WO2017049994A1 - 一种高光谱图像角点检测方法与系统 - Google Patents

一种高光谱图像角点检测方法与系统 Download PDF

Info

Publication number
WO2017049994A1
WO2017049994A1 PCT/CN2016/090015 CN2016090015W WO2017049994A1 WO 2017049994 A1 WO2017049994 A1 WO 2017049994A1 CN 2016090015 W CN2016090015 W CN 2016090015W WO 2017049994 A1 WO2017049994 A1 WO 2017049994A1
Authority
WO
WIPO (PCT)
Prior art keywords
point
corner
function
hyperspectral image
response
Prior art date
Application number
PCT/CN2016/090015
Other languages
English (en)
French (fr)
Inventor
李岩山
石伟
谢维信
张勇
Original Assignee
深圳大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 深圳大学 filed Critical 深圳大学
Publication of WO2017049994A1 publication Critical patent/WO2017049994A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/0002Inspection of images, e.g. flaw detection
    • G06T7/0004Industrial image inspection
    • G06T7/0008Industrial image inspection checking presence/absence
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/10Image acquisition modality
    • G06T2207/10032Satellite or aerial image; Remote sensing
    • G06T2207/10036Multispectral image; Hyperspectral image
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/30Subject of image; Context of image processing
    • G06T2207/30108Industrial image inspection

