信息集成中编码曝光技术优化研究

发布时间：2018-03-06 12:44

本文选题：编码曝光　切入点：图像复原　出处：《国防科学技术大学》2014年博士论文　论文类型：学位论文

【摘要】：遥感成像系统信息集成需要对各种平台、各种波段的成像系统进行有效地整合与统一,要求各分系统能够提供足够精确的信息源。基于无人机平台的可见光遥感成像系统作为信息集成的重要组成部分,其成像运动模糊问题会影响到整个遥感成像系统的决策支持能力。随着无人机技术的快速发展,基于无人机平台的遥感成像系统在遥感成像领域发挥着越来越重要的作用。因此,解决遥感系统信息集成中的无人机可见光遥感图像运动模糊问题具有迫切的现实需求。无人机可见光遥感图像运动模糊主要是由成像设备与拍摄场景之间大尺度相对运动引起的。针对图像的运动模糊,当前主要有两种解决思路:一是采用后处理方式的图像复原算法;二是采用高速成像技术。然而,传统模糊图像复原是一种病态问题,对于大尺度的运动模糊图像的复原效果不够理想。高速成像技术需要特殊硬件设备的支持,价格昂贵且应用范围有限。近年来,计算摄影领域中发展出一种编码曝光成像技术,创造性地将病态的模糊图像复原问题转变为一种良态问题,为运动模糊图像复原提供了新的发展空间。同时,受压缩感知理论的启发,编码曝光的随机采样方式有望以低帧率的采样实现高速成像。本文以对编码曝光技术的优化研究为出发点,从模糊图像复原和高速成像两个方面深入探讨了如何在成像设备与拍摄场景之间存在大尺度相对运动的情况下获取目标清晰图像的问题。论文的主要创新点及取得的研究成果包括:(1)提出了一种基于勒让德序列和遗传算法的编码曝光最优码字搜索方法。编码曝光最优码字的搜索空间会随着码字长度呈指数级增长,当前的码字搜索方法只能用于长度较短码字的搜索。本文将信息论领域中低互相关度的勒让德序列用于编码曝光成像中,通过对勒让德序列进行简单的旋转和延拓操作进一步提高了码字的性能,得到了改进勒让德序列。将改进勒让德序列作为遗传算法的初始种群,利用遗传算法的交叉和变异算子实现了对改进勒让德序列的进一步优化,得到了编码曝光最优码字。该码字搜索方法可以很好地满足编码曝光技术对于码字可逆性的要求,且能大大降低复原图像的噪声。由于该方法充分利用了勒让德序列可以由公式快速计算得到的优势,使得该方法效率很高,适合搜索长度较长的码字。(2)提出了一种考虑CCD传感器噪声的编码曝光最优码字搜索方法。编码曝光模式下CCD传感器噪声比较严重,从实用化的角度出发,在寻找最优码字时需要充分考虑传感器噪声的影响。本文首先基于仿射噪声模型系统分析了编码曝光相机的传感器噪声成分;其次给出了光子噪声条件下最优码字结构和复原图像信噪比增益关系的解析表达式,并对真实的编码曝光相机进行了噪声标定;针对已有的码字搜索方法多采用近似穷极搜索,时间效率低的缺点,最后提出了考虑CCD噪声的寻找最优码字的适应度函数,并采用遗传算法搜索编码曝光相机的最优码字。该算法不但提高了码字搜索效率,并且大幅度地提高了复原图像的信噪比。(3)提出了一种基于图像能量谱统计信息分析的编码曝光图像模糊尺度自动估计方法。编码曝光图像模糊信息是非连续的,使得当前流行的模糊尺度估计方法失去效力。本文通过对编码曝光模糊图像能量谱统计数据进行分析,发现复原图像的能量谱统计数据点发散程度与图像复原过程中输入的模糊尺度的正确程度紧密相关。即只有使用正确的模糊尺度时,经过反卷积运算后得到的复原图像能量谱统计数据点的发散程度最小。基于此,本文对图像能量谱统计数据点进行多项式拟合,用残差平方和的大小来代表复原图像能量谱统计数据点的发散程度。编码曝光图像的复原可以采用快速的直接反卷积算法,因此本文基于迭代寻优的思路来估计编码曝光模糊图像的模糊尺度。给定一个初始的模糊尺度,使用全局搜索算法,则能快速地得到编码曝光模糊图像正确的模糊尺度。(4)提出了一种基于单个编码曝光相机进行高速视频重建的方法。充分考虑到在现代传感器设备上的易实现能力,进一步探索了使用单个编码曝光相机实现高速视频帧重建的可能性。考虑到高速视频信号的空间冗余和时间冗余存在很大差别,本文提出使用克罗内克积构建三维双曲线小波基来同时稀疏表示高速视频信号的空间冗余和时间冗余。基于不同尺度的分解,双曲线小波基可以同时建模高速视频各个维度上不同的信号结构,恰好迎合了高速视频信号在空间维和时间维上具有不同平滑度的本质属性。为了进一步提高重建视频帧的质量,本文充分利用了1?凸优化模型可以进一步融合先验知识的特性,在重建模型中加入高速视频信号帧间相关性的全变分正则化项,以进一步建模视频的时间冗余。最后通过最小化一个凸优化问题实现了从低帧率的编码曝光图像序列中重建出高质量的高速视频帧。
[Abstract]:Remote sensing imaging system is the need for a variety of information integration platform, imaging system of various bands were effectively integrated and unified, requirements of each system can provide enough accurate information sources. An important part of optical remote sensing imaging system based on UAVs as information integration, the image motion blur problem will affect the whole remote sensing imaging system the decision support ability. With the rapid development of UAV technology, remote sensing imaging system of UAV platform is playing an increasingly important role in the field of Remote Sensing Imagery Based on remote sensing system. Therefore, to solve the information integration in UAV remote sensing image motion blur problem has the urgent practical needs. The UAV remote sensing image motion blur is mainly composed of between the imaging device and the large scale scene caused by the relative motion. According to the motion blurred image, the main There are two solutions: one is image restoration algorithm using postprocessing method; the two is the use of high-speed imaging technology. However, traditional image restoration is an ill posed problem for large scale motion blurred image restoration effect is not ideal. The high-speed imaging technique requires special hardware support, price and application the scope is limited. In recent years, the development of a computational photography field exposure imaging encoding technology, creatively change fuzzy image restoration problem is a kind of morbid state and problems, provide new development space for the restoration of blurred image motion. At the same time, inspired by the theory of compressed sensing, random sampling method of encoding exposure is expected to in order to achieve high speed sampling of low frame rate imaging. This paper to study the optimization of the encoding lithography technology as the starting point, from the aspects of fuzzy image restoration and high-speed imaging of two is discussed How to get a clear image of target existence problem of large scale of relative motion between the imaging device and the filming of the scene. The main innovation points and research achievements include: (1) proposed a Legendre sequence and genetic algorithm encoding code word searching method based on optimal exposure exposure. The optimal codeword search space encoding as the codeword length grows exponentially, the searching method can only be used for short length