CN104834931A

CN104834931A - Improved SIFT algorithm based on wavelet transformation

Info

Publication number: CN104834931A
Application number: CN201510111168.4A
Authority: CN
Inventors: 茅正冲; 王丹; 唐雨玉
Original assignee: Jiangnan University
Current assignee: Jiangnan University
Priority date: 2015-03-13
Filing date: 2015-03-13
Publication date: 2015-08-12

Abstract

In order to solve the problem that the SIFT algorithm runs too long and the matching rate is not high, the invention proposes an improved SIFT algorithm. On the basis of the original classic SIFT algorithm, the two-dimensional Mallat fast wavelet transform algorithm is introduced to reconstruct the low-frequency components of the image, and then the number of Gaussian pyramid groups is adjusted to reduce the number of downsampling. Finally, the optimized RANSAC algorithm is used to eliminate mismatching points. . The improved algorithm not only reduces the matching time, but also improves the matching rate, which is better than the original SIFT algorithm.

Description

An Improved Scale Invariant Feature Matching Algorithm Based on Wavelet Transform

技术领域technical field

本发明涉及计算机视觉中的图像匹配领域，具体指的是一种基于小波变换的改进的尺度不变特征匹配算法。The invention relates to the field of image matching in computer vision, and specifically refers to an improved scale-invariant feature matching algorithm based on wavelet transform.

背景技术Background technique

图像匹配在近几十年来一直是人们研究的热点和难点，它是在变换空间中寻找一种或多种变换，使来自不同时间、不同传感器或者不同视角的同一场景的两幅或者多幅图像在空间上一致，目前已经应用于许多领域，其中应用最广泛的是图像配准领域和运动目标识别与跟踪领域。因此图像匹配技术占据着至关重要的地位。Image matching has been a hot and difficult point of research in recent decades. It is to find one or more transformations in the transformation space, so that two or more images of the same scene from different times, different sensors or different perspectives It is spatially consistent and has been used in many fields, among which the most widely used are image registration and moving target recognition and tracking. Therefore, image matching technology occupies a crucial position.

由于拍摄时间、拍摄角度、自然环境的变化，使得拍摄的图像受到各种噪声的影响。在这种条件下，匹配算法如何达到精度高、匹配正确率高、速度快、鲁棒性强以及并行实现成为人们追求的目标。针对这一问题，很多国内外学者进行了广泛的研究。基于匹配的基本原理将图像匹配算法分为四种：基于灰度相关的匹配、基于特征的匹配、基于模型的匹配以及基于变换域的匹配。目前，研究最广泛的是基于特征的匹配，其中最经典的算法是由D.G.Lowe于2004年完善总结的SIFT(尺度不变特征匹配)算法，该算法对平移、旋转、尺度缩放、亮度变化均能保持很好的不变性。后来，很多学者依据SIFT算法提出了不同的改进算法。Due to changes in shooting time, shooting angle, and natural environment, the captured image is affected by various noises. Under such conditions, how to achieve high precision, high matching accuracy, fast speed, strong robustness and parallel implementation of matching algorithms has become the goal pursued by people. Aiming at this problem, many domestic and foreign scholars have conducted extensive research. Based on the basic principle of matching, the image matching algorithm is divided into four types: matching based on gray correlation, matching based on feature, matching based on model and matching based on transform domain. At present, the most widely studied is feature-based matching, and the most classic algorithm is the SIFT (Scale Invariant Feature Matching) algorithm, which was perfected and summarized by D.G. Lowe in 2004. Can maintain good invariance. Later, many scholars proposed different improved algorithms based on the SIFT algorithm.

