WO2018081929A1 - 一种高光谱遥感图像特征提取和分类方法及其*** - Google Patents

一种高光谱遥感图像特征提取和分类方法及其*** Download PDF

Info

Publication number
WO2018081929A1
WO2018081929A1 PCT/CN2016/104268 CN2016104268W WO2018081929A1 WO 2018081929 A1 WO2018081929 A1 WO 2018081929A1 CN 2016104268 W CN2016104268 W CN 2016104268W WO 2018081929 A1 WO2018081929 A1 WO 2018081929A1
Authority
WO
WIPO (PCT)
Prior art keywords
pixel
binary
spatial
classification
dimensional
Prior art date
Application number
PCT/CN2016/104268
Other languages
English (en)
French (fr)
Inventor
贾森
胡杰
邓琳
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
Priority to PCT/CN2016/104268 priority Critical patent/WO2018081929A1/zh
Publication of WO2018081929A1 publication Critical patent/WO2018081929A1/zh
Priority to US15/978,189 priority patent/US10509984B2/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/13—Satellite images
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/50—Extraction of image or video features by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/60—Type of objects
    • G06V20/64—Three-dimensional objects
    • G06V20/647—Three-dimensional objects by matching two-dimensional images to three-dimensional objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
    • G06V10/467—Encoded features or binary features, e.g. local binary patterns [LBP]

