基于单应性估计的多帧超声TOFD图像配准与降噪技术研究

发布时间：2018-05-11 15:59

本文选题：小波变换 + 阈值函数　；参考：《西南交通大学》2017年硕士论文

【摘要】：图像配准技术是图像处理与模式识别领域中非常重要的一部分。图像配准在目标检测、超分辨率重建、运动轨迹估计、特征识别、医学影像、地理遥感图像等领域都有广泛的应用。每一种配准方法都会针对研究领域特定的需要而设计,所有的图像配准方法的共性就是要在不同的空间中寻找一种最优的空间变换模型,这种变换能够使两幅图像在某种意义上达到匹配。对于不同的应用领域,图像特征匹配具有不同的意义,就需要采用具有针对性的方法。根据模板图像和待配准图像之间的关系,将图像配准分为多源图像配准、基于模板的配准、多角度图像配准、时间序列图像配准四大类。本文主要讨论基于模板图像的多帧图像配准,在技术层面主要是基于图像特征的图像配准算法。采用一种叠加的图像配准技术来进行讨论和研究,来提升图像的信噪比等相关参数。本文针对超声TOFD图像在信噪比提升方面进行了讨论和研究。系统的介绍和分析了小波图像降噪算法相关技术;采用贝叶斯估计理论进行噪声参数估计;接下来,本文提出了一种新的阈值函数对高频小波系数进行压缩处理。然后,将处理后的细节小波系数结合近似小波系数进行反小波变换,得到新的信号;最后,采用基于最大相关的图像配准方法消除数据采集过程中的偏差,经过多幅配准图像的叠加得到最终的去噪图像数据。相对于传统方法,信噪比提高了 12.2dB,提升幅度约为99%。本文采用单应性估计的方法进行图像配准。这里,对单应性矩阵的计算方法进行了讨论和分析。分别采用SIFT算法和SURF算法提取特征点和特征描述向量的计算,生成模板图像和待配准图像之间的对应匹配关系。然后分别采用有向线性变换和Bundle Adjustment方法进行单应性矩阵的计算。对于超声TOFD图像质量的提升,在采用降噪算法对图像进行预处理后,采用基于图像特征的方法进行图像配准。首先,采用DoG算子提取图像中的特征点,然后对每个特征点采用SIFT算法进行标记;在模板图像(fixed image)和待配准图像(float image)之间的特征点通过最小欧式距离原则形成的对应关系,然后采用RANSAC算法对形成的对应关系进行优化,将其中部分误匹配点去除,计算出两幅图像之间的单应性矩阵。最后,利用单应性矩阵进行仿射变换,完成图像配准。利用单应性矩阵进行图像配准可以对图像进行伸缩、旋转等变换。相对于基于最大相关方法的配准结果,信噪比提升了约2dB,提升幅度约为10%。为了克服经典SIFT、SURF等配准算法中的不足,本文讨论了一种多单应性矩阵局部图像配准的方法。提出一种AGAST-DAISY特征点提取标记的方法,结合红黑树匹配算法实现图像配准。首先采用AGAST算法对图像中的特征点进行提取,在选取特征点标记时采用DAISY描述子,为了加快运算速度,本文提出一种基于三阶统计量的主成分提取算法进行降维,然后在模板图像和待配准图像之间采用最小欧式距离原则进行特征点匹配。对于采用传统RANSAC算法的随机性带来的不利影响,本文采用基于聚类一致性原则的方法对其进行改进。首先,分别采用在Jeffrey散度和Gap统计量原则下选取聚类数目,通过多种实验数据的计算机模拟验证,两种原则下选取出的聚类数目相同。在确定聚类数目后,采用C-means模糊聚类的方法分别对模板图像和待配准图像进行聚类,形成C个聚类中心。采用红黑树数据结构进行匹配,即对应的每个类都会形成一个单应性变换矩阵。通过对应类的单应性矩阵仿射变换,实现图像的局部配准,有效的消除了传统RANSAC算法随机性带来的影响。从而提高了配准精度。对于超声T0FD图像在信噪比方面,随着叠加数目的增加,信噪比提升1dB左右,均方根误差低于最大相关配准方法0.01左右,约为最大相关配准方法的85%。对于土星图像,相对于传统方法信噪比大约有10dB的提升,对比度噪声比提升了 10左右,均方根误差约下降了 0.02,大约为传统方法的50%。对于火星图像,信噪比有5dB的提升,对比度噪声比提升了约45,均方根误差下降了 0.02左右,约为传统方法的60%。
[Abstract]:Image registration is a very important part of image processing and pattern recognition. Image registration is widely used in the fields of target detection, super resolution reconstruction, motion trajectory estimation, feature recognition, medical image, geographic remote sensing image and so on. Each registration method will be designed for the specific needs of the research field, all of which are designed. The similarity of image registration methods is to find an optimal space transformation model in different spaces. This transformation can make the two images match in a certain sense. For different application fields, the image feature matching has different meanings. It is necessary to adopt a targeted method. The relationship between quasi images is divided into multi source image registration, based on template registration, multi angle image registration, and time series image registration. This paper mainly discusses the multi frame image registration based on template image, and mainly is image registration algorithm based on image feature on the technical level. A superimposed image registration is adopted in this paper. The technology is discussed and studied to improve the signal to noise ratio and other related parameters. In this paper, the ultrasonic TOFD image is discussed and studied in the aspect of signal to noise ratio enhancement. The system is introduced and analyzed the correlation technology of wavelet image denoising algorithm; Bayesian estimation theory is used to estimate the noise parameters. The new threshold function compresses the high frequency wavelet coefficients. Then, the wavelet coefficients after the processing are combined with the approximate wavelet coefficients to carry out the anti wavelet transform, and the new signals are obtained. Finally, the error in the data acquisition process is eliminated by the image registration method based on the maximum correlation, and the superposition of multiple registration images is finally obtained. Compared to the traditional method, the signal to noise ratio is improved by 12.2dB and the amplitude of the lifting is about 99%.. This paper uses the method of single