陈抒瑢等提出了Contourlet-SIFT特征匹配算法，对尺度空间下旋转不变特征进行Contourlet变换后再进行匹配，但是计算量偏大，不满足实时性要求。曹娟等提出了基于D²OG特征点检测算子的改进SIFT特征匹配算法，适用于图像信息丰富且对实时性要求较高的场合，但是算法提取的匹配点对数相对较少，限制了此算法处理的图像类型。杨幸芳提出了以街区距离代替欧氏距离作为特征描述符之间的相似性度量，降低了相似性度量公式的时间复杂度，但是没有提高鲁棒性。于丽莉提出了一种基于图像Radon变化的改进的SIFT特征匹配算法，降低了SIFT特征向量的维数，提高了特征匹配效率，但是在实际场景使用时性能有待提高。孔军提出了多尺度特征提取的双目视觉匹配，虽然匹配率得到了提高，但是匹配耗时较长，时效性较差。为了进一步提高正确匹配率、增强鲁棒性及实时性，本发明提出了基于小波变换的改进的尺度不变特征匹配算法。Chen Shurong and others proposed the Contourlet-SIFT feature matching algorithm, which performs Contourlet transformation on the rotation-invariant features in the scale space and then performs matching, but the calculation amount is too large and does not meet the real-time requirements. Cao Juan et al. proposed an improved SIFT feature matching algorithm based on the D ² OG feature point detection operator, which is suitable for occasions with rich image information and high real-time requirements, but the number of matching point pairs extracted by the algorithm is relatively small, which limits The type of image that this algorithm processes. Yang Xingfang proposed to replace the Euclidean distance with the block distance as the similarity measure between feature descriptors, which reduces the time complexity of the similarity measure formula, but does not improve the robustness. Yu Lili proposed an improved SIFT feature matching algorithm based on image Radon changes, which reduces the dimension of SIFT feature vectors and improves the efficiency of feature matching, but the performance needs to be improved when used in actual scenes. Kong Jun proposed multi-scale feature extraction for binocular vision matching. Although the matching rate has been improved, the matching takes a long time and the timeliness is poor. In order to further improve the correct matching rate, enhance robustness and real-time performance, the present invention proposes an improved scale-invariant feature matching algorithm based on wavelet transform.

发明内容Contents of the invention

本发明的目的在于针对现有方法中存在正确匹配率不高、鲁棒性不强、算法运行时间过长等问题，在原来经典尺度不变特征匹配算法的基础上，提出了基于小波变换的改进的尺度不变特征匹配算法技术，开发出一种鲁棒性强，正确匹配率高，适用于实时性要求高的场景中的图像匹配方法。The purpose of the present invention is to solve the existing problems such as low correct matching rate, low robustness, and long running time of the algorithm in the existing method. On the basis of the original classical scale-invariant feature matching algorithm, a wavelet transform-based The improved scale-invariant feature matching algorithm technology has developed an image matching method with strong robustness and high correct matching rate, which is suitable for scenes with high real-time requirements.

本发明是基于如下考虑：为了解决原SIFT算法运行时间过长，匹配率不高的问题，提出了一种改进的SIFT算法。在原来经典的SIFT算法的基础上，引入了二维Mallat快速小波变换算法，重建图像的低频成分，再对高斯金字塔组数进行调整，减少降采样次数，最后通过优化的RANSAC算法剔除误匹配点。改进后的算法不仅减少了匹配耗时，而且匹配率也得到了提高。The present invention is based on the following considerations: in order to solve the problems of long running time and low matching rate of the original SIFT algorithm, an improved SIFT algorithm is proposed. On the basis of the original classic SIFT algorithm, the two-dimensional Mallat fast wavelet transform algorithm is introduced to reconstruct the low-frequency components of the image, and then the number of Gaussian pyramid groups is adjusted to reduce the number of downsampling. Finally, the optimized RANSAC algorithm is used to eliminate mismatching points. . The improved algorithm not only reduces the matching time, but also improves the matching rate.

