两阶段反卷积图像去模糊算法的DSP实现

发布时间：2018-04-13 16:52

本文选题：图像去模糊 + 交替方向法　；参考：《华中科技大学》2016年硕士论文

【摘要】：运动、散焦或大气湍流等非理想因素导致成像系统采集到的图像产生不同程度的模糊,而传统的目标识别方法主要依赖于目标自身的特征与表达,模糊的图像必然将加大目标检测与识别的难度。如何提高模糊图像的分辨率,为后期信息处理提供有效数据,具有重要的理论与实际意义。两阶段图像去模糊算法可以有效复原模糊图像,首先利用自然图像稀疏性先验估计模糊核,再利用非盲反卷积求解清晰图像,采用交替方向法将原本复杂的优化问题转换为若干个具有封闭解的子问题进行迭代求解;为了加快算法的收敛速度,同时防止求解过程中陷入局部极小值,采用了一种由粗到精的多尺度图像金字塔求解框架。本文在DSP平台(TMS320C6657)上实现了两阶段去模糊算法。首先分析了DSP的硬件资源、开发环境,设计了需使用的DSP外设的驱动软件。进一步,分析算法中主要模块的处理流程,对关键步骤的实现进行了详细的阐述,并通过实验确定算法参数的选取规则。最后,提出了一种基于图像水平及竖直梯度图的多节点并行优化方式,并调用DSPLIB函数库对算法中出现的FFT、IFFT、矩阵转置及向量点积等操作进行优化,同时利用EDMA3实现算法中数据块搬移、图像边界延拓及矩阵循环平移等操作。仿真和实测实验表明了两阶段图像去模糊算法在DSP上的实现对模糊图像的复原效果与matlab复原效果基本相同,而且经过对程序一系列的优化处理,算法的执行速度明显提高。
[Abstract]:Some non-ideal factors, such as motion, defocusing or atmospheric turbulence, cause the images collected by the imaging system to be blurred to varying degrees, while the traditional methods of target recognition mainly depend on the characteristics and expressions of the target itself.Blurred images will inevitably increase the difficulty of target detection and recognition.How to improve the resolution of blurred images and provide effective data for the later information processing has important theoretical and practical significance.The two-stage image de-blurring algorithm can effectively restore the blurred image. Firstly, the sparse priori estimation of the fuzzy kernel is used in the natural image, and then the non-blind deconvolution is used to solve the clear image.In order to speed up the convergence of the algorithm and avoid falling into a local minimum value, the alternating direction method is used to transform the original complex optimization problem into several subproblems with closed solutions.A multi-scale image pyramid solution framework from coarse to fine is used.In this paper, a two-stage de-blur algorithm is implemented on DSP platform TMS320C6657.Firstly, the hardware resource and development environment of DSP are analyzed, and the driver software of DSP peripheral is designed.Furthermore, the processing flow of the main modules in the algorithm is analyzed, and the key steps are described in detail, and the selection rules of the algorithm parameters are determined by experiments.Finally, a multi-node parallel optimization method based on image horizontal and vertical gradients is proposed, and the DSPLIB function library is called to optimize the operations such as FFTF, matrix transposition and vector dot product.At the same time, EDMA3 is used to realize data block moving, image boundary extension and matrix cyclic translation.The simulation and experimental results show that the effect of the two-stage image de-blurring algorithm on DSP is basically the same as that of the matlab restoration algorithm, and the speed of the algorithm is improved obviously through a series of optimization processing of the program.
【学位授予单位】：华中科技大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】