Definitions

  • the invention relates to the field of image processing, and in particular to a method and system for extracting and classifying features of hyperspectral remote sensing images.
  • Hyperspectral remote sensing images obtained by sensing remote sensing sensors on the ground in the visible, near-infrared, mid-infrared and thermal infrared bands of the electromagnetic spectrum not only provide spatial geometric information of the features, but also contain rich reflections. Spectral information specific to physical properties. Therefore, the feature and knowledge of the spectral object, texture, shape and other features of the object object are extracted from the hyperspectral remote sensing image, and the hyperspectral classification technique for feature recognition is born.
  • hyperspectral remote sensing images contain Rich spatial structure information, therefore, comprehensive consideration of the spatial and spectral information of hyperspectral data can effectively improve the classification accuracy of features, and obtain a classification map with better spatial continuity.
  • the classification of hyperspectral remote sensing images combined with spectral-spatial features has become a hot topic.
  • the core problem lies in how to extract spatial structure information such as texture, shape, object and semantics, and how to realize the organic combination of spectral information and spatial features. According to the combination of spectral features and spatial features, spectral-spatial classification can be roughly divided into two strategies: combined processing and fusion processing.
  • the combined processing strategy inputs the extracted spatial features together with the spectral features into the classifier to obtain the classification results (referred to as spatial information preprocessing), or uses the image segmentation method to regularize the original classification results to obtain the same space.
  • a higher quality classification map (referred to as spatial information post processing).
  • the former mainly includes morphological analysis, spatial filtering based on edge preservation and sparse representation, and other spatial features. Take the method, the latter mainly includes methods such as multiple logistic regression and hypergraph generation.
  • the combined spectral-spatial feature classification method has better classification effect and lower computational complexity.
  • the hyperspectral remote sensing image itself is a three-dimensional structure.
  • the spatial features and spectral features obtained by the combined spatial information processing method are separated, and the context relationship between the spectral and spatial structures is neglected.
  • the spatial information post-processing method is greatly affected by the classification result. That is, if most of a certain type of feature is misclassified, the use of post-processing methods will aggravate this error.
  • the fusion processing strategy obtains a spatially integrated characterization by directly performing mathematical operations on a set of pre-defined multi-scale kernel or three-dimensional structure filters with the original hyperspectral data. Because this method treats the three-dimensional hyperspectral remote sensing image as a whole, it can fully exploit the context relationship between the spectral domain and the spatial domain, and has gained more and more attention in recent years. However, since the spectral and spatial distribution structure of the features is usually unknown, it is necessary to define enough scale or three-dimensional structure filters to obtain sufficient spatial-spectral integrated representation features, resulting in extremely high feature dimensions and feature redundancy. The large classification process is very time consuming and reduces the practicability of the algorithm.
  • the spatial-spectral context relationship is neglected in the prior art, which is insufficient and inaccurate in characterizing and extracting the internal structure and statistical relationship of data spectrum and spatial information, and contains a large amount of redundant information.
  • the object of the present invention is to provide a method and system for extracting and classifying features of hyperspectral remote sensing images, and aiming at solving the spatial-spectral context relationship in the prior art, which is sensitive to noise and redundant in model features. And the problem of high computational complexity.
  • the invention provides a feature extraction and classification method for hyperspectral remote sensing images, which mainly comprises:
  • the binarization step compares the gray value of the six vertex pixels of the regular octahedron with the pixel value of the center point one by one, and if the absolute value of the difference between the two is less than the preset discriminant threshold, the corresponding vertex pixel mark Is 1, otherwise marked as 0, to form a binary pattern T1 of the local spatial structure of the central pixel, where T1 ⁇ t(s(g 0 -g c ,s(g 1 -g c ),...,s( g 5 -g c )), g c represents the pixel value of the center point;
  • the encoding step for the binary mode having the same spatial topology, is uniquely labeled using the number of 1 of the binary patterns to obtain the three-dimensional local binary encoding corresponding to the binary patterns, namely: Where the spatial topology of the binary mode passes To be measured, the same binary value ⁇ their spatial patterns indicate the same topology, where, g i, g j represent the binary mode of the i, j th sampling point, g i and g j and i ⁇ adjacent j,i,j ⁇ 0,1,2,3,4,5 ⁇ ;
  • the statistical step after obtaining the three-dimensional local binary pattern coding of each pixel, the eight three-dimensional parts of 0, 1, 2, 3, 4, 5, 6, and 7 are counted in the nxn rectangular neighborhood of each pixel in the pixel.
  • the histogram features of each pixel in the pixel are sequentially connected in series to obtain a three-dimensional local binary pattern characteristic corresponding to the pixel;
  • the classification step sends the obtained three-dimensional local binary pattern features to the classifier for classification.
  • the present invention also provides a hyperspectral remote sensing image feature extraction and classification system, the system comprising:
  • a sampling module for using six vertices of a regular octahedral domain of each pixel as sampling points of a three-dimensional local binary pattern, and using the gray distribution T of the six sampling points to describe a local spatial spectral structure of the pixel, wherein, T ⁇ t(g 0 , g 1 , g 2 , g 3 , g 4 , g 5 ), representing the gray scale distribution of these six sampling points, g 0 , g 1 , g 2 , g 3 , g 4 , g 5 Representing the pixel values of the six sampling points respectively;
  • a binarization module for comparing gray values of six vertex pixels of a regular octahedron with pixel values of a center point, and if the absolute value of the difference is less than a preset discriminant threshold, the corresponding vertex is The pixel is marked as 1, otherwise marked as 0, to form a binary pattern T1 of the local spatial spectral structure of the central pixel, where T1 ⁇ t(s(g 0 -g c ,s(g 1 -g c ),..., s(g 5 -g c )), g c represents the pixel value of the center point;
  • An encoding module for uniquely marking a binary pattern having the same spatial topology using the number of ones of the binary patterns to obtain a three-dimensional local binary encoding corresponding to the binary patterns, namely: Where the spatial topology of the binary mode passes To be measured, the same binary value ⁇ their spatial patterns indicate the same topology, where, g i, g j represent the binary mode of the i, j th sampling point, g i and g j and i ⁇ adjacent j,i,j ⁇ 0,1,2,3,4,5 ⁇ ;