本发明基于小波变换的改进的尺度不变特征匹配算法的技术方案如下：The technical scheme of the improved scale-invariant feature matching algorithm based on wavelet transform in the present invention is as follows:

(1)对待配准的两幅图像进行二维Mallat小波变换分解，得到图像的低频成分和水平与垂直高频成分，舍弃经小波变换分解后的高频成分，对图像的低频成分进行重构，得到新的图像；(1) Decompose the two images to be registered by two-dimensional Mallat wavelet transform to obtain the low-frequency components and horizontal and vertical high-frequency components of the image, discard the high-frequency components decomposed by wavelet transform, and reconstruct the low-frequency components of the image , get a new image;

(2)利用两幅新图像构造尺度空间(DoG)，通过图像与不同尺度的高斯差分卷积核生成高斯金字塔，由于新图像经过小波变换，舍弃了部分信息，在构建高斯差分金字塔时，减少降采样次数，比原始生成层数少一层，减少构造尺度空间的时间；(2) Use two new images to construct the scale space (DoG), and generate a Gaussian pyramid through the image and Gaussian difference convolution kernels of different scales. Since the new image undergoes wavelet transformation, part of the information is discarded. When constructing the Gaussian difference pyramid, reduce The number of downsampling is one layer less than the original generation layer, which reduces the time for constructing the scale space;

(3)在DoG空间寻找关键点，把中间的检测点和它同尺度的8个相邻点和上下相邻尺度对应的18点共26个点进行比较，确保最终能够获得足够多的关键点。(3) Find key points in the DoG space, compare the middle detection point with 8 adjacent points of the same scale and 18 points corresponding to the upper and lower adjacent scales, a total of 26 points, to ensure that enough key points can be obtained in the end .

(4)由于DoG值对噪声和边缘比较敏感，在上面DoG尺度空间中检测到局部极值点还要经过进一步的检验才能精确定位为特征点。通过拟合三维二次函数来精确确定关键点的位置和尺度。(4) Since the DoG value is sensitive to noise and edges, the local extremum points detected in the above DoG scale space need further inspection before they can be accurately positioned as feature points. Precisely determine the position and scale of key points by fitting a 3D quadratic function.

其梯度值m(x,y)和方向θ(x,y)的计算表达式为The calculation expression of its gradient value m(x, y) and direction θ(x, y) is

{[L(x+1,y)-L(x-1,y)]²+[L(x,y+1)-L(x,y-1)]²}^1/2 {[L(x+1,y)-L(x-1,y)] ² +[L(x,y+1)-L(x,y-1)] ² } ^1/2

$θ θ ((x x,, y the y)) = = {tan the tan}^{- - 11} [[\frac{L L ((x x,, y the y + + 11)) - - L L ((x x,, y the y - - 11))}{L L ((x x + + 11,, y the y)) - - L L ((x x - - 11,, y the y))}]]$

(5)特征点描述通过对关键点周围区域进行图像分块，计算各块内的梯度直方图，生成独特性的向量描述符。为了增强匹配的稳健性，对每个关键点使用4×4共16个种子点来描述，这样对于一个关键点就可以产生128个数据，最终形成128维的SIFT特征向量。(5) Feature point description By dividing the image into blocks around the key point, the gradient histogram in each block is calculated to generate a unique vector descriptor. In order to enhance the robustness of matching, a total of 16 seed points of 4×4 are used to describe each key point, so that 128 data can be generated for one key point, and finally a 128-dimensional SIFT feature vector is formed.

(6)当两幅图像的SIFT特征向量生成后，下一步我们采用关键点特征向量的欧式距离来作为两幅图像中关键点的相似性判定度量。取样本特征点和样本特征点欧氏距离最近的特征点，在这两个特征点中，如果最近的距离除以次近的距离少于某个比例阈值，则接受这一对匹配点。降低这个比例阈值，SIFT匹配点数目会减少，但更加稳定。在计算特征点之间的欧氏距离时，采用了BBF算法来处理128维的特征向量。(6) After the SIFT feature vectors of the two images are generated, in the next step we use the Euclidean distance of the key point feature vectors as the similarity judgment measure of the key points in the two images. Take the sample feature point and the feature point with the closest Euclidean distance to the sample feature point. Among the two feature points, if the nearest distance divided by the second closest distance is less than a certain ratio threshold, the pair of matching points is accepted. Lowering this ratio threshold will reduce the number of SIFT matching points, but it will be more stable. When calculating the Euclidean distance between the feature points, the BBF algorithm is used to process the 128-dimensional feature vector.