Definitions

  • the invention belongs to the field of corner detection, and in particular relates to a method and system for detecting corner points of a hyperspectral image.
  • hyperspectral images have been widely used in many fields, including remote sensing, machine vision, artificial intelligence, medicine, and astronomy. Different from grayscale images and color images, hyperspectral images are superimposed by images of multiple wavelength bands such as visible light, near-infrared, short-wave infrared, medium-wave infrared, and thermal infrared. Therefore, hyperspectral images can provide more information than grayscale images and color images.
  • Feature extraction is a common method of image analysis. However, the feature extraction methods of existing hyperspectral images are still immature, mainly to extract the global features of hyperspectral images, that is, all the targets in the whole image as a whole, but global Feature extraction cannot solve the more complex target recognition and segmentation problems.
  • the local features of an image are important features that distinguish one object from another, especially for machine vision cognition.
  • the corner point is one of the most common local features. In the general sense, the corner point refers to the point where the gray level changes sharply in the image or the intersection point of the contour boundary in the image.
  • the corner points reflect the key information in the image and play an important role in the understanding and analysis of the image. Especially the Harris corner, because of the invariance of rotation and grayscale changes, and the principle is simple, so the application is very extensive.
  • Moravec corner detection has many limitations, such as no rotation invariance.
  • Harris replaced the movement of the luminance block direction with a differential operator based on Moravec, and constructed a second-order moment matrix with structural information. Calculate the response value of each point by the determinant and straightness of the matrix, given The local maximum point in a neighborhood is the corresponding corner point.
  • the Harris corner detection algorithm has invariance to rotation and grayscale changes, and the detection rate is higher than the Moravec corner, so it is widely used.
  • Mikolajczyk and Schmid proposed the Harris-Laplacian detection operator and the Harris-Affine detection operator by means of the scale concept.
  • the Harris-Laplacian detection operator combines the Harris corner detection operator with the Gaussian scale space, and uses the idea of iterative estimation of the affine invariant neighborhood proposed by Lindeberg to increase the scale invariance of the corner feature.
  • the Harris-Affine detection operator can automatically detect image features under affine transformation and has affine invariance.
  • SZShen et al. proposed a fast adaptive Harris corner detection algorithm that overcomes the feature point omission caused by the uncertainty of the threshold value of the corner response or by setting the adaptive corner response threshold. Extract the wrong question.
  • B-spline function instead of Gaussian function to smooth the image and improve the accuracy of corner extraction.
  • Ivan Laptev et al. proposed a space-time Harris corner detection algorithm for video images. Considering the time domain, the expansion from the angular point of the airspace to the corner of the space-time domain was realized.
  • the existing Harris corner detection methods are all for grayscale images or video images, and there is no Harris corner detection method for hyperspectral images; that is, the existing hyperspectral image analysis methods are still insufficient and cannot be very good.
  • the key information for obtaining hyperspectral images is not limited to grayscale images or video images, and there is no Harris corner detection method for hyperspectral images; that is, the existing hyperspectral image analysis methods are still insufficient and cannot be very good.
  • the technical problem to be solved by the present invention is to provide a method and system for detecting hyperspectral image corner points, which aims to solve the problem that the current Harris corner detection method can only be used for grayscale images or video images and cannot be used for hyperspectral images. Improves the recognition of hyperspectral image patterns.
  • the present invention is achieved by a method for detecting a hyperspectral image corner point, the method comprising the steps of:
  • Step A a structure on the hyper-spectral image point p 0 f (x, y, z) of the point and its neighborhood weighting of the correlation function p 1;
  • Step B constructing a corner response function according to the weighted correlation function
  • Step C calculating, according to the corner response function, a Harris corner response value of a point p 0 in the hyperspectral image f(x, y, z) and a Harris corner point response value of all points in the neighborhood;
  • Step D if the Harris corner response value of a point p 0 in the hyperspectral image f(x, y, z) is greater than the Harris corner point response value of all points in the neighborhood, the point p 0 is The Harris corner of the hyperspectral image f(x, y, z).
  • weighted correlation function is:
  • the point p 0 is a pixel in the hyperspectral image f(x, y, z) whose coordinates are (x, y, z), and f(x, y, z) is a pixel value corresponding to the point p 0 ;
  • Point p 1 coordinates are (x + ⁇ x, y + ⁇ y, z + ⁇ z), f (x + ⁇ x, y + ⁇ y, z + ⁇ z) is the pixel value corresponding to point p 1 ;
  • the window function ⁇ (x, y, z) uses a Gaussian weighting function as follows:
  • is the scale factor of the Gaussian function
  • l is the length of the window function moving in the x direction
  • m is the length of the window function moving in the y direction
  • r is the length of the window function moving in the z direction
  • f x , f y , f z represent the gradient of the image f(x, y, z) in the three directions x, y, z, respectively
  • represents the Gaussian weighting function ⁇ (x, y, z)
  • A, B, C, D, E, and F correspond to the respective elements of the matrix M.
  • corner response function is:
  • ⁇ 1 , ⁇ 2 , and ⁇ 3 are the eigenvalues of the matrix M, respectively.
  • f x , f y , f z are calculated by the following formula:
  • R x , R y , and R z are gradient operators in three directions of horizontal, vertical, and spectral, respectively, and the size is 5*5*5. Convolution symbol;
  • the horizontal gradient operator R x (:,:, 3) expression is as follows, the other positions of R x are 0:
  • the vertical gradient operator R y (:,:,3) has the following expression, and all other positions of R y are 0:
  • the gradient operator R z (3,3,:) expression in the spectral direction is as follows, and the other positions of R z are all 0:
  • f xy f x ⁇ f y
  • f yz f y ⁇ f z
  • f xz f x ⁇ f z ;
  • ⁇ (x, y, z) is a Gaussian weighting function
  • is the scale factor of the Gaussian function
  • 1.
  • the invention also provides a hyperspectral image corner point detection system, the system comprising: a weighted correlation function construction module, a corner point response function construction module, a corner point response value calculation module, and a corner point judgment module; wherein:
  • Weighted correlation function constructing module for constructing a weighted correlation function on the point p 1 p 0 of a point and its neighbors on hyperspectral image f (x, y, z) ;
  • a corner response function constructing module is configured to construct a corner point response function according to the weighted correlation function
  • the corner response value calculation module is configured to calculate a Harris corner response value of a point p 0 in the hyperspectral image f(x, y, z) according to the corner point response function and Harris of all points in the neighborhood thereof Corner response value;
  • the corner point judging module is configured to determine whether a point p 0 is a Harris corner of the hyperspectral image f(x, y, z), if a point p in the hyperspectral image f(x, y, z) The Harris corner response of 0 is greater than the Harris corner response of all points in its neighborhood, then the point p 0 is the Harris corner.
  • weighted correlation function is:
  • the point p 0 is a pixel in the hyperspectral image f(x, y, z) whose coordinates are (x, y, z), and f(x, y, z) is a pixel value corresponding to the point p 0 ;
  • Point p 1 coordinates are (x + ⁇ x, y + ⁇ y, z + ⁇ z), f (x + ⁇ x, y + ⁇ y, z + ⁇ z) is the pixel value corresponding to point p 1 ;
  • the window function ⁇ (x, y, z) uses a Gaussian weighting function as follows:
  • is the scale factor of the Gaussian function
  • l is the length of the window function moving in the x direction
  • m is the length of the window function moving in the y direction
  • r is the length of the window function moving in the z direction
  • f x , f y , f z represent the gradient of the image f(x, y, z) in the three directions x, y, z, respectively
  • represents the Gaussian weighting function ⁇ (x, y, z)
  • A, B, C, D, E, and F correspond to the respective elements of the matrix M.
  • corner response function is:
  • ⁇ 1 , ⁇ 2 , and ⁇ 3 are the eigenvalues of the matrix M, respectively.
  • f x , f y , f z are calculated by the following formula:
  • R x , R y , and R z are gradient operators in three directions of horizontal, vertical, and spectral, respectively, and the size is 5*5*5. Convolution symbol;
  • the horizontal gradient operator R x (:,:, 3) expression is as follows, the other positions of R x are 0:
  • the vertical gradient operator R y (:,:,3) has the following expression, and all other positions of R y are 0:
  • the gradient operator R z (3,3,:) expression in the spectral direction is as follows, and the other positions of R z are all 0:
  • f xy f x ⁇ f y
  • f yz f y ⁇ f z
  • f xz f x ⁇ f z ;
  • ⁇ (x, y, z) is a Gaussian weighting function
  • is the scale factor of the Gaussian function
  • 1.
  • the present invention has the beneficial effects that the present invention provides a hyperspectral image corner point detection method and system, and calculates a corner of a point and its neighbor in a hyperspectral image by defining a Harris corner response function. Point response value and compare the size of the response value to determine whether it is a corner point; using this detection method can obtain the key information of the hyperspectral image, which can better analyze the hyperspectral image and improve the pattern recognition effect of the hyperspectral image.
