| 题名 | 多光谱成像系统图像处理关键技术研究 |
| 作者 | 张艳超 |
| 学位类别 | 博士 |
| 答辩日期 | 2015-05 |
| 授予单位 | 中国科学院大学 |
| 导师 | 孙强 |
| 关键词 | 多光谱成像 自动调焦 电子稳像 图像分割 投影微分 投影特征峰匹配 偏微分方程 C-V模型 |
| 其他题名 | Research on the Key Technologies of Image Processingbased on Multi-spectral Imaging System |
| 学位专业 | 光学工程 |
| 中文摘要 | 多光谱成像系统主要用于对同一观测目标的多个波段同时进行成像比对，以实现对其成分组成、分布状态等特性进行观测分析。相对于传统的单一波段成像系统，多光谱成像系统能够获得更为全面的目标光谱信息。因此，在地质勘探、目标侦查、生命状态观察、环境监测等众多领域等到了广泛的应用。对于多光谱成像系统，保证在每个波段都能够清晰成像是后续信息提取、特性分析的重要前提。而在实际使用过程中，往往由于噪声干扰、系统离焦、平台抖动等因素，导致系统像质下降，必须采取一定的技术手段对上述问题进行相应处理。由于图像处理手段相对于硬件调整方法，具有成本低、方法灵活、实时性强、改进空间大等优点，在成像系统、尤其是多光谱成像系统中，图像处理技术发挥了越来越重要的作用。 本文在课题组自主研发的多光谱成像系统硬件平台上，重点针对多光谱成像系统的快速自动调焦技术、电子稳像技术、以及目标识别技术等几个关键技术进行了深入研究，提出了相应的图像处理算法。此外，根据研究内容与系统硬件特点，进行了相应软件控制系统的设计与实现。本文的主要研究内容和研究成果归纳如下： 第一，自动对焦方面，对当前经典的自动对焦算法进行了比较分析。针对现有算法的不足，先后提出了基于对数功率谱的离焦深度自动对焦法、基于SUSAN边缘算子的对焦深度自动对焦法以及基于投影微分法的对焦深度自动对焦方法。经过比较分析，最终选定投影微分法作为系统的自动对焦方法。该方法首先是将对焦窗口内的数据做x与y方向投影，对这两个方向投影数组的微分1范数均值求均方根，并将其作为该帧图像的清晰度评价值；然后，结合经典的爬山搜索算法，完成系统的自动对焦过程。然后，利用该多光谱成像系统平台，对该方法的性能进行了相应的验证与评价。实验结果显示，与经典的Brenner、能量梯度以及Roberts清晰度评价算法相比，投影微分算法在保证具有同样优良的单峰性、无偏性、较高的灵敏度等基本特性的同时，算法时间仅分别为以上三种算法的0.67、0.33和0.33倍，能够更好的满足系统对自动对焦的高精度与实时性要求。 第二，电子稳像方面，提出了一种基于投影最大特征峰匹配的稳像算法。该方法首先将参考帧与当前帧等分为若干区域子块，利用灰度投影计算式分别计算每个子块的水平投影和垂直投影；然后，依次将两帧图像相应子块中对应的垂直和水平投影最大特征峰的位置差值，作为相应子块的水平和垂直运动矢量；再次，根据帧间运动矢量变化程度进行运动矢量修正；最后，根据各子块的运动矢量统计结果进行全局运动估计。将该算法应用于多光谱成像系统中，在保证相同的稳像效果的前提下，投影特征峰匹配算法与经典的灰度投影算法相比节省了几乎全部的搜索时间，取得了较好的稳像效果和更好的实时性。 第三，目标识别方面，本文在现有经典的C-V模型目标分割方法的基础上，对该模型进行了改进。该模型通过将每次迭代得到的距离函数的最大值引入C-V模型的Dirac函数，对该函数进行自适应参数修正，以拓宽活动轮廓线的有效作用范围，进而大大的降低了分割算法的迭代次数。实验结果表明，改进的C-V模型在其终止条件下得到了较理想的分割效果，与经典的C-V模型相比，降低了初始曲线位置对最终分割结果的影响，且新模型的收敛速度在原有的基础上提高了至少7倍。改进的C-V模型在实时性及全局性方面都得到了明显改进，进一步提高了该算法在多光谱成像仪的图像分割方面的自适应性。 第四，根据研究内容与系统特点，研制了多光谱成像系统的控制软件。该软件系统具有如下功能：多个波段图像数据的实时采集与显示、手/自动调焦功能、手/自动曝光功能、电子稳像功能、目标分割与识别功能、以及图像与视频序列的存储功能。保证了系统按照预设的功能顺利运行。 |
| 英文摘要 | Multi-spectral imaging system is mainly used for contrast imaging for several spectral bands of the same object, synchronously, in order to monitor and analysis the composition and distribution. Compared with the traditional single band imager, multi-spectral imager can obtain more spectral information; therefore, Multi-spectral imaging system has been widely used in many fields, such as geology exploring, object investigating, life state observing, and environmental monitoring. As a multi-spectral imaging system with a good performance, imaging clearly is at the forefront of the following information extracting and characteristic analyzing. However, noise interfering, optical system defocusing, platform vibrating and other related factors lead to the poor imaging quality, so it’s very necessary to solve above problem accordingly. Relative to the method of hardware adjustment, image processing play a more and more important role in the imaging system, especially in the multi-spectral imaging, because of the low cost, flexible method, real-time, improvement of large space, and so on. Several key technologies about auto-focusing, digital image stabilization and object recognition are researched systematically and some new algorithm of image processing are proposed accordingly in this paper, based on the multi-spectral platform which is researched and developed independently by our team. In addition, the software for the system is designed and realized, according to the research content and the device hardware characteristics. The main research contents and achievements in this paper are summarized as follows: The first, in the aspect of auto-focusing, several classic algorithms have been analyzed comparatively. Aimed at the existed defects of above algorithms, some new methods were proposed such as depth from defocusing based on logarithmic power spectrum, depth from focusing based on SUSAN edge detector, and depth from focusing based on differential