(7)在对所有特征点进行粗匹配之后，使用RANSAC算法估计两个图像对之间的单位变换矩阵并将其作为几何约束，进而去除一些误匹配点，完成图像之间的精确匹配，提高匹配效率。(7) After roughly matching all the feature points, use the RANSAC algorithm to estimate the unit transformation matrix between the two image pairs and use it as a geometric constraint, and then remove some mismatching points, complete the precise matching between images, and improve matching efficiency.

本发明的有益效果：本发明采用二维Mallat小波变换对图像进行分解与重构，舍弃图像的包含大量噪声和少量有用信息的高频成分，仅对图像的低频成分进行重构。经二维小波变换处理后的待配准图像，一方面可以减少每次参与匹配的像素点，提高了匹配速度，另一方面减少了弱匹配点，从而使误匹配率下降。本发明在构造尺度空间时，由于高斯差分金字塔的最后一层包含了很少的兴趣点，对最终的匹配结果没有太大的影响，而且经过图像小波变换处理后包含的信息也减少一部分，所以可以减少降采样次数，使高斯差分金字塔的层数减少一层，减少构造尺度空间的时间，同时还可以去除部分误匹配点，提高匹配效率。本发明在对图像所有特征点进行粗匹配后使用RANSAC算法估计两个图像对之间的单位变换矩阵并将其作为几何约束，进而去除一些误匹配点，完成图像之间的精确匹配，提高匹配效率。Beneficial effects of the present invention: the present invention uses two-dimensional Mallat wavelet transform to decompose and reconstruct images, abandons high-frequency components of images containing a large amount of noise and a small amount of useful information, and only reconstructs low-frequency components of images. The image to be registered after the two-dimensional wavelet transform can reduce the number of pixels involved in each match and improve the matching speed. When the present invention constructs the scale space, since the last layer of the Gaussian difference pyramid contains few interest points, it does not have much influence on the final matching result, and the information contained after image wavelet transform processing is also reduced, so The number of downsampling can be reduced, the number of layers of the Gaussian difference pyramid can be reduced by one layer, and the time for constructing the scale space can be reduced. At the same time, some mismatching points can be removed to improve the matching efficiency. The present invention uses the RANSAC algorithm to estimate the unit transformation matrix between two image pairs after performing rough matching on all feature points of the image, and uses it as a geometric constraint, and then removes some mismatching points, completes accurate matching between images, and improves matching efficiency.

附图说明Description of drawings

图1是SIFT算法步骤。Figure 1 shows the steps of the SIFT algorithm.

具体实施方式Detailed ways

为使本发明的目的、技术方案和优点更加清楚明白，下面结合具体实例，并参照附图，对本发明的具体实施方式作详细说明，本发明包含但不限于所举实例。In order to make the purpose, technical solutions and advantages of the present invention clearer, the specific implementation of the present invention will be described in detail below in conjunction with specific examples and with reference to the accompanying drawings. The present invention includes but is not limited to the cited examples.

本发明的具体步骤如下：Concrete steps of the present invention are as follows:

1、二维Mallat小波变换1. Two-dimensional Mallat wavelet transform

Mallat在小波变换多分辨率分析理论与图像处理的应用研究中受到塔式算法的启发，提出了信号的塔式多分辨率分解与重构的快速算法，即为马拉特(Mallat)算法。Inspired by the tower algorithm in the application research of wavelet transform multi-resolution analysis theory and image processing, Mallat proposed a fast algorithm for tower multi-resolution decomposition and reconstruction of signals, which is the Mallat algorithm.

