【作者】 唐峥远；

【导师】 姚莉秀；

【作者基本信息】 上海交通大学 ， 模式识别与智能系统， 2012， 硕士

【摘要】 随着科学技术的发展,红外设备和红外技术得到了越来越广泛的应用,红外图像的处理与识别随之也扮演越来越重要的角色。在红外图像中目标距离较远且噪声干扰较大,对红外小目标的检测和跟踪技术提出了巨大的挑战。近年来,伴随着压缩传感的兴起,稀疏表示理论作为压缩传感的基础也受到了广泛的关注。它不再受限于传统的傅里叶变换、小波变换等对于基函数的构造,而是采用冗余的超完备字典来实现对信号的表示。在超完备字典中,基原子的个数总是大于它自身的维数,因此寻找采用最少的基原子来对信号进行最优表示的方法,即稀疏表示应运而生。虽然稀疏表示的理论仍然在进一步的探索和完善中,但是其在信号处理领域的应用已经收获了许多优秀的成果,展现出了未来发展的巨大潜能。本文正是基于上述背景,采用基于图像稀疏表示的技术对红外小目标检测和跟踪进行了深入的研究。在前人对于红外小目标研究的基础上,本文进一步提出了新的检测和跟踪算法:1.采用图像稀疏表示理论来获得待测子图像块在超完备字典下的稀疏表示系数,其中超完备字典由修正高斯灰度模型建立。由于目标子图像块和背景子图像块在稀疏域中表现出显著的差异性,因此通过评估整幅图像的稀疏集中程度可以获得图像中目标分布的描述。最终通过简单的阈值操作就可以分离出所需的目标集,提高了红外小目标检测性能。2.更进一步地,对于稀疏表示理论中字典构造进行了研究,利用字典学习的方法成功获得了最优表示的红外小目标字典。基于该最优字典可以获得目标子图像在稀疏域中的最优表示,从而能够十分精确地重建图像。因此,对于重建后的残差图像进行分析就可以高效地抑制背景中的杂波和噪声,提高图像的信噪比。3.基于上述稀疏表示在红外目标检测中表现出的对噪声抗干扰能力较强和对遮挡的不敏感性,在粒子滤波的框架下提出了对红外目标进行跟踪的方法。该方法通过构建目标子空间和表示噪声与遮挡的平凡子空间,利用稀疏表示作为目标观测模型,提高了对目标描述的鲁棒性;同时,采用在线学习的方法对目标子空间进行更新,使其进一步适应背景中光照等因素引起的变化,提升了跟踪算法的有效性和稳健性。更多还原

【Abstract】 With the development of modern science, infrared related techniques have been extensively used and infrared image processing and recognition also play a more and more important role. Because of far distance when equipment capturing object and large noise in the environment, small target detection and tracking is always a challenge area.Recently, compressed sensing which is based on sparse representation theory attracts huge attentions. Instead of being constrained to traditional basis construction such as Fourier transform and wavelet transform, sparse representation uses overcomplete dictionary to represent signal. A normal overcomplete dictionary always obtain more atoms than the dimension of itself, and therefore sparse representation search for the least number of atoms in a single representation. Although the theory is still not fully explored and need for further improvement, lots of its applications reach excellent result and imply a promising future.Based on such background, this paper approaches small target detection and tracking via sparse representation. The proposed algorithms are as follows:1. Apply sparse representation theory to obtain sparse coefficients of testing image based on a small target dictionary built by Modified Gaussian Intensity Model. Then, taking advantage of the difference between representations of target and background sub-images, sparse concentration index can be used to describe target distribution in the image. Finally, a simple threshold can easily reveal target class and improve detection performance. 2. Furthermore, train an optimal small target dictionary ground on research of dictionary learning. Upon that, the best representation of the dictionary has ability to rebuild image in a high accuracy and similarity. As a result, residual between original and rebuilt image could achive higher signal-to-noise ratio by background and clutter suppression.3. Benifit from outstanding antinoise ability and occlusion insensitivity of sparse representation, a new infrared target tracking technique is proposed under particle filter framework. The algorithm employs sparse representation as observation model which is implemented by building target and noise subspace, and successfully enhances object decription outcome. Meanwhile, online learning is used to dynamically update the target subspace in order to adapt to background variation and improve the efficiency and robust of the algorithm.更多还原