codeword search. In this paper, information theory in the field of low cross-correlation sequence of Legendre encoding for exposure imaging, by simple rotation and extension operation to further improve the performance of the code on the Legendre sequence, improved Legendre the improved Legendre sequence. Sequence as the initial population of genetic algorithm, the improved Legens using crossover and mutation operator of genetic algorithm To further optimize the de sequence, obtained the optimal exposure. The encoding codeword searching method can well satisfy the encoding technology for the reversible exposure code requirements, and can reduce the noise image. The method makes full use of the Legendre sequence can be calculated by the formula of fast advantage, which makes the method efficiency high, suitable for search long codewords. (2) proposed an optimal exposure searching method of CCD sensor noise is considered. The exposure mode encoding encoding of CCD noise is more serious, from the practical angle, fully considering the influence of sensor noise need to search the optimal codeword. Based on affine noise model analysis of sensor noise component encoding exposure camera; then gives the optimal codeword structure conditions under photon noise and image SNR gain The analytical expression of interests, and the real encoding of noise exposure camera calibration; the searching method has more than one using approximate search time, low efficiency, and finally put forward the CCD noise is considered to find the optimal codeword fitness function, and genetic algorithm is adopted to search the optimal codeword encoding exposure camera. The algorithm not only improves the codeword search efficiency, and greatly improve the image SNR. (3) proposed a statistical information analysis of energy spectrum image encoding exposure image fuzzy scale automatic estimation method based on fuzzy information encoding. Exposure image is not continuous, the fuzzy scale current estimation method of losing effectiveness. Based on the analysis of fuzzy image encoding exposure energy spectrum data, found that the energy spectrum of the statistical data of image restoration and image point of divergence degree Closely related to the degree of fuzzy input scale correctly in the course of restoration. Only use fuzzy scale correctly, obtained by deconvolution restoration after image energy spectrum statistics point of divergence is minimized. Based on this, this paper fitting the image energy spectrum data, and the residual square size to represent the divergence the degree of image energy spectrum statistical data points. Encoding image restoration can be used direct exposure deconvolution algorithm is fast, so the iterative optimization method based on Fuzzy scale to estimate the exposure encoding fuzzy image. Given an initial fuzzy scale, using a global search algorithm, it can quickly get the correct exposure encoding fuzzy scale the blurred image. (4) proposed a single encoding method for high speed video camera exposure reconstruction based on fully considering the transfer in the modern. The sense of easy ability to achieve device on, to further explore the possibility of using a single encoding reconstruction high-speed video frame exposure camera. Considering the different spatial and temporal redundancies of high-speed video signal, the paper proposes using the Kronecker product line to construct three-dimensional hyperbolic wavelet and sparse representation of spatial redundancy and temporal redundancy of high-speed video signal based on the different decomposition scales. Wavelet basis can also signal, the hyperbolic structure of the various dimensions of different modeling of high speed video, just to meet the high speed video signal has a different nature smoothness in the space and time dimension. In order to further improve the quality of the reconstructed video frame, this paper makes full use of the 1 Characteristics of convex optimization model can be further? Incorporating prior knowledge into high speed video signal, the inter frame correlation in the reconstruction model of total variation regularization In order to further model the temporal redundancy of video, and finally minimize a convex optimization problem, a high quality video frame is reconstructed from the low frame rate encoded exposure image sequence.

【学位授予单位】：国防科学技术大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TP751

【参考文献】