使用两个相同的一维小波函数和一维尺度函数的乘积构造二维小波基，生成的尺度函数和三个小波函数分别为：Use the product of two identical one-dimensional wavelet functions and one-dimensional scaling functions to construct a two-dimensional wavelet base, and the generated scaling functions and three wavelet functions are:

$\{\begin{matrix} φ φ ((x x,, y the y)) = = φ φ ((x x)) φ φ ((y the y)) \\ {ψ ψ}^{11} ((x x,, y the y)) = = φ φ ((x x)) ψ ψ ((y the y)) \\ {ψ ψ}^{22} ((x x,, y the y)) = = ψ ψ ((x x)) φ φ ((y the y)) \\ {ψ ψ}^{33} ((x x,, y the y)) = = ψ ψ ((x x)) ψ ψ ((y the y)) \end{matrix}$

设为待分析的图像信号，其二维逼近图像为set up is the image signal to be analyzed, and its two-dimensional approximation image is

${A A}_{i i} f f = = {A A}_{j j + + 11} f f + + {D D.}_{j j + + 11}^{11} f f + + {D D.}_{j j + + 11}^{22} f f + + {D D.}_{j j + + 11}^{33} f f$

式中In the formula

$\{\begin{matrix} {A A}_{j j + + 11} f f = = {Σ Σ}_{m m = = - - \infty \infty}^{\infty \infty} {Σ Σ}_{n no = = - - \infty \infty}^{\infty \infty} {c c}_{j j + + 11} ((m m,, n no)) {φ φ}_{j j + + 11} ((m m,, n no)) \\ {D D.}_{j j + + 11}^{i i} f f = = {Σ Σ}_{m m = = - - \infty \infty}^{\infty \infty} {Σ Σ}_{n no = = - - \infty \infty}^{\infty \infty} {d d}_{j j + + 11}^{i i} ((m m,, n no)) {φ φ}_{j j + + 11} ((m m,, n no)) \end{matrix} ((i i = = 1,2,3 1,2,3))$

利用尺度函数和小波函数的正交性，可得：Using the orthogonality of scaling function and wavelet function, we can get:

${c c}_{j j + + 11} ((m m,, n no)) = = {Σ Σ}_{k k = = - - \infty \infty}^{\infty \infty} {Σ Σ}_{l l = = - - \infty \infty}^{\infty \infty} h h ((k k - - 22 m m)) h h ((l l - - 22 n no)) {c c}_{j j} ((k k,, l l))$

上式说明，j+1尺度空间的尺度系数c_j+1(m,n)可以由j尺度空间的尺度系数c_j(k,l)经二维滤波器系数进行加权求和得到。The above formula shows that the scale coefficient c _j+1 (m,n) of the j+1 scale space can be obtained by weighting and summing the scale coefficient c _j (k,l) of the j scale space through the two-dimensional filter coefficients.

又again

$\{\begin{matrix} {d d}_{j j + + 11}^{11} = = {Σ Σ}_{k k = = - - \infty \infty}^{\infty \infty} {Σ Σ}_{l l = = - - \infty \infty}^{\infty \infty} h h ((k k - - 22 m m)) g g ((l l - - 22 n no)) {c c}_{j j} ((k k,, l l)) \\ {d d}_{j j + + 11}^{22} = = {Σ Σ}_{k k = = - - \infty \infty}^{\infty \infty} {Σ Σ}_{l l = = - - \infty \infty}^{\infty \infty} g g ((k k - - 22 m m)) h h ((l l - - 22 n no)) {c c}_{j j} ((k k,, l l)) \\ {d d}_{j j + + 11}^{33} = = {Σ Σ}_{k k = = - - \infty \infty}^{\infty \infty} {Σ Σ}_{l l = = - - \infty \infty}^{\infty \infty} g g ((k k - - 22 m m)) g g ((l l - - 22 n no)) {c c}_{j j} ((k k,, l l)) \end{matrix}$

引入矩阵算子，令H_r和H_c分别表示用尺度滤波器系数对阵列的行和列作用的算子，G_r和G_c分别表示用小波滤波器系数对行和列作用的算子，则二维Mallat分解算法为Introducing a matrix operator, let H _r and H _c respectively denote the array with scaling filter coefficients The operator acting on the rows and columns of , G _r and G _c represent the operators acting on the rows and columns with wavelet filter coefficients respectively, then the two-dimensional Mallat decomposition algorithm is

