【作者】 熊波；

【导师】 尹周平； 李文龙；

【作者基本信息】 华中科技大学 ， 机械电子工程， 2012， 博士

【摘要】 图像处理技术是电子封装设备(如无线射频识别(RFID:Radio Frequency Identification)封装设备)开发的关键技术。实际应用中,受环境、机械振动等因素的影响,数字图像在获取和传输过程中会不可避免地被噪声污染,严重影响视觉定位与RFID芯片的精密操作Pick-and-Place,进而影响整台设备的运行性能。传统的图像去噪方法在去除噪声的同时难以保存图像的边缘等细节特征,为此,本文系统地研究了脉冲噪声、高斯噪声以及两者混合噪声的保边缘图像去噪算法,并应用到自主开发的RFID设备及视觉软件平台FAMT_MV中。主要研究内容和创新性成果包括：(1)提出了基于八方向搜索的自适应加权均值脉冲噪声(EDS-AWM:Eight Directional Searching-Adaptive Weighted Mean)滤除方法,解决了中值类滤波方法的缺陷。论文深入分析了中值类脉冲噪声滤波方法的缺陷：1)在滤波阶段,通常选用未受污染的像素计算中值时,未考虑像素的分布均匀性；2)滤波的结果完全决定于一个中值像素,只考虑了局部像素之间的大小关系,而没有考虑像素与像素之间的其它关联性以及非局部信息。为此,论文提出了基于八方向的搜索方法：以被处理像素为中心,沿着八个不同的方向进行搜索,成功解决了像素分布的均匀性问题,同时引入了非局部的思想；引入了用于去除高斯噪声的加权均值滤波的思想,提出了基于空间相似性和灰度相似性的自适应权重计算,实现了对脉冲噪声的滤波。由于EDS-AWM在滤波过程中使用了更多的像素及其它相关信息,提高了算法的保边缘能力。(2)提出了基于权重对称技术和滑动平均的快速非局部均值高斯噪声滤除算法(FNLM:Fast Nonlocal-means),解决了非局部均值(NLM:Nonlocal-means)的效率问题。原始NLM效率低下,无法得到实际应用,基于此,论文提出了权重对称技术和滑动平均的加速方法,提高了算法的效率。考虑到算法在加速的同时会引起质量下降,论文提出的方法结合了改进权重函数,提高了算法对异常图像块的处理能力,降低了异常图像块对去噪质量的影响；并从统计学的角度,研究了残余图像的性质,从理论上证明残余图像中含有一定的结构信息,分析其对提高图像去噪质量的可行性,提出了一种可行的应用方案,使得算法的质量得到了进一步的提升。在相似性图像块的选取问题上,提出了基于结构相似性(SSIM:Structural Similarity)的选择方法,提高了选取的稳定性及对噪声的鲁棒性。(3)提出了基于稳健异常度比率和非局部均值的混合噪声滤除方法(ROR-NLM: Robust Outlyingness Ratio-Nonlocal-means),解决了噪声检测和NLM用于脉冲噪声去除的问题。论文首次提出了用于描述像素异常程度的稳健异常度比率(ROR:Robust Outlyingness Ratio)统计量,并根据ROR将图像中的像素分成四个不同的异常度水平,在每一个水平中采用不同的噪声检测方式,构建了全新的噪声检测框架。在检测的过程中,为了提高检测方法的精度以及对噪声密度的稳健性,论文引入了由粗到精和迭代的检测策略。由于NLM原始是用于高斯噪声的去除,通过引入参考图像,论文成功解决了将NLM应用于脉冲噪声去除中的问题,首次实现了NLM对脉冲噪声的去除,并结合提出的噪声检测方法构建了新的混合噪声去除框架ROR-NLM。(4)在RFID封装设备对图像处理技术的需求上,数字制造装备与技术国家重点实验室开发了统一的视觉软件平台FAMT_MV。将以上研究成果融于FAMT_MV中,并对RFID设备中的图像进行了实验验证,为在同一平台下实现不同图像处理技术的应用提供了技术支撑。更多还原

【Abstract】 The image processing technology is one of the critical technologies in the electronic packaging equipments such as Radio Frequency Identification(RFID) equipments. In the practical applications, digital images could be contaminated by noise during image acquisition and transmission due to the environment, mechanical vibration and some other factors, and severely impede vision location and the precision operation(Pick-and-Place) of the chips. Then, the performance of the RFID equipments is affected. Traditional image denoising methods did not well preserve the details when removing the noise, to this end, this paper systematically studies the denoising algorithms about impulse noise, Gaussian noise, and mixed noise, and these developed algorithms are applied to the independent development RFID equipment and the vision software platform FAMT_MV. The major research works and contributions of the thesis are as follows:(1)A novel algorithm for removing impulse noise, i.e. Eight Directional Searching-Adaptive Weighted Mean(EDS-AWM), is proposed, and it solves the drawbacks of the median-based algorithms. This thesis deeply analyzes the drawbacks of the median-based algorithms:1)in the filtering stage, the median-based algorithms did not consider the uniformity of the distribution of the pixels which used for calculating the median; 2)the filtering result was absolutely from the median pixel, and only considered the size relations of the local pixels, did not take into account other relativity and nonlocal information. On the basis, the thesis proposes a new searching method based on eight directions, that is, searching uncorrupted pixels in the eight different directions around the processed pixel, and introduces the nonlocal spirit; the spirit of the weighted mean filtering, which is used for Gaussian noise, is applied to remove impulse noise, and the weights can be adaptively calculated by using some weight function based on space similarity and gray similarity. For filtering with more pixels and other relative information, the EDS-AWM algorithm preserves more details.(2)A novel algorithm for removing Gaussian noise, i.e. Fast Nonlocal-means(FNLM), is proposed, and it solves the efficiency of the Nonlocal-means(NLM). The efficiency of the original Nonlocal-means(NLM) was very low, and it is difficult for using NLM in practical applications. In order to solve this problem, this thesis proposes a way to speed up NLM based on weight symmetry and moving average. Considering the quality may be reduced when speeding up, the presented method combines the modified weight function, and the ability of the algorithm to deal with the outlying image patches is improved. Furthermore, the thesis studies the properties of the residual image from the statistics perspective, and proves that the residual image contains some structural information in theory. And then propose a feasible framework about using the residual image for improving the denoised quality. For selecting the similar image patches, the thesis proposes a new method based on SSIM (Structural Similarity), and the stability and the robustness to the noise of the method are improved.(3)A novel algorithm for removing mixed noise based on Robust Outlyingness Ratio-Nonlocal-means(ROR-NLM) is proposed, and it solves the impulse noise detection and the problem of using NLM for removing impulse noise. The thesis proposes a new statistics called ROR for measuring the outlyingness of each pixel in the image at the first time, and the pixels are divided into four different outlyingness levels based on the ROR. Then the noise is detected in each level with different rules. In order to improve the detection precision and the robustness to the noise density, the coarse to fine strategy and the iterative strategy are used. Meanwhile, the NLM is successfully extended to remove the impulse noise by introducing a reference image for the first time, and a new framework for removing mixed noise is proposed by combining the new detection method.(4)According to the needs of the image processing technology in the RFID equipment, the State Key Laboratory of Digital Manufacturing Equipment and Technology developed the universal image processing software platform FAMT_MV. The above researches are integrated in the FAMT_MV, and the algorithms are tested by using the practical images from the RFID equipment. The FAMT_MV provides support for achieving different image processing operations in the uniform platform.更多还原