projection. Through analysis and comparison, the third one is selected finally. In this method, firstly, calculate the projective values in x-direction and y-direction of the involved image’s focusing window data, derive the 1th norm of the two arrays’ first order differential values, compute the mean of the two set of 1st norm data, and make the Rms of the two means as the definition evaluation value of this image. And then, combined with the classic Mountain Climb-searching method, the auto-focusing process is finished. Finally, the performance of the method is tested and evaluated with the use of the system platform. Experimental results indicate that the auto-focusing method based on differential projection can be realized accurately, and has the same effect with the classic Brenner, Energy gradient and Roberts gradient algorithm, approximately. However, the running time of differential projection is only 0.67 times of Brenner, 0.33 times of Energy gradient, and 0.33 times of Roberts gradient. The method can meet the high-precision auto-focusing requirements very well. The second, in the aspect of digital image stabilization, a new algorithm based on the characteristic peak of projection matching is proposed. Firstly, several sub image blocks are divided about the reference frame and the current frame, and the horizontal projection and vertical projection of each sub image block are calculated according to the projection calculation formula. Secondly, the horizontal and vertical position differences of the corresponding sub blocks’ maximal projection characteristic peaks are calculated as motion vectors. And then, the motion vectors are corrected according to the motion vector variance ratio between frame and frame. Finally, the global motion vectors are derived based on the sub motion vectors. The characteristic peak of projection matching algorithm being applied to the multi-spectral imager, almost all of the search time is saved, in the premise to the same purpose. The ideal image sequence is obtained. In addition, the better real-time characteristic is obtained, too. The third, for the sake of the object recognition requirement, an improved C-V model is proposed, compared to the classical C-V model for image segmentation. In this model, the Dirac function’ parameter is corrected adaptively, by introducing the maximum value of distance function in each iteration. In this way, the effective range of active contour is broadened, and the number of iterations is reduced. The experimental results show that the ideal segmentation effect is obtained by the improved C-V model, with the iteration termination condition. Compared with the classic C-V model, the influence on segmentation, because of the initial contour position differences, is cut down. In addition, the convergence speed is improved 7 times. The characteristics of real time and global nature both become better. Therefore, the robustness of multi-spectral imager segmentation is improved, accordingly. The fourth, according to the research content and the device hardware characteristics, the control software is developed and realized comprehensively. And then, the following functions is realized: the real-time acquisition and display of multi-spectral data, manual focusing and automatic focusing, manual exposure and automatic exposure, digital image stabilization, image segmentation and target recognition, and storage of images and videos. Finally, the system runs normally with the help of the software. |
| 公开日期 | 2015-12-24 |
| 内容类型 | 学位论文 |
| 源URL | [http://ir.ciomp.ac.cn/handle/181722/48956]  |
| 专题 | 长春光学精密机械与物理研究所_中科院长春光机所知识产出 |
| 推荐引用方式 GB/T 7714 | 张艳超. 多光谱成像系统图像处理关键技术研究[D]. 中国科学院大学. 2015. |
| 个性服务 |
| 查看访问统计 |
| 相关权益政策 |
| 暂无数据 |
| 收藏/分享 |
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。
修改评论