$\{\begin{matrix} {C C}_{j j + + 11} = = {H h}_{r r} {H h}_{c c} {C C}_{j j} \\ {D D.}_{j j + + 11}^{11} = = {H h}_{r r} {H h}_{c c} {C C}_{j j} \\ {D D.}_{j j + + 11}^{22} = = {G G}_{r r} {H h}_{c c} {C C}_{j j} \\ {D D.}_{j j + + 11}^{33} = = {G G}_{r r} {G G}_{c c} {C C}_{j j} \end{matrix} ((j j = = 0,1 0,1,, . . . . . . . .,, J J))$

二维Mallat重构算法为The two-dimensional Mallat reconstruction algorithm is

${C C}_{j j} = = {H h}_{r r}^{* *} {H h}_{c c}^{* *} {C C}_{j j + + 11} + + {H h}_{r r}^{* *} {G G}_{c c}^{* *} {D D.}_{j j + + 11}^{* *} + + {G G}_{r r}^{* *} {H h}_{c c}^{* *} {D D.}_{j j + + 11}^{22} + + {G G}_{r r}^{* *} {G G}_{c c}^{* *} {D D.}_{j j + + 11}^{33}$

利用上述二维Mallat小波变换算法对图像进行分解与重构，得到新的图像。The above-mentioned two-dimensional Mallat wavelet transform algorithm is used to decompose and reconstruct the image to obtain a new image.

2、改进的SIFT算法2. Improved SIFT algorithm

SIFT算法是在不同的尺度空间上查找关键点，并计算出关键点的方向。SIFT算法的实现包括两个方面，一个是图像特征点的生成，另一个是不同图像之间SIFT特征点的匹配。SIFT特征提取可分解为四个步骤，如图1所示。The SIFT algorithm is to find key points in different scale spaces and calculate the direction of the key points. The implementation of SIFT algorithm includes two aspects, one is the generation of image feature points, and the other is the matching of SIFT feature points between different images. SIFT feature extraction can be decomposed into four steps, as shown in Figure 1.

2.1构造尺度空间2.1 Construct scale space

一幅二维图像I(x,y)的尺度空间L(x,y,σ)定义为一个变化尺度的高斯函数与原图像的卷积，即The scale space L(x,y,σ) of a two-dimensional image I(x,y) is defined as the convolution of a Gaussian function with varying scales and the original image, namely

L(x,y,σ)＝G(x,y,σ)*I(x,y)L(x,y,σ)=G(x,y,σ)*I(x,y)

式中，σ为尺度空间的空间尺度因子，G(x,y,σ)为高斯核函数，其定义为where σ is the spatial scale factor of the scale space, G(x,y,σ) is the Gaussian kernel function, which is defined as

$G G ((x x,, y the y,, σ σ)) = = \frac{11}{{22 πσ πσ}^{22}} \cdot \cdot exp exp ((- - (({x x}^{22} + + {y the y}^{22})) / / {22 σ σ}^{22}))$

为了能在尺度空间中检测到稳定的关键点，使用高斯差分(DoG)算子近似尺度归一化的拉普拉斯——高斯(LoG)算子。通过图像与不同尺度的高斯差分卷积核生成：In order to detect stable keypoints in the scale space, the difference of Gaussian (DoG) operator is used to approximate the scale-normalized Laplacian-Gaussian (LoG) operator. Generated by image and Gaussian difference convolution kernels of different scales:

D(x,y,σ)＝L(x,y,kσ)-L(x,y,σ)D(x,y,σ)=L(x,y,kσ)-L(x,y,σ)

高斯金字塔共O组、S层，一般取为4组、5层，每组第一层是由上一组最后一层降采样得到的。在构造尺度空间之前，本文已先利用二维小波变换对待配准图像进行预处理，舍弃了高频成分，只保留了低频成分对其进行重建，所以无需做和原图大小时一样的降采样次数。因此，本发明针对高斯金字塔的组数进行调整，减少降采样次数，使高斯差分金字塔的层数减少一层。The Gaussian pyramid has a total of O groups and S layers, which are generally taken as 4 groups and 5 layers. The first layer of each group is obtained by down-sampling the last layer of the previous group. Before constructing the scale space, this paper has used the two-dimensional wavelet transform to preprocess the registration image, discarding the high-frequency components, and only retaining the low-frequency components for reconstruction, so there is no need to down-sample the same size as the original image frequency. Therefore, the present invention adjusts the number of groups of the Gaussian pyramid, reduces the number of times of downsampling, and reduces the number of layers of the Gaussian difference pyramid by one layer.

2.2极值点检测2.2 Extreme point detection

关键点是由DoG空间的局部极值点组成的。为了寻找DoG函数的极值点，把中间的检测点和它同尺度的8个相邻点和上下相邻尺度对应的18点共26个点进行比较，最终确保在尺度空间和二维图像空间都能够检测到极值点。Keypoints are composed of local extremum points in the DoG space. In order to find the extreme point of the DoG function, the middle detection point is compared with its 8 adjacent points of the same scale and 18 points corresponding to the upper and lower adjacent scales, a total of 26 points, and finally ensure that in the scale space and the two-dimensional image space can detect extreme points.

2.3关键点定位2.3 Key point positioning

由于DoG值对噪声和边缘比较敏感，因此，在上面DoG尺度空间中检测到局部极值点还要经过进一步的检验才能精确定位为特征点。为了增强匹配的稳定性、提高抗噪声能力，通过拟合三维二次函数来精确确定关键点的位置和尺度，通过去除低对比度的关键点和不稳定的边缘响应点。Since the DoG value is sensitive to noise and edges, the local extremum points detected in the above DoG scale space need further inspection before they can be accurately positioned as feature points. In order to enhance the stability of matching and improve the anti-noise ability, the position and scale of key points are accurately determined by fitting a three-dimensional quadratic function, and low-contrast key points and unstable edge response points are removed.

通过求解每个极值点的梯度来为关键点赋予方向，采用梯度直方图统计法来确定关键点的方向，其梯度值m(x,y)和方向θ(x,y)的计算表达式为By solving the gradient of each extreme point to assign direction to the key point, using the gradient histogram statistical method to determine the direction of the key point, the calculation expression of the gradient value m(x,y) and direction θ(x,y) for

2.4特征描述子生成2.4 Feature Descriptor Generation

特征点描述通过对关键点周围区域进行图像分块，计算各块内的梯度直方图，生成独特性的向量描述符。为了增强匹配的稳健性，对每个关键点使用4×4共16个种子点来描述，这样对于一个关键点就可以产生128个数据，最终形成128维的SIFT特征向量。The feature point description divides the image into blocks around the key point, calculates the gradient histogram in each block, and generates a unique vector descriptor. In order to enhance the robustness of matching, a total of 16 seed points of 4×4 are used to describe each key point, so that 128 data can be generated for one key point, and finally a 128-dimensional SIFT feature vector is formed.

2.5特征点匹配2.5 Feature point matching

当两幅图像的SIFT特征向量生成后，下一步我们采用关键点特征向量的欧式距离来作为两幅图像中关键点的相似性判定度量。取样本特征点和样本特征点欧氏距离最近的特征点，在这两个特征点中，如果最近的距离除以次近的距离少于某个比例阈值，则接受这一对匹配点。降低这个比例阈值，SIFT匹配点数目会减少，但更加稳定。在计算特征点之间的欧氏距离时，采用了BBF算法来处理128维的特征向量。After the SIFT feature vectors of the two images are generated, in the next step we use the Euclidean distance of the key point feature vectors as the similarity judgment measure of the key points in the two images. Take the sample feature point and the feature point with the closest Euclidean distance to the sample feature point. Among the two feature points, if the nearest distance divided by the second closest distance is less than a certain ratio threshold, the pair of matching points is accepted. Lowering this ratio threshold will reduce the number of SIFT matching points, but it will be more stable. When calculating the Euclidean distance between the feature points, the BBF algorithm is used to process the 128-dimensional feature vector.

在对所有特征点进行粗匹配之后，使用RANSAC算法估计两个图像对之间的单位变换矩阵并将其作为几何约束，进而去除一些误匹配点，完成图像之间的精确匹配，提高匹配效率。After rough matching of all feature points, the RANSAC algorithm is used to estimate the unit transformation matrix between the two image pairs and use it as a geometric constraint, and then remove some mismatching points, complete the precise matching between images, and improve the matching efficiency.

Claims

1. the scale invariant feature matching algorithm based on the improvement of wavelet transformation, it is characterized in that carrying out two-dimentional Mallat wavelet transformation to two width images subject to registration decomposes, obtain the low-frequency component of image and level and vertical high frequency composition, give up the radio-frequency component after wavelet transformation decomposes, the low-frequency component of image is reconstructed, obtains new image; Utilize two width new images tectonic scale spaces (DoG), gaussian pyramid is generated by the difference of Gaussian convolution kernel of image and different scale, because new images is through wavelet transformation, give up partial information, when building difference of Gaussian pyramid, reduce down-sampled number of times, one deck fewer than the original generation number of plies, reduce the time in tectonic scale space; Find key point in DoG space, the check point of centre and it with 8 consecutive point of yardstick and neighbouring yardstick corresponding 18 totally 26 points compare, guarantee finally can obtain abundant key point.After the SIFT feature vector of two width images generates, next step adopts the Euclidean distance of key point proper vector to be used as the similarity determination tolerance of key point in two width images.Sampling eigen point and the nearest unique point of sample characteristics point Euclidean distance, in these two unique points, if nearest distance is less than certain proportion threshold value except distance near in proper order, then accept this pair match point.Reduce this proportion threshold value, SIFT match point number can reduce, but more stable.When calculating the Euclidean distance between unique point, have employed BBF algorithm to process the proper vector of 128 dimensions.After all unique points are slightly mated, use RANSAC algorithm estimate two images between identity transformation matrix and it can be used as geometrical constraint, and then remove some Mismatching points, complete the exact matching between image, improve matching efficiency.

2. the scale invariant feature matching algorithm of the improvement based on wavelet transformation according to claim 1, it is characterized in that: adopt two-dimentional Mallat wavelet transformation to carry out decomposition and reconstruction to image, give up the radio-frequency component comprising much noise and a small amount of useful information of image, only the low-frequency component of image is reconstructed.When tectonic scale space, because the pyramidal last one deck of difference of Gaussian contains little point of interest, too much influence is not had to final matching result, and the information comprised after image wavelet transform process also reduces a part, so down-sampled number of times can be reduced, the pyramidal number of plies of difference of Gaussian is made to reduce one deck.Use after all unique points of image being carried out to thick coupling RANSAC algorithm estimate two images between identity transformation matrix and it can be used as geometrical constraint, and then remove some Mismatching points, complete the exact matching between image.

3. the scale invariant feature matching algorithm of the improvement based on wavelet transformation according to claim 1 and 2, it is characterized in that: described two-dimentional Mallat wavelet transformation decomposes image, give up the radio-frequency component of image, only the low-frequency component of image is reconstructed.

4. the scale invariant feature matching algorithm of the improvement based on wavelet transformation according to claim 1 and 2, it is characterized in that: when tectonic scale space, decrease down-sampled number of times, make the pyramidal number of plies of difference of Gaussian reduce one deck, reduce the redundant computation of algorithm, improve real-time.

5. the scale invariant feature matching algorithm of the improvement based on wavelet transformation according to claim 1 and 2, is characterized in that: use RANSAC algorithm to remove part Mismatching point, improve accuracy and the accuracy of coupling.