration method based on Harris corner was proposed. The first step is Harris corner detection in the reference and sensed images. At the second step, the corner points are matched based on neighborhood correlation and affine invariance of Mahalanobis distance. The final step, based on the set of correctly matched corner point pairs, the accurate transformation parameters between the reference and sensed images are estimated. Experiments demonstrated that our image registration method can register images with a big rotation angle, shift, or gray level difference, and have both high registration accuracy and speed. Discrete wavelet transform has a shift variance problem. To solve this problem, a shift invariant fusion approach based on discrete wavelet frames was developed. The method defines two fusion measurements, improved-neighborhood-entropy for lowpass subband and across-band-neighborhood-space-frequency for highpass subbands. It effectively extracts salient features at different scales and directions and fuse them. Experimental results demonstrated that satisfactory results were obtained using the proposed method. In order to satisfy the requirement of more directions in image fusion, based on the study of steerable pyramid’s principle and characters, another shift invariant fusion approach based on steerable pyramid was presented. This method makes use of the multi-orientation and invariant property of steerable pyramid to obtain better results than the method based on discrete wavelet frames. An image fusion approach based on dual tree complex wavelet transform was proposed in Chapter 5. This method provides approximate shift invariance and good directional selectivity with limited redundancy. It obtained a good fusion performance and dramatically decreased the computational complexity. How to assess image fusion performance is important for evaluating image fusion method. First, the evaluation measurements proposed are reviewed, and the qualitative and quantitative comparison rules are set up. Furthermore, use the proposed methods to evaluate the methods developed in this thesis. Finally, the effect of the size of neighborhood window, decomposition level and fusion rules on the performance were investigated, and some conclusions were concluded.更多还原