  • FIG. 1 is a flowchart of a method for detecting a Harris corner point of a hyperspectral image according to an embodiment of the present invention
  • FIG. 2 is a schematic diagram showing a representation of a hyperspectral image provided by the prior art
  • FIG. 3 is a schematic diagram of a high-spectral image Harris corner detection system according to an embodiment of the present invention.
  • the point When performing hyperspectral Harris corner detection, it can be detected whether the point is a Harris corner by determining the correlation between a point in the hyperspectral image and its neighborhood; specifically, if a point has a small correlation with its neighborhood, That is, in the neighborhood, the gray level of the image changes greatly, then the point is the Harris corner; otherwise, it is not the corner.
  • the degree of correlation between a point and its neighbor is not well measured. Therefore, the present invention derives from the function of correlation, and then derives and deforms to finally obtain a corner response function, thereby converting to using a corner response. Function to determine if a point is a Harris corner.
  • the hyperspectral image corner detection method is specifically described below, as shown in FIG. 1, the method includes the following step:
  • Step A a structure on the hyper-spectral image point p 0 f (x, y, z) of the point and its neighborhood weighting of the correlation function p 1;
  • the hyperspectral image F whose spatial domain size is M*N and the number of spectral bands is n, can be expressed as:
  • f(x, y, z) represents the function of the hyperspectral image
  • (x, y, z) represents the 3-dimensional coordinates
  • x and y represent the spatial domain coordinates, 0 ⁇ x ⁇ M, 0 ⁇ y ⁇ N
  • z Indicates the spectral domain coordinates, 0 ⁇ z ⁇ n.
  • the xOy plane and its parallel plane are two-dimensional images, which represent a certain band of the hyperspectral image, that is, the airspace, and the z-axis represents each band, that is, the spectral domain.
  • the hyperspectral image is different from the two-dimensional gray image, and the hyperspectral image is three-dimensional, which is one-dimensional spectral information based on the two-dimensional image. Therefore, the method of extracting Harris corner points from gray image cannot be directly applied to hyperspectral images, and the relationship between spectra needs to be measured.
  • p 0 be a pixel in the hyperspectral image f(x, y, z) whose coordinates are (x, y, z), and point p 1 is a point on the neighborhood of p 0 whose coordinates are (x) + ⁇ x, y+ ⁇ y, z+ ⁇ z), then the p 0 and p 1 correlation functions are defined as follows:
  • f(x, y, z) is a pixel value corresponding to the point p 0
  • f(x + ⁇ x, y + ⁇ y, z + ⁇ z) is a pixel value corresponding to the point p 1 .
  • the hyperspectral image of different bands has correlation between the pixels at the same position, For inter-spectral correlation.
  • the feature points of the hyperspectral image proposed by the invention are mainly used in the application fields of hyperspectral image pattern recognition, image registration, etc., therefore, the present invention requires that the extracted feature points need to be distinct in the spatial and spectral domains.
  • the present invention uses a window function to convolve with a hyperspectral image to determine the correlation between a hyperspectral pixel point and its neighborhood. Therefore, based on equation (2), the weighted correlation function is defined as follows;
  • p 0 be a pixel in the hyperspectral image f(x, y, z) whose coordinates are (x, y, z), and point p 1 is a point on the neighborhood of p 0 whose coordinates are (x) + ⁇ x, y+ ⁇ y, z+ ⁇ z), then the p 0 and p 1 weighted correlation functions are defined as follows
  • f(x, y, z) is a pixel value corresponding to point p 0
  • f(x+ ⁇ x, y+ ⁇ y, z+ ⁇ z) is a pixel value corresponding to point p 1 ;
  • the window function ⁇ (x, y, z) uses a Gaussian weighting function as follows:
  • is the scale factor of the Gaussian function.
  • Equation (3) For the convolution operation, l is the length of the window moving in the x direction, m is the length of the window moving in the y direction, and r is the length of the window moving in the z direction. In the embodiment of the present invention, the values of l, m, and r are taken. 1, the window size is 3*3*3.
  • Step B constructing a corner point response function according to the weighted correlation function; the specific implementation process is as follows.
  • f(u+ ⁇ x, v+ ⁇ y, p+ ⁇ z) is obtained by translating the image f(u, v, p) ( ⁇ x, ⁇ y, ⁇ z); for f(u+ ⁇ x, v + ⁇ y,p+ ⁇ z) performs Taylor series expansion and takes a first-order approximation as follows:
  • f x , f y , f z are the gradients of the various points of the image f(x, y, z) in the three directions x, y, z, ie
  • equation (6) can be approximated as:
  • f x 2 , f y 2 , f z 2 represent the squares of the gradients f x , f y , f z of the hyperspectral image in the x, y, and z directions, respectively, and f xy represents the f x and f y Product, f yz represents the product of f y and f z , f xz represents the product of f x and f z , and ⁇ is the Gaussian weighting function ⁇ (x, y, z) in equation (4),
  • A, B, C, D, E, and F correspond to the respective elements of the matrix M.
  • Equation (10) M is a real symmetric second-order moment matrix, reflecting the structural information of the image.
  • the magnitudes of the three eigenvalues ⁇ 1 , ⁇ 2 , ⁇ 3 of the second-order moment matrix M are in the image.
  • ⁇ 1 , ⁇ 2 , and ⁇ 3 are both large and approximately equal, the weighted correlation function of equation (6) is large in all directions, and the point (x, y, z) is the angle in the hyperspectral image. point.
  • Step C calculating, according to the corner response function, a Harris corner response value of a point p 0 in the hyperspectral image f(x, y, z) and a Harris corner point response value of all points in the neighborhood;
  • Step D if the Harris corner response value of a point p 0 in the hyperspectral image f(x, y, z) is greater than the Harris corner point response value of all points in the neighborhood, the point p 0 is The Harris corner of the hyperspectral image f(x, y, z).
  • Step 1 Calculate the gradients f x , f y , f z of the points of the hyperspectral image f(x, y, z) in the three directions of x, y, and z according to the equation (8).
  • f x , f y , f z can be obtained by convolving the hyperspectral image f(x, y, z) with a gradient operator, namely:
  • f(x, y, z) is a hyperspectral image
  • R x , R y , and R z are gradient operators in three directions of horizontal, vertical, and spectral, respectively, and the size is 5*5*5. Is a convolution symbol.
  • the present invention designs the horizontal gradient operator R x (:, :, 3) as follows, and the other positions of R x are all 0:
  • the vertical gradient operator R y (:,:,3) is designed as follows, and all other positions of R y are 0:
  • the gradient operator R z (3,3,:) in the spectral direction is designed as follows, and the other positions of R z are all 0:
  • step 2 the product of f x , f y , f z is calculated according to the following formula, and a gradient matrix N is constructed.
  • f x , f y , f z represent the gradient of the hyperspectral image in the x, y, and z directions
  • f xy represents the product of f x and f y
  • f yz represents the product of f y and f z
  • f xz Represents the product of f x and f z .
  • the gradient matrix N is constructed from the product of the individual gradients calculated in equation (18), namely:
  • Step 3 in order to construct the second-order moment matrix M, it is known from equation (10) that Gaussian weighting is required for each element in the gradient matrix N of equation (19) using a Gaussian weighting function ⁇ (x, y, z). Then, the Gaussian weighting function ⁇ (x, y, z) can be directly convoluted with the gradient matrix N, and the formula is as follows:
  • Step 4 calculating each pixel f(x 0 , y 0 , z 0 ) ⁇ f(x, y, z) according to the value of each parameter in the matrix M calculated by the equation (11) and combining the equation (20)
  • Step 5 the comparison point (x 0, y 0, z 0) Harris corner response values and other points in its neighborhood 3 * 3 * 3, if the point (x 0, y 0, z 0) Harris corner The point response value is greater than the corner response value of all other points in the neighborhood of its 3*3*3, then the point (x 0 , y 0 , z 0 ) is the corner point of the hyperspectral image.
  • the center point of the window has a local maximum Harris corner response value, and if so, returns the center point.
  • the position is the position of the corner point.
  • the present invention also provides a hyperspectral image corner detection system.
  • the system includes: a weighted correlation function construction module, a corner response function construction module 2, a corner response value calculation module 3, Corner point determination module 4; wherein:
  • Weighting a correlation function construction module configured for weighting the correlation function with respect to a point p 0 on the hyper-spectral image f (x, y, z) of the point and its neighborhood of p 1;
  • the corner response function constructing module 2 is configured to construct a corner point response function according to the weighted correlation function
  • the corner response value calculation module 3 is configured to calculate a Harris corner response value of a point p 0 in the hyperspectral image f(x, y, z) according to the corner point response function and all points on the neighborhood thereof Harris corner response value;
  • the corner point judging module 4 is configured to determine whether a point p 0 is a Harris corner of the hyperspectral image f(x, y, z), if a point in the hyperspectral image f(x, y, z) The Harris corner response of p 0 is greater than the Harris corner response of all points in its neighborhood, then the point p 0 is the Harris corner.
  • the invention provides a method and system for detecting hyperspectral image corner points, which solves the problem that the current Harris corner detection method can only be used for gray image or video image and cannot be used for hyperspectral image;
  • the key information of the hyperspectral image can be obtained to better analyze the hyperspectral image and improve the pattern recognition effect of the hyperspectral image.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Quality & Reliability (AREA)
  • Image Analysis (AREA)
  • Magnetic Resonance Imaging Apparatus (AREA)
  • Medicines Containing Antibodies Or Antigens For Use As Internal Diagnostic Agents (AREA)

Abstract

一种高光谱图像角点检测方法与系统,所述方法包括下述步骤:步骤A,构造关于高光谱图像f(x,y,z)中的某点p 0和其邻域上的点p 1的加权相关性函数(A);步骤B,根据所述加权相关性函数构造一个角点响应函数(B);步骤C,根据所述角点响应函数计算所述高光谱图像f(x,y,z)中的某点p 0的Harris角点响应值和其邻域上所有点的Harris角点响应值(C);步骤D,若所述高光谱图像f(x,y,z)中的某点p 0的Harris角点响应值大于其邻域上所有点的Harris角点响应值,则该点p 0即为所述高光谱图像f(x,y,z)的Harris角点(D)。该检测方法能获取高光谱图像的关键信息从而能更好地对高光谱图像进行分析,提高了高光谱图像模式识别效果。

Description

一种高光谱图像角点检测方法与系统 技术领域
本发明属于角点检测领域,尤其涉及一种高光谱图像角点检测方法与系统。
背景技术
近年来高光谱图像在很多领域都有着广泛的应用,包括遥感、机器视觉、人工智能、医学、天文学等。与灰度图像和彩色图像不同的是,高光谱图像是由可见光、近红外、短波红外、中波红外、热红外等多个波段的图像叠加而成。因此,高光谱图像能够比灰度图像和彩色图像提供更多的信息。特征提取是一种常见的图像分析的方法,然而现有的高光谱图像的特征提取方法还不成熟,主要是提取高光谱图像的全局特征,即将整个图像中的所有目标作为一个整体,但是全局特征提取不能解决较为复杂的目标识别和分割问题。
图像的局部特征是能够区别一个物体与另一个物体的重要特征,特别是对于机器视觉认知来说,这是十分重要的。而角点则是十分常见的局部特征之一,通常意义上,角点是指图像中灰度变化剧烈的点或者图像中轮廓边界的相交点。角点反映了图像中的关键信息,对于图像的理解和分析有很重要的作用。尤其是Harris角点,因为对旋转和灰度变化具有不变性,而且原理简单,所以应用十分广泛。
关于图像局部特征的研究可以追溯到20世纪70年代。1977年Moravec提出了角点特征,通过灰度自相关函数来考虑一个像素和其邻域像素的相似性,但Moravec角点检测有很多局限性,如不具备旋转不变性等等。1988年Harris在Moravec的基础上用微分算子代替了亮度块方向的移动,构建了带有结构信息的二阶矩矩阵。通过该矩阵的行列式及直迹来计算各个点的响应值,在给定 的一邻域中局部最大值点,即为对应的角点。Harris角点检测算法具有对旋转和灰度变化的不变性,而且检测率高于Moravec角点,所以被广泛应用。而Mikolajczyk和Schmid借助尺度的概念,提出了Harris-Laplacian检测算子和Harris-Affine检测算子。Harris-Laplacian检测算子将Harris角点检测算子与高斯尺度空间相结合,利用Lindeberg提出的通过迭代估计仿射不变性邻域的思想,给角点特征增加了尺度不变性。Harris-Affine检测算子能自动检测仿射变换下的图像特征,具有仿射不变性。2011年,S.Z.Shen等人提出了一种快速自适应的Harris角点检测算法,通过设置自适应的角点响应阈值,克服了之前由于角点响应的阈值的不确定而造成的特征点遗漏或者提取错误的问题。之后Y.H.Yang等人在Harris角点原有算法的基础上,提出了用B-样条函数代替高斯函数对图像进行平滑,提高了角点提取的准确性。
2003年,Ivan Laptev等人提出了视频图像的时空Harris角点检测算法,考虑了时域的情况,实现了从空域角点到时空域角点的扩展。
然而,现有的Harris角点检测方法都是针对灰度图像或视频图像的,没有用于高光谱图像的Harris角点检测方法;即现有的高光谱图像分析方法还存在不足,不能很好的获取高光谱图像的关键信息。
发明内容
本发明所要解决的技术问题在于提供一种高光谱图像角点检测方法与系统,旨在解决目前的Harris角点检测方法只能用于灰度图像或视频图像而不能用于高光谱图像的问题,提高了高光谱图像模式识别效果。
本发明是这样实现的,一种高光谱图像角点检测方法,所述方法包括下述步骤:
步骤A,构造关于高光谱图像f(x,y,z)中的某点p0和其邻域上的点p1的加权相关性函数;
步骤B,根据所述加权相关性函数构造一个角点响应函数;
步骤C,根据所述角点响应函数计算所述高光谱图像f(x,y,z)中的某点p0的Harris角点响应值和其邻域上所有点的Harris角点响应值;
步骤D,若所述高光谱图像f(x,y,z)中的某点p0的Harris角点响应值大于其邻域上所有点的Harris角点响应值,则该点p0即为所述高光谱图像f(x,y,z)的Harris角点。
进一步地,所述加权相关性函数为:
Figure PCTCN2016090015-appb-000001
其中,点p0是高光谱图像f(x,y,z)中的一个像素,其坐标为(x,y,z),f(x,y,z)为点p0对应的像素值;点p1坐标为(x+Δx,y+Δy,z+Δz),f(x+Δx,y+Δy,z+Δz)为点p1对应的像素值;
窗函数ω(x,y,z)采用高斯加权函数,如下所示:
Figure PCTCN2016090015-appb-000002
而由上式可得:
Figure PCTCN2016090015-appb-000003
其中,σ为高斯函数的尺度因子;
Figure PCTCN2016090015-appb-000004
为卷积运算符号,l为窗函数沿x方向移动的长度,m为窗函数沿y方向移动的长度,r为窗函数沿z方向移动的长度,l=1,m=1,r=1。
进一步地,所述加权相关性函数中
Figure PCTCN2016090015-appb-000005
表示为
Figure PCTCN2016090015-appb-000006
即:
Figure PCTCN2016090015-appb-000007
而,
Figure PCTCN2016090015-appb-000008
则,
Figure PCTCN2016090015-appb-000009
其中,
Figure PCTCN2016090015-appb-000010
式中,fx,fy,fz分别表示图像f(x,y,z)在x,y,z三个方向上的梯度,即
Figure PCTCN2016090015-appb-000011
Figure PCTCN2016090015-appb-000012
Figure PCTCN2016090015-appb-000013
上式中,ω表示高斯加权函数ω(x,y,z),
Figure PCTCN2016090015-appb-000014
为卷积符号,A、B、C、D、E、F分别对应矩阵M的各个元素。
进一步地,所述角点响应函数为:
R=det(M)-k(trace(M))3=(ABC+2DEF-BE2-AF2-CD2)-k(A+B+C)3
其中,k=0.04,k为经验常数;det(M)表示矩阵M的行列式,trace(M)表示矩阵M的迹,其表达式如下:
det(M)=λ1λ2λ3=ABC+2DEF-BE2-AF2-CD2
trace(M)=λ123=A+B+C
其中,λ1、λ2、λ3分别为矩阵M的特征值。
进一步地,所述fx,fy,fz由如下公式计算:
Figure PCTCN2016090015-appb-000015
Figure PCTCN2016090015-appb-000016
Figure PCTCN2016090015-appb-000017
式中,Rx、Ry、Rz分别为水平、垂直、光谱三个方向的梯度算子,且大小均为5*5*5,
Figure PCTCN2016090015-appb-000018
为卷积符号;
其中,水平梯度算子Rx(:,:,3)表达式如下,Rx的其他位置均为0:
Figure PCTCN2016090015-appb-000019
垂直梯度算子Ry(:,:,3)表达式如下,Ry的其他位置均为0:
Figure PCTCN2016090015-appb-000020
光谱方向的梯度算子Rz(3,3,:)表达式如下,Rz的其他位置均为0:
Figure PCTCN2016090015-appb-000021
进一步地,矩阵M由如下公式计算:
Figure PCTCN2016090015-appb-000022
其中,
Figure PCTCN2016090015-appb-000023
fxy=fx·fy,fyz=fy·fz,fxz=fx·fz
ω(x,y,z)为高斯加权函数,σ为高斯函数的尺度因子,σ=1。
本发明还提供了一种高光谱图像角点检测系统,所述系统包括:加权相关性函数构造模块、角点响应函数构造模块、角点响应值计算模块、角点判断模块;其中:
加权相关性函数构造模块用于构造关于高光谱图像f(x,y,z)中的某点p0和其邻域上的点p1的加权相关性函数;
角点响应函数构造模块用于根据所述加权相关性函数构造一个角点响应函数;
角点响应值计算模块用于根据所述角点响应函数计算所述高光谱图像f(x,y,z)中的某点p0的Harris角点响应值和其邻域上所有点的Harris角点响应值;
角点判断模块用于判断某点p0是否为所述高光谱图像f(x,y,z)的Harris角点,若所述高光谱图像f(x,y,z)中的某点p0的Harris角点响应值大于其邻域上所有点的Harris角点响应值,则该点p0即为Harris角点。
进一步地,所述加权相关性函数为:
Figure PCTCN2016090015-appb-000024
其中,点p0是高光谱图像f(x,y,z)中的一个像素,其坐标为(x,y,z),f(x,y,z)为点p0对应的像素值;点p1坐标为(x+Δx,y+Δy,z+Δz),f(x+Δx,y+Δy,z+Δz)为点p1对应的像素值;
窗函数ω(x,y,z)采用高斯加权函数,如下所示:
Figure PCTCN2016090015-appb-000025
而由上式可得:
Figure PCTCN2016090015-appb-000026
其中,σ为高斯函数的尺度因子;
Figure PCTCN2016090015-appb-000027
为卷积运算符号,l为窗函数沿x方向移动的长度,m为窗函数沿y方向移动的长度,r为窗函数沿z方向移动的长度,l=1,m=1,r=1。
进一步地,所述加权相关性函数中
Figure PCTCN2016090015-appb-000028
表示为
Figure PCTCN2016090015-appb-000029
即:
Figure PCTCN2016090015-appb-000030
而,
Figure PCTCN2016090015-appb-000031
则,
Figure PCTCN2016090015-appb-000032
其中,
Figure PCTCN2016090015-appb-000033
式中,fx,fy,fz分别表示图像f(x,y,z)在x,y,z三个方向上的梯度,即
Figure PCTCN2016090015-appb-000034
Figure PCTCN2016090015-appb-000035
Figure PCTCN2016090015-appb-000036
上式中,ω表示高斯加权函数ω(x,y,z),
Figure PCTCN2016090015-appb-000037
为卷积符号,A、B、C、D、E、F分别对应矩阵M的各个元素。
进一步地,所述角点响应函数为:
R=det(M)-k(trace(M))3=(ABC+2DEF-BE2-AF2-CD2)-k(A+B+C)3
其中,k=0.04,k为经验常数;det(M)表示矩阵M的行列式,trace(M)表示 矩阵M的迹,其表达式如下:
det(M)=λ1λ2λ3=ABC+2DEF-BE2-AF2-CD2
trace(M)=λ123=A+B+C
其中,λ1、λ2、λ3分别为矩阵M的特征值。
进一步地,所述fx,fy,fz由如下公式计算:
Figure PCTCN2016090015-appb-000038
Figure PCTCN2016090015-appb-000039
Figure PCTCN2016090015-appb-000040
式中,Rx、Ry、Rz分别为水平、垂直、光谱三个方向的梯度算子,且大小均为5*5*5,