  • a statistical module configured to count 8 of 0, 1, 2, 3, 4, 5, 6, and 7 in the nxn rectangular neighborhood of each pixel in the pixel after obtaining the three-dimensional local binary pattern encoding of each pixel.
  • a series module for sequentially concatenating the histogram features of each pixel in the pixel to obtain a three-dimensional local binary pattern characteristic corresponding to the pixel;
  • a classification module is configured to send the obtained three-dimensional local binary pattern features to the classifier for classification.
  • the technical solution provided by the invention extends the two-dimensional local binary pattern (LBP) to the three-dimensional LBP, fully utilizes the spatial-spectral context relationship in the hyperspectral remote sensing map, and introduces a relaxation threshold discriminating operation, The noise has good robustness.
  • the rotation-invariant three-dimensional LBP model proposed by the present invention considers the essential features of hyperspectral remote sensing images, and has the advantages of strong pertinence, simple operation, and high computational efficiency.
  • FIG. 1 is a flow chart of a feature extraction and classification method for hyperspectral remote sensing images according to an embodiment of the present invention
  • FIG. 2 is a schematic diagram of a sampling model using a regular octahedron approximate spherical surface in a three-dimensional LBP according to an embodiment of the present invention
  • FIG. 3 is a schematic diagram of a process of binarizing a local spatial spectrum structure according to an embodiment of the present invention
  • FIG. 4 is a schematic diagram of a binary mode of seven different spatial topologies in an embodiment of the present invention.
  • FIG. 5 is a schematic structural diagram of a hyperspectral remote sensing image feature extraction and classification system 10 according to an embodiment of the present invention.
  • FIG. 1 is a flowchart of a feature extraction and classification method for hyperspectral remote sensing images according to an embodiment of the present invention.
  • the sampling step, the binarization step, the encoding step, the statistical step, the series step, and the sorting step in the method respectively correspond to steps S1-S6 in FIG. 1 .
  • step S1 six vertices of the regular octahedral domain of each pixel are taken as sampling points of the three-dimensional local binary pattern, and the local spatial spectral structure of the pixel is described by using the gray distribution T of the six sampling points, wherein T ⁇ t (g 0, g 1, g 2, g 3, g 4, g 5), which represents six samples gradation distribution, g 0, g 1, g 2, g 3, g 4, g 5 The pixel values of the six sample points are respectively indicated.
  • a three-dimensional Local Binary Patterns (LBP) model is used to sample the spherical neighbors of the center pixel with a radius R, and a regular octahedron is used to approximate the spherical surface, that is, the sampling points are six regular octahedrons. Vertex, if the radius R is 1, the six sampling points are the six pixels of the center pixel left, right, up, down, front, and back. The gray scale distribution of the six sampling points represents the local spatial spectrum structure of the central pixel.
  • Figure 2 shows a sampling model using a regular octahedron approximate spherical surface in a three-dimensional LBP.
  • step S2 the gray value of the six vertex pixels of the regular octahedron is compared with the pixel value of the center point one by one, and if the absolute value of the difference between the two is less than the preset discriminant threshold, the corresponding vertex pixel mark is 1, 0 otherwise labeled, to form a binary pattern configuration of the local blank spectrum of Tl center pixel, wherein, T1 ⁇ t (s (g 0 -g c, s (g 1 -g c), ..., s ( g 5 -g c )), g c represents the pixel value of the center point.
  • the binarization is binarization of the local spatial spectral structure, and in order to calculate the grayscale distribution T of the six sampling points, for each pixel, the circumscribed octahedron neighborhood with a radius of R is R
  • the gray value of the vertex pixel is compared with the pixel value of the center point, where the center point is the center point of the regular octahedron, and if the absolute value of the difference between the gray value of the vertex pixel and the center pixel value is less than a preset discriminating threshold ⁇ , the vertex pixel is marked as 1, otherwise it is marked 0.
  • the vertices of the marked positive octahedral neighborhood are called the binary mode T1 of the local spatial spectral structure of the central pixel, and the binary mode T1 can be expressed as
  • g c represents the pixel value of the center point
  • ⁇ represents a preset discriminant threshold
  • the local spatial spectrum structure of the center pixel is binarized using the gradation value of the center pixel. If the absolute value of the difference between the gray value of the sample point and the gray value of the center pixel is greater than the discrimination threshold, the sample point is marked as 1, otherwise the flag is 0.
  • the local spatial spectrum structure of the central pixel is binarized and called its binary mode.
  • Figure 3 shows the process of binarization of a local spatial spectrum structure. The left side of FIG. 3 shows a local spatial spectrum structure of the center point pixel with a gray value of 145, and the right side of FIG. 3 shows the binary mode corresponding to the binarization.
  • step S3 for the binary mode having the same spatial topology, the number of 1s of the binary patterns is uniquely labeled to obtain the three-dimensional local binary coding corresponding to the binary patterns, namely: Where the spatial topology of the binary mode passes To measure, the binary patterns with the same threshold value indicate that their spatial topologies are the same.
  • g i and g j respectively represent the i-th and j-th sampling points in the binary mode, and gi is adjacent to g j and i ⁇ j , i, j ⁇ 0,1,2,3,4,5 ⁇ .
  • the spatial topology of the binary mode is calculated using the following expression:
  • the encoding is a three-dimensional local binary pattern encoding, and in order to distinguish the local spatial spectral structure around each pixel, the binary pattern is marked with a different encoding. Since different binary modes may correspond to the same local spatial spectral structure, in order to distinguish this phenomenon, the binary mode is coded according to the spatial topology of the binary mode.
  • the spatial topology of the binary mode refers to the distribution of 1 in the three-dimensional space corresponding to the binary mode.
  • Six different sampling modes are generated for six sampling points.
  • the 64 binary modes are classified into eight according to their spatial topologies. These eight modes correspond to different local spatial spectral structures, and 0 and 1, respectively. 2, 3, 4, 5, 6, and 7 are marked.
  • Figure 4 shows a binary mode diagram of seven different spatial topologies of 0, 1, 2, 3, 4, 5, and 6, including 1, 6, 12, 8, 12, 6, and 1 binary modes.
  • the solid point represents 1
  • the hollow point represents 0.
  • 1 is close together in the regular octahedral neighborhood and is not in line, the difference is that the number of 1 is different.
  • step S4 after obtaining the three-dimensional local binary pattern encