图像实时复原技术的研究与应用

发布时间：2018-05-03 01:00

本文选题：图像复原 + 图像去雾　；参考：《中国科学院研究生院(光电技术研究所)》2016年博士论文

【摘要】：图像复原技术作为图像处理中交叉学科,在图像处理领域中,一直是最重要和最基本的研究课题之一,具有很强的理论价值和工程应用价值。本文研究的图像复原技术将图像复原理论与工程实践密切结合,目的是在复原质量的可靠保证前提下,尽可能满足实时处理场合的图像复原需求。复原算法主要针对图像复原处理中的细节保护和噪声抑制这一矛盾以及图像复原处理的快速性要求,进一步提高算法的性能和增强算法的实用性。本文的研究工作主要基于快速复原算法和高速实现两个方向展开,针对快速复原算法和当前主流的实现平台进行研究。研究常见的降质模型和快速复原算法,针对离焦和G类降质模型进行分析,论证此类降质对于自解卷积降质估计(SeDDaRA)算法的有效性。并在此基础上利用SeDDaRA算法对此类降质图像进行了有效复原。研究基于概率模型的图像盲复原算法,分析代表性算法MIA(乘性迭代算法)的原理和特点,提出了改进性的迭代模型,并从数学模型证明算法的有效性。针对乘性迭代算法收敛性慢,SeDDaRA单步逆滤波估计性不足的问题,提出了基于SeDDaRA算法的乘性迭代方法。进一步根据图像序列中的目标图像短时不变这一先验,提出了面向视频序列的两步式快速复原方法。通过选择更优的迭代路径实现快速迭代复原,实验结果表明,仅需要增加少量的计算量就可以大量减少迭代次数。研究基于暗通道的快速去雾方法,图像去雾从广义上讲,同属于图像复原问题。针对现有算法计算复杂度高,复原时间慢问题,提出基于双尺度的暗通道去雾复原算法,高尺度暗通道用于计算光照系数,低尺度暗通道用于计算透射系数。在此基础上,针对有雾的观测图像提出了在FPGA资源约束的条件下进行快速复原处理的电路结构以及实现方法。研究用高级语言进行FPGA设计的开发工具ImpluseC,有效地利用了FPGA的资源,较好地解决了浮点运算和时序设计问题,在Xilinx Spartan3平台上能够获得实时稳定的复原结果。在探索实时图像复原算法的基础上,利用FPGA平台实现了图像去雾电路。研究OpenCL技术,利用GPU平台提高图像复原算法的速度。长此以来,复原算法的复杂性是制约算法应用的一个重要方面,因此算法的实时性研究是本文算法研究的一个重要方向。对于快速复原算法,分析其计算共性,针对二维FFT和其他特殊函数,研究低复杂度的计算方法。经过算法复杂度分析,复原算法中近70%的计算量来自于2D-FFT运算。利用傅里叶变换的周期性,以及频域数据的共轭对称性减少2D-FFT中冗余数据的计算量,在此基础上,设计用于快速复原的FFT计算模块,用于复原算法的GPU加速。对于相同浮点处理能力的平台,加速比能够达到4倍以上。在此基础上,对GPU在存储层次上进行了分析,探究PCIE和GPU传输带宽的瓶颈,以及存储器访问的方式,进而找到影响OpenCL程序性能的关键。然后在分析两步式快速迭代复原算法的基础上,总结该算法的特点,对算法的计算模块进行拆解,在AMD 7400平台和通用的小型化GPU平台上找到快速有效的实现方式。通过充分利用GPU的计算资源,算法有效移植到GPU平台,两种平台对算法进行加速后,分别能够获得6倍和30倍的可观加速比。
[Abstract]:Image restoration technology, as a cross subject in image processing, has been one of the most important and basic research topics in the field of image processing. It has a strong theoretical value and engineering application value. The image restoration technology studied in this paper combines the image restoration theory with the engineering practice, and aims to ensure the reliability of the restoration quality. On the premise, we can meet the needs of image restoration in real-time processing as far as possible. The restoration algorithm mainly aims at the contradiction of detail protection and noise suppression in image restoration and the fast requirement of image restoration processing, and further improves the performance of the algorithm and the practicability of the enhancement algorithm. The research work of this paper is mainly based on fast recovery. The algorithm and high speed realization are carried out in two directions. The fast restoration algorithm and the current mainstream implementation platform are studied. The common descending model and the fast recovery algorithm are studied. The analysis of the defocus and G class reduction model is carried out to demonstrate the effectiveness of this kind of degradation for the self deconvolution reduction estimation (SeDDaRA) algorithm. This kind of degraded image is effectively restored with the SeDDaRA algorithm. The blind restoration algorithm based on probability model is studied. The principle and characteristics of the representative algorithm MIA (multiplicative iterative algorithm) are analyzed. An improved iterative model is proposed, and the validity of the algorithm is proved from the mathematical model. The convergence of the multiplicative iterative algorithm is slow and the SeDDaRA single step is used. In this paper, a multiplicative iterative method based on SeDDaRA algorithm is proposed. The two step fast restoration method for video sequence is proposed based on a priori the short time invariant of the target image in the image sequence. Adding a small amount of computation can reduce the number of iterations in a large amount. A fast fog removal method based on dark channels is studied. The image demogging is in the broad sense, and it belongs to the image restoration problem. In view of the high complexity and slow recovery time of the existing algorithms, a double scale dark channel fog restoration algorithm is proposed, and the high scale dark channel is used for calculation. The illumination coefficient and the low scale dark channel are used to calculate the transmission coefficient. On this basis, the circuit structure and implementation method for fast recovery processing under the condition of FPGA resource constraints are proposed for the misty observed images. The development tool of the advanced language for FPGA design, ImpluseC, is used to effectively utilize the resources of FPGA. The problem of floating point operation and time series design is solved. On the Xilinx Spartan3 platform, the real-time and stable recovery results can be obtained. On the basis of exploring the real-time image restoration algorithm, the image demogging circuit is realized by using the FPGA platform. The OpenCL technology is studied and the speed of the image restoration algorithm is improved by using the GPU platform. It is an important aspect to restrict the application of the algorithm, so the study of the real-time performance of the algorithm is an important direction in the study of this algorithm. For the fast restoration algorithm, it analyzes its common calculation and studies the calculation method of low complexity for two dimensional FFT and other special functions. After the algorithm complexity analysis, the calculation of nearly 70% of the restoration algorithm is made. The quantity comes from the 2D-FFT operation. Using the periodicity of Fourier transform and the conjugate symmetry of the frequency domain data to reduce the amount of redundant data in the 2D-FFT, the FFT computing module for fast recovery is designed for the GPU acceleration of the restoration algorithm. The acceleration ratio can reach more than 4 times for the platform with the same floating point. On this basis, GPU is analyzed on the storage level, the bottleneck of the PCIE and GPU transmission bandwidth and the way of memory access are explored, and then the key to the performance of the OpenCL program is found. Then, on the basis of the analysis of the two step fast iterative restoration algorithm, the characteristics of the algorithm are summarized, and the calculation module of the algorithm is dismantled in AM. The D 7400 platform and the universal miniaturized GPU platform find a fast and effective implementation. By making full use of the computing resources of GPU, the algorithm is effectively transplanted to the GPU platform, and the two platforms can get 6 times and 30 times the significant acceleration ratio respectively after the acceleration of the algorithm.

【学位授予单位】：中国科学院研究生院(光电技术研究所)
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】