Figure PCTCN2016090015-appb-000041
为卷积符号;
其中,水平梯度算子Rx(:,:,3)表达式如下,Rx的其他位置均为0:
Figure PCTCN2016090015-appb-000042
垂直梯度算子Ry(:,:,3)表达式如下,Ry的其他位置均为0:
Figure PCTCN2016090015-appb-000043
光谱方向的梯度算子Rz(3,3,:)表达式如下,Rz的其他位置均为0:
Figure PCTCN2016090015-appb-000044
进一步地,矩阵M由如下公式计算:
Figure PCTCN2016090015-appb-000045
其中,
Figure PCTCN2016090015-appb-000046
fxy=fx·fy,fyz=fy·fz,fxz=fx·fz
ω(x,y,z)为高斯加权函数,σ为高斯函数的尺度因子,σ=1。
本发明与现有技术相比,有益效果在于:本发明提供了一种高光谱图像角点检测方法与系统,通过定义Harris角点响应函数,计算高光谱图像中某点及其邻域的角点响应值并比较响应值的大小来判断是否为角点;采用这种检测方法能获取高光谱图像的关键信息从而能更好地对高光谱图像进行分析,提高了高光谱图像模式识别效果。
附图说明
图1是本发明实施例提供的高光谱图像Harris角点检测方法的流程图;
图2是现有技术提供的高光谱图像的表示示意图;
图3是本发明实施例提供的高光谱图像Harris角点检测系统的示意图。
具体实施方式
为了使本发明的目的、技术方案及优点更加清楚明白,以下结合附图及实施例,对本发明进行进一步详细说明。应当理解,此处所描述的具体实施例仅仅用以解释本发明,并不用于限定本发明。
在进行高光谱Harris角点检测时,可以通过判断高光谱图像中某点与其邻域的相关性来检测该点是否为Harris角点;具体地,如果某点与其邻域的相关性较小,即在其邻域范围内,图像的灰度发生了很大的变化,则该点为Harris角点;反之,则不为角点。但是,事实上,某点与其邻域的相关性程度不能很好地去衡量,所以,本发明从相关性的函数出发,推到、变形最终得到角点响应函数,从而转换为利用角点响应函数来判断某点是否为Harris角点。
下面具体介绍高光谱图像角点检测方法,如图1所示,所述方法包括下述 步骤:
步骤A,构造关于高光谱图像f(x,y,z)中的某点p0和其邻域上的点p1的加权相关性函数;
下面先介绍关于高光谱图像的相关性函数;
高光谱图像F,其空域尺寸为M*N,光谱波段数为n,则它可以表示为:
F=f(x,y,z)    (1)
式中,f(x,y,z)表示高光谱图像的函数,(x,y,z)表示3维坐标,x和y表示空域坐标,0<x<M,0<y<N,z表示光谱域坐标,0<z<n。如图2所示,xOy平面及其平行面为二维图像,表示高光谱图像的某个波段,即空域,z轴表示各个波段,即光谱域。
显然,高光谱图像与二维灰度图像不同,高光谱图像是三维的,是在二维图像的基础上多了一维光谱信息。所以不能直接将灰度图像提取Harris角点的方法应用到高光谱图像上,还需要衡量光谱间的关系。
根据二维图像相关性的定义及高光谱图像的特性,我们首先给出高光谱图像的相关性函数。
设点p0是高光谱图像f(x,y,z)中的一个像素,其坐标为(x,y,z),点p1为p0的邻域上的点,其坐标为(x+Δx,y+Δy,z+Δz),则p0和p1相关性函数定义如下:
c(Δx,Δy,Δz)=[f(x,y,z)-f(x+Δx,y+Δy,z+Δz)]2    (2)
式中,f(x,y,z)为点p0对应的像素值,f(x+Δx,y+Δy,z+Δz)为点p1对应的像素值。
高光谱图像的某一波段,存在空间相关性,即像素间相关性,这也是在灰度图像的Harris角点检测算法中所用到的。同时,由于不同波段的高光谱图像空间结构基本相同,且其灰度值是同一成像物在各个波段的反射强度值,因此不同波段的高光谱图像在相同位置上的像素间存在相关性,称为谱间相关性。
下面介绍关于高光谱图像的加权相关性函数;
本发明所提的高光谱图像特征点主要用于高光谱图像模式识别、图像配准等应用领域中,因此本发明要求所提取的特征点需要空域和光谱域上具有明显 的变化。本发明采用窗函数与高光谱图像进行卷积来判断高光谱像素点与其邻域的相关性,因此,在式(2)的基础上,对加权相关性函数进行如下定义;
设点p0是高光谱图像f(x,y,z)中的一个像素,其坐标为(x,y,z),点p1为p0的邻域上的点,其坐标为(x+Δx,y+Δy,z+Δz),则p0和p1加权相关性函数定义如下
Figure PCTCN2016090015-appb-000047
其中,f(x,y,z)为点p0对应的像素值,f(x+Δx,y+Δy,z+Δz)为点p1对应的像素值;
窗函数ω(x,y,z)采用高斯加权函数,如下所示:
Figure PCTCN2016090015-appb-000048
而由式(4)可得:
Figure PCTCN2016090015-appb-000049
式中,σ为高斯函数的尺度因子。
式(3)中,
Figure PCTCN2016090015-appb-000050
为卷积运算,l为窗口沿x方向移动的长度,m为窗口沿y方向移动的长度,r为窗口沿z方向移动的长度,本发明实施例中,l、m和r的值均取1,即窗口大小为3*3*3。
步骤B,根据所述加权相关性函数构造一个角点响应函数;具体实现过程如下,
将公式(3)中的
Figure PCTCN2016090015-appb-000051
表示为
Figure PCTCN2016090015-appb-000052
即:
Figure PCTCN2016090015-appb-000053
式(6)中,f(u+Δx,v+Δy,p+Δz)为将图像f(u,v,p)平移(Δx,Δy,Δz)后得到;对f(u+Δx,v+Δy,p+Δz)进行泰勒级数展开,并取一阶近似,具体如下:
Figure PCTCN2016090015-appb-000054
式中,fx,fy,fz是图像f(x,y,z)的各个点在x、y、z三个方向上的梯度,即
Figure PCTCN2016090015-appb-000055
因此,式(6)可近似为:
Figure PCTCN2016090015-appb-000056
其中,
Figure PCTCN2016090015-appb-000057
式中,fx 2,fy 2,fz 2分别表示高光谱图像在x、y、z三个方向的梯度fx,fy,fz的平方,fxy表示fx与fy的乘积,fyz表示fy与fz的乘积,fxz表示fx与fz的乘积,ω为式(4)中的高斯加权函数ω(x,y,z),
Figure PCTCN2016090015-appb-000058
为卷积符号,A、B、C、D、E、F分别对应矩阵M的各个元素。
而角点响应函数的公式为:
R=det(M)-k(trace(M))3=(ABC+2DEF-BE2-AF2-CD2)-k(A+B+C)3    (11)
式中,k为经验常数,这里取值0.04;det(M)为矩阵M的行列式,trace(M)为矩阵M的迹,其表达式如下:
det(M)=λ1λ2λ3=ABC+2DEF-BE2-AF2-CD2    (12)
trace(M)=λ123=A+B+C    (13)
由式(10)可知,M为实对称的二阶矩矩阵,反映了图像的结构信息,事实上,二阶矩矩阵M的三个特征值λ1、λ2、λ3的大小与图像中的角点、边缘和平面存在着对应关系。如果λ1、λ2、λ3均较大,且近似相等,则式(6)的加权相关函数在所有方向均较大,则点(x,y,z)即为高光谱图像中的角点。但是,要判断三个特征值是否均很大,而且还要求矩阵特征值,显然比较繁琐,所以在实际判断某点是否为高光谱图像中的角点时,一般不采用这种方式判断。
步骤C,根据所述角点响应函数计算所述高光谱图像f(x,y,z)中的某点p0的Harris角点响应值和其邻域上所有点的Harris角点响应值;
步骤D,若所述高光谱图像f(x,y,z)中的某点p0的Harris角点响应值大于其邻域上所有点的Harris角点响应值,则该点p0即为所述高光谱图像f(x,y,z)的Harris角点。
具体地,下面介绍关于求Harris角点响应值进而实现角点检测的算法:
步骤1,根据式(8),计算高光谱图像f(x,y,z)各个点在x、y、z三个方向的梯度fx,fy,fz。在实际计算中,fx,fy,fz可以通过将高光谱图像f(x,y,z)与梯度算子卷积而得,即:
Figure PCTCN2016090015-appb-000059
式中,f(x,y,z)为高光谱图像,Rx、Ry、Rz分别为水平、垂直、光谱三个方向的梯度算子,且大小均为5*5*5,
Figure PCTCN2016090015-appb-000060
为卷积符号。
本发明对水平梯度算子Rx(:,:,3)设计如下,Rx的其他位置均为0:
Figure PCTCN2016090015-appb-000061
垂直梯度算子Ry(:,:,3)设计如下,Ry的其他位置均为0:
Figure PCTCN2016090015-appb-000062
光谱方向的梯度算子Rz(3,3,:)设计如下,Rz的其他位置均为0:
Figure PCTCN2016090015-appb-000063
步骤2,根据下述式子,计算fx,fy,fz两两间的乘积,并构建梯度矩阵N。
fx 2=fx·fx,fy 2=fy·fy,fz 2=fz·fz,fxy=fx·fy,fyz=fy·fz,fxz=fx·fz    (18)
其中,fx,fy,fz表示高光谱图像在x、y、z三个方向的梯度,fxy表示fx与fy的乘积,fyz表示fy与fz的乘积,fxz表示fx与fz的乘积。
由式(18)中计算的各个梯度的乘积构建梯度矩阵N,即:
Figure PCTCN2016090015-appb-000064
步骤3,为了构建二阶矩矩阵M,由式(10)可知,需要使用高斯加权函数ω(x,y,z)对式(19)梯度矩阵N中的各个元素进行高斯加权。则可以直接将高斯加权函数ω(x,y,z)与梯度矩阵N卷积求得,公式如下:
Figure PCTCN2016090015-appb-000065
式中,ω(x,y,z)为式(4)中的高斯加权函数,其中,σ为高斯函数的尺度因子,这里取σ=1。
步骤4,根据式(11)并结合式(20)计算出的矩阵M中的各参数的值计算每个像元f(x0,y0,z0)∈f(x,y,z)的Harris角点响应值R(x0,y0,z0)。
步骤5,比较点(x0,y0,z0)和其3*3*3的邻域中其他点的Harris角点响应值,如果点(x0,y0,z0)的Harris角点响应值大于其3*3*3的邻域中其余所有点的角点 响应值,则点(x0,y0,z0)即为高光谱图像的角点。在实验中,通过沿着x、y、z三个方向移动一个3*3*3的窗口,判断窗口的中心点是否具有局部最大的Harris角点响应值,如果有,则返回该中心点的位置,即为角点的位置。
本发明还提供了一种高光谱图像角点检测系统,如图3所示,所述系统包括:加权相关性函数构造模块1、角点响应函数构造模块2、角点响应值计算模块3、角点判断模块4;其中:
加权相关性函数构造模块1用于构造关于高光谱图像f(x,y,z)中的某点p0和其邻域上的点p1的加权相关性函数;
角点响应函数构造模块2用于根据所述加权相关性函数构造一个角点响应函数;
角点响应值计算模块3用于根据所述角点响应函数计算所述高光谱图像f(x,y,z)中的某点p0的Harris角点响应值和其邻域上所有点的Harris角点响应值;
角点判断模块4用于判断某点p0是否为所述高光谱图像f(x,y,z)的Harris角点,若所述高光谱图像f(x,y,z)中的某点p0的Harris角点响应值大于其邻域上所有点的Harris角点响应值,则该点p0即为Harris角点。
本发明提供的一种高光谱图像角点检测方法与系统,解决了目前的Harris角点检测方法只能用于灰度图像或视频图像而不能用于高光谱图像的问题;采用这种检测方法能获取高光谱图像的关键信息从而能更好地对高光谱图像进行分析,提高了高光谱图像模式识别效果。
以上所述仅为本发明的较佳实施例而已,并不用以限制本发明,凡在本发明的精神和原则之内所作的任何修改、等同替换和改进等,均应包含在本发明的保护范围之内。

Claims (12)

  1. 一种高光谱图像角点检测方法,其特征在于,所述方法包括下述步骤:
    步骤A,构造关于高光谱图像f(x,y,z)中的某点p0和其邻域上的点p1的加权相关性函数;
    步骤B,根据所述加权相关性函数构造一个角点响应函数;
    步骤C,根据所述角点响应函数计算所述高光谱图像f(x,y,z)中的某点p0的Harris角点响应值和其邻域上所有点的Harris角点响应值;
    步骤D,若所述高光谱图像f(x,y,z)中的某点p0的Harris角点响应值大于其邻域上所有点的Harris角点响应值,则该点p0即为所述高光谱图像f(x,y,z)的Harris角点。
  2. 如权利要求1所述的角点检测方法,其特征在于,所述加权相关性函数为:
    Figure PCTCN2016090015-appb-100001
    其中,点p0是高光谱图像f(x,y,z)中的一个像素,其坐标为(x,y,z),f(x,y,z)为点p0对应的像素值;点p1坐标为(x+Δx,y+Δy,z+Δz),f(x+Δx,y+Δy,z+Δz)为点p1对应的像素值;
    窗函数ω(x,y,z)采用高斯加权函数,如下所示:
    Figure PCTCN2016090015-appb-100002
    而由上式可得:
    Figure PCTCN2016090015-appb-100003
    其中,σ为高斯函数的尺度因子;
    Figure PCTCN2016090015-appb-100004
    为卷积运算符号,l为窗函数沿x方向移动的长度,m为窗函数沿y方向 移动的长度,r为窗函数沿z方向移动的长度,l=1,m=1,r=1。
  3. 如权利要求2所述的角点检测方法,其特征在于,所述加权相关性函数中
    Figure PCTCN2016090015-appb-100005
    表示为
    Figure PCTCN2016090015-appb-100006
    即:
    Figure PCTCN2016090015-appb-100007
    而,
    Figure PCTCN2016090015-appb-100008
    则,
    Figure PCTCN2016090015-appb-100009
    其中,
    Figure PCTCN2016090015-appb-100010
    式中,fx,fy,fz分别表示图像f(x,y,z)在x,y,z三个方向上的梯度,即
    Figure PCTCN2016090015-appb-100011
    Figure PCTCN2016090015-appb-100012
    Figure PCTCN2016090015-appb-100013
    上式中,ω表示高斯加权函数ω(x,y,z),
    Figure PCTCN2016090015-appb-100014
    为卷积符号,A、B、C、D、E、F分别对应矩阵M的各个元素。
  4. 如权利要求3所述的角点检测方法,其特征在于,所述角点响应函数为:
    R=det(M)-k(trace(M))3=(ABC+2DEF-BE2-AF2-CD2)-k(A+B+C)3
    其中,k=0.04,k为经验常数;det(M)表示矩阵M的行列式,trace(M)表示矩阵M的迹,其表达式如下:
    det(M)=λ1λ2λ3=ABC+2DEF-BE2-AF2-CD2
    trace(M)=λ123=A+B+C
    其中,λ1、λ2、λ3分别为矩阵M的特征值。
  5. 如权利要求3所述的角点检测方法,其特征在于,所述fx,fy,fz由如下公式计算:
    Figure PCTCN2016090015-appb-100015
    Figure PCTCN2016090015-appb-100016
    Figure PCTCN2016090015-appb-100017
    式中,Rx、Ry、Rz分别为水平、垂直、光谱三个方向的梯度算子,且大小均为5*5*5,
    Figure PCTCN2016090015-appb-100018
    为卷积符号;
    其中,水平梯度算子Rx(:,:,3)表达式如下,Rx的其他位置均为0:
    Figure PCTCN2016090015-appb-100019
    垂直梯度算子Ry(:,:,3)表达式如下,Ry的其他位置均为0:
    Figure PCTCN2016090015-appb-100020
    光谱方向的梯度算子Rz(3,3,:)表达式如下,Rz的其他位置均为0:
    Figure PCTCN2016090015-appb-100021
  6. 如权利要求5所述的角点检测方法,其特征在于,矩阵M由如下公式计算:
    Figure PCTCN2016090015-appb-100022
    其中,
    Figure PCTCN2016090015-appb-100023
    fxy=fx·fy,fyz=fy·fz,fxz=fx·fz
    ω(x,y,z)为高斯加权函数,σ为高斯函数的尺度因子,σ=1。
  7. 一种高光谱图像角点检测系统,其特征在于,所述系统包括:加权相关性函数构造模块、角点响应函数构造模块、角点响应值计算模块、角点判断模块;其中:
    加权相关性函数构造模块用于构造关于高光谱图像f(x,y,z)中的某点p0和其邻域上的点p1的加权相关性函数;
    角点响应函数构造模块用于根据所述加权相关性函数构造一个角点响应函数;
    角点响应值计算模块用于根据所述角点响应函数计算所述高光谱图像f(x,y,z)中的某点p0的Harris角点响应值和其邻域上所有点的Harris角点响应值;
    角点判断模块用于判断某点p0是否为所述高光谱图像f(x,y,z)的Harris角点,若所述高光谱图像f(x,y,z)中的某点p0的Harris角点响应值大于其邻域上所有点的Harris角点响应值,则该点p0即为Harris角点。
  8. 如权利要求7所述的角点检测系统,其特征在于,所述加权相关性函数为:
    Figure PCTCN2016090015-appb-100024
    其中,点p0是高光谱图像f(x,y,z)中的一个像素,其坐标为(x,y,z),f(x,y,z) 为点p0对应的像素值;点p1坐标为(x+Δx,y+Δy,z+Δz),f(x+Δx,y+Δy,z+Δz)为点p1对应的像素值;
    窗函数ω(x,y,z)采用高斯加权函数,如下所示:
    Figure PCTCN2016090015-appb-100025
    而由上式可得:
    Figure PCTCN2016090015-appb-100026
    其中,σ为高斯函数的尺度因子;
    Figure PCTCN2016090015-appb-100027
    为卷积运算符号,l为窗函数沿x方向移动的长度,m为窗函数沿y方向移动的长度,r为窗函数沿z方向移动的长度,l=1,m=1,r=1。
  9. 如权利要求8所述的角点检测系统,其特征在于,所述加权相关性函数中
    Figure PCTCN2016090015-appb-100028
    表示为
    Figure PCTCN2016090015-appb-100029
    即:
    Figure PCTCN2016090015-appb-100030
    而,
    Figure PCTCN2016090015-appb-100031
    则,
    Figure PCTCN2016090015-appb-100032
    其中,
    Figure PCTCN2016090015-appb-100033
    式中,fx,fy,fz分别表示图像f(x,y,z)在x,y,z三个方向上的梯度,即
    Figure PCTCN2016090015-appb-100034
    Figure PCTCN2016090015-appb-100035
    Figure PCTCN2016090015-appb-100036
    上式中,ω表示高斯加权函数ω(x,y,z),
    Figure PCTCN2016090015-appb-100037
    为卷积符号,A、B、C、D、E、F分别对应矩阵M的各个元素。
  10. 如权利要求9所述的角点检测系统,其特征在于,所述角点响应函数为:
    R=det(M)-k(trace(M))3=(ABC+2DEF-BE2-AF2-CD2)-k(A+B+C)3
    其中,k=0.04,k为经验常数;det(M)表示矩阵M的行列式,trace(M)表示矩阵M的迹,其表达式如下:
    det(M)=λ1λ2λ3=ABC+2DEF-BE2-AF2-CD2
    trace(M)=λ123=A+B+C
    其中,λ1、λ2、λ3分别为矩阵M的特征值。
  11. 如权利要求9所述的角点检测系统,其特征在于,所述fx,fy,fz由如下公式计算:
    Figure PCTCN2016090015-appb-100038
    Figure PCTCN2016090015-appb-100039
    Figure PCTCN2016090015-appb-100040
    式中,Rx、Ry、Rz分别为水平、垂直、光谱三个方向的梯度算子,且大小均为5*5*5,
    Figure PCTCN2016090015-appb-100041
    为卷积符号;
    其中,水平梯度算子Rx(:,:,3)表达式如下,Rx的其他位置均为0:
    Figure PCTCN2016090015-appb-100042
    垂直梯度算子Ry(:,:,3)表达式如下,Ry的其他位置均为0:
    Figure PCTCN2016090015-appb-100043
    光谱方向的梯度算子Rz(3,3,:)表达式如下,Rz的其他位置均为0:
    Figure PCTCN2016090015-appb-100044
  12. 如权利要求11所述的角点检测系统,其特征在于,矩阵M由如下公式计算:
    Figure PCTCN2016090015-appb-100045
    其中,
    Figure PCTCN2016090015-appb-100046
    fxy=fx·fy,fyz=fy·fz,fxz=fx·fz
    ω(x,y,z)为高斯加权函数,σ为高斯函数的尺度因子,σ=1。
PCT/CN2016/090015 2015-09-25 2016-07-14 一种高光谱图像角点检测方法与系统 WO2017049994A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201510621881.3 2015-09-25
CN201510621881.3A CN105139412B (zh) 2015-09-25 2015-09-25 一种高光谱图像角点检测方法与系统

Publications (1)

Publication Number Publication Date
WO2017049994A1 true WO2017049994A1 (zh) 2017-03-30

Family

ID=54724744

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2016/090015 WO2017049994A1 (zh) 2015-09-25 2016-07-14 一种高光谱图像角点检测方法与系统

Country Status (2)

Country Link
CN (1) CN105139412B (zh)
WO (1) WO2017049994A1 (zh)

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108960012A (zh) * 2017-05-22 2018-12-07 中科创达软件股份有限公司 特征点检测方法、装置及电子设备
CN109191537A (zh) * 2018-07-19 2019-01-11 成都智慧海派科技有限公司 一种视觉定位移动设备屏幕的方法
CN110188777A (zh) * 2019-05-31 2019-08-30 余旸 一种基于数据挖掘的多周期钢轨伤损数据对齐方法
CN110232390A (zh) * 2019-06-13 2019-09-13 长安大学 一种变化光照下图像特征提取方法
CN110490913A (zh) * 2019-07-22 2019-11-22 华中师范大学 基于角点与单线段编组的特征描述算子进行影像匹配方法
CN110728285A (zh) * 2019-08-23 2020-01-24 北京正安维视科技股份有限公司 一种基于动态变频的快速角点检测方法
CN111047614A (zh) * 2019-10-10 2020-04-21 南昌市微轲联信息技术有限公司 一种基于特征提取的复杂场景图像的目标角点提取方法
CN111311657A (zh) * 2020-03-12 2020-06-19 广东电网有限责任公司广州供电局 一种基于改进角点主方向分配的红外图像同源配准方法
CN111949270A (zh) * 2020-07-27 2020-11-17 中国建设银行股份有限公司 流程机器人的运行环境变化感知方法及装置
CN113129351A (zh) * 2021-03-10 2021-07-16 西安理工大学 一种基于光场傅里叶视差分层的特征检测方法
CN114821044A (zh) * 2022-05-31 2022-07-29 中煤科工机器人科技有限公司 一种基于梯度变换的方形指针式仪表示数识别方法
CN117635507A (zh) * 2024-01-26 2024-03-01 深圳市精森源科技有限公司 一种塑胶颗粒在线视觉检测方法及系统

Families Citing this family (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105139412B (zh) * 2015-09-25 2018-04-24 深圳大学 一种高光谱图像角点检测方法与系统
CN106340010B (zh) * 2016-08-22 2019-02-15 电子科技大学 一种基于二阶轮廓差分的角点检测方法
CN106303501B (zh) * 2016-08-23 2018-12-04 深圳市捷视飞通科技股份有限公司 基于图像稀疏特征匹配的立体图像重构方法及装置
CN106529472B (zh) * 2016-10-24 2019-08-02 深圳大学 基于大尺度高分辨率高光谱图像的目标探测方法及装置
CN107230220B (zh) * 2017-05-26 2020-02-21 深圳大学 一种新的时空Harris角点检测方法及装置
CN107742114B (zh) * 2017-11-09 2021-11-19 深圳大学 高光谱图像特征检测方法及装置
CN108009550B (zh) * 2017-11-09 2021-01-22 深圳大学 基于光谱曲线拟合的高光谱图像特征检测方法及装置
CN108492306A (zh) * 2018-03-07 2018-09-04 鞍钢集团矿业有限公司 一种基于图像轮廓的x型角点提取方法
CN109118473B (zh) * 2018-07-03 2022-04-12 深圳大学 基于神经网络的角点检测方法、存储介质与图像处理系统
CN109360264B (zh) * 2018-08-30 2023-05-26 深圳大学 图像统一模型的建立方法与装置
CN110006844A (zh) * 2019-05-22 2019-07-12 安徽大学 基于函数性主元分析的近红外光谱特征提取方法和系统
CN110751048B (zh) * 2019-09-20 2022-06-14 华中科技大学 基于图像特征自动选择谱线的激光探针分类方法及装置
CN111724425B (zh) * 2020-05-28 2023-11-10 交通运输部东海航海保障中心上海航标处 航标图拼接方法、装置及终端设备
CN112967241B (zh) * 2021-02-26 2023-09-12 西安理工大学 一种基于局部梯度引导的高光谱图像异常检测方法

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102800099A (zh) * 2012-07-19 2012-11-28 北京市遥感信息研究所 多特征多级别的可见光与高光谱图像高精度配准方法
CN105139412A (zh) * 2015-09-25 2015-12-09 深圳大学 一种高光谱图像角点检测方法与系统

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102800097B (zh) * 2012-07-19 2015-08-19 中国科学院自动化研究所 多特征多级别的可见光与红外图像高精度配准方法
CN102819839B (zh) * 2012-07-19 2015-06-03 北京市遥感信息研究所 多特征多级别的红外与高光谱图像的高精度配准方法
CN103824076A (zh) * 2014-02-28 2014-05-28 苏州大学 一种图像尺度不变特征的检测提取方法和系统
CN104318548B (zh) * 2014-10-10 2017-02-15 西安电子科技大学 一种基于空间稀疏度和sift特征提取的快速图像配准实现方法
CN104700359A (zh) * 2015-03-20 2015-06-10 南京大学 像平面不同极轴方向图像序列进行超分辨率重建的方法

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102800099A (zh) * 2012-07-19 2012-11-28 北京市遥感信息研究所 多特征多级别的可见光与高光谱图像高精度配准方法
CN105139412A (zh) * 2015-09-25 2015-12-09 深圳大学 一种高光谱图像角点检测方法与系统

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
LIU, YING ET AL.: "Optimization of Harris Corner Detection Algorithm", JOURNAL OF YUNNAN MINZU UNIVERSITY ( NATURAL SCIENCES EDITION, 31 March 2011 (2011-03-31), ISSN: 1672-8513 *
PAVITHRA, S. ET AL.: "Agile Segmentation and Classification for Hyper Spectral Image Using Harris Corner Detector", INTERNATIONAL JOURNAL OF SCIENTIFIC RESEARCH, vol. 4, no. 7, 31 July 2015 (2015-07-31), XP055373315, ISSN: 2277-8179 *
SI , MEILING: "Multispectral Remote Sensing Image Registration and Fusion Scheme Research", ELECTRONIC TECHNOLOGY & INFORMATION SCIENCE , CHINA MASTER'S THESES FULL-TEXT DATABASE, 2012, 15 January 2012 (2012-01-15), ISSN: 1674-0246 *

Cited By (21)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108960012B (zh) * 2017-05-22 2022-04-15 中科创达软件股份有限公司 特征点检测方法、装置及电子设备
CN108960012A (zh) * 2017-05-22 2018-12-07 中科创达软件股份有限公司 特征点检测方法、装置及电子设备
CN109191537A (zh) * 2018-07-19 2019-01-11 成都智慧海派科技有限公司 一种视觉定位移动设备屏幕的方法
CN110188777A (zh) * 2019-05-31 2019-08-30 余旸 一种基于数据挖掘的多周期钢轨伤损数据对齐方法
CN110188777B (zh) * 2019-05-31 2023-08-25 东莞先知大数据有限公司 一种基于数据挖掘的多周期钢轨伤损数据对齐方法
CN110232390A (zh) * 2019-06-13 2019-09-13 长安大学 一种变化光照下图像特征提取方法
CN110232390B (zh) * 2019-06-13 2022-10-14 长安大学 一种变化光照下图像特征提取方法
CN110490913A (zh) * 2019-07-22 2019-11-22 华中师范大学 基于角点与单线段编组的特征描述算子进行影像匹配方法
CN110490913B (zh) * 2019-07-22 2022-11-22 华中师范大学 基于角点与单线段编组的特征描述算子进行影像匹配方法
CN110728285A (zh) * 2019-08-23 2020-01-24 北京正安维视科技股份有限公司 一种基于动态变频的快速角点检测方法
CN111047614A (zh) * 2019-10-10 2020-04-21 南昌市微轲联信息技术有限公司 一种基于特征提取的复杂场景图像的目标角点提取方法
CN111047614B (zh) * 2019-10-10 2023-09-29 南昌市微轲联信息技术有限公司 一种基于特征提取的复杂场景图像的目标角点提取方法
CN111311657A (zh) * 2020-03-12 2020-06-19 广东电网有限责任公司广州供电局 一种基于改进角点主方向分配的红外图像同源配准方法
CN111311657B (zh) * 2020-03-12 2023-04-25 广东电网有限责任公司广州供电局 一种基于改进角点主方向分配的红外图像同源配准方法
CN111949270A (zh) * 2020-07-27 2020-11-17 中国建设银行股份有限公司 流程机器人的运行环境变化感知方法及装置
CN113129351A (zh) * 2021-03-10 2021-07-16 西安理工大学 一种基于光场傅里叶视差分层的特征检测方法
CN113129351B (zh) * 2021-03-10 2023-08-11 西安理工大学 一种基于光场傅里叶视差分层的特征检测方法
CN114821044A (zh) * 2022-05-31 2022-07-29 中煤科工机器人科技有限公司 一种基于梯度变换的方形指针式仪表示数识别方法
CN114821044B (zh) * 2022-05-31 2024-05-03 中煤科工机器人科技有限公司 一种基于梯度变换的方形指针式仪表示数识别方法
CN117635507A (zh) * 2024-01-26 2024-03-01 深圳市精森源科技有限公司 一种塑胶颗粒在线视觉检测方法及系统
CN117635507B (zh) * 2024-01-26 2024-04-09 深圳市精森源科技有限公司 一种塑胶颗粒在线视觉检测方法及系统

Also Published As

Publication number Publication date
CN105139412B (zh) 2018-04-24
CN105139412A (zh) 2015-12-09

Similar Documents

Publication Publication Date Title
WO2017049994A1 (zh) 一种高光谱图像角点检测方法与系统
US11244197B2 (en) Fast and robust multimodal remote sensing image matching method and system
Chen et al. Building change detection with RGB-D map generated from UAV images
Li et al. LNIFT: Locally normalized image for rotation invariant multimodal feature matching
Han et al. Fusion of color and infrared video for moving human detection
CN111079556A (zh) 一种多时相无人机视频图像变化区域检测及分类方法
Li et al. Extracting man-made objects from high spatial resolution remote sensing images via fast level set evolutions
Liu et al. A contrario comparison of local descriptors for change detection in very high spatial resolution satellite images of urban areas
CN108154159B (zh) 一种基于多级检测器的具有自恢复能力的目标跟踪方法
JP6899189B2 (ja) ビジョンシステムで画像内のプローブを効率的に採点するためのシステム及び方法
CN109919960B (zh) 一种基于多尺度Gabor滤波器的图像连续边缘检测方法
CN107909018B (zh) 一种稳健的多模态遥感影像匹配方法和系统
CN110245566B (zh) 一种基于背景特征的红外目标远距离追踪方法
CN107240130A (zh) 遥感影像配准方法、装置及系统
Li et al. Semiautomatic airport runway extraction using a line-finder-aided level set evolution
CN113763274B (zh) 一种联合局部相位锐度定向描述的多源图像匹配方法
Li et al. A Harris corner detection algorithm for multispectral images based on the correlation
Chen et al. Scene segmentation of remotely sensed images with data augmentation using U-net++
Yang et al. A research of feature-based image mosaic algorithm
Liu et al. Keypoint matching by outlier pruning with consensus constraint
Peng et al. Seamless UAV hyperspectral image stitching using optimal seamline detection via graph cuts
CN110738098A (zh) 一种目标的识别定位与锁定跟踪方法
Liu et al. SAR image matching based on speeded up robust feature
Liu et al. A novel adaptive weights proximity matrix for image registration based on R-SIFT
Yao et al. Real-time multiple moving targets detection from airborne IR imagery by dynamic Gabor filter and dynamic Gaussian detector

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 16847876

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 17/07/2018)

122 Ep: pct application non-entry in european phase

Ref document number: 16847876

Country of ref document: EP

Kind code of ref document: A1