stress estimation to register the image. Here, the calculation method of the single echo matrix is discussed and analyzed. The calculation of the feature point and the feature description vector is obtained by using the SIFT algorithm and the SURF algorithm respectively. The corresponding matching relationship between the template image and the image to be registered. Then the mono homography matrix is calculated using the directed linear transformation and the Bundle Adjustment method. For the improvement of the quality of the ultrasonic TOFD image, the image registration based on the image feature is adopted after the image is preprocessed with the denoising algorithm. The feature points in the image are extracted by the DoG operator, and then the SIFT algorithm is used to mark each feature point. The corresponding relationship between the template image (fixed image) and the registration image (float image) is formed by the minimum Euclidean distance principle, and then the corresponding relationship is optimized by the RANSAC algorithm. The single echo matrix between two images is calculated. Finally, the image registration is completed by using the homography matrix to carry out affine transformation. The image registration can be telescopic and rotated by using the single stress matrix. The signal to noise ratio is improved by about 2dB, and the signal to noise ratio is improved by about 2dB. The amplitude is about 10%. in order to overcome the shortcomings of the classical SIFT, SURF and other registration algorithms. In this paper, a method of local image registration of multiple homography matrix is discussed. A method of extracting AGAST-DAISY feature points is proposed, which combines the red black tree matching algorithm to realize the image registration. First, the AGAST algorithm is used to extract the feature points in the image. The DAISY descriptor is used in the selection of feature points. In order to speed up the operation speed, a principal component extraction algorithm based on three order statistics is proposed to reduce the dimension, and then the feature points are matched with the minimum Euclidean distance principle between the template image and the image to be registered. In this paper, we adopt the method of clustering consistency principle to improve it. Firstly, the number of clustering is selected under the principle of Jeffrey divergence and Gap statistics, and the number of clusters selected under the two principles is the same. The number of clustering is the same. After determining the number of clustering, the number of clustering is used to be fuzzy. Clustering methods are used to cluster the template image and the image to be registered to form C cluster centers. Using the red black tree data structure to match, that is, each class will form a homography transformation matrix. The local registration of the image is realized by the affine transformation of the corresponding homography matrix of the corresponding class, which effectively eliminates the traditional RANSAC calculation. The accuracy of registration is improved. As to the signal to noise ratio of the ultrasonic T0FD image, with the increase of the number of superposition, the signal to noise ratio is raised about 1dB, the root mean square error is less than 0.01 of the maximum correlation registration method, and the 85%. pair of the largest correlation registration method is about the image noise ratio of the traditional method relative to the traditional method. With the enhancement of 10dB, the contrast noise ratio increased by about 10, and the mean square root error decreased by 0.02. It was about the traditional method of 50%. for Mars images, the signal to noise ratio was raised by 5dB, the contrast noise ratio was raised about 45, the root mean square error was reduced by about 0.02, about the traditional method of 60%..

【学位授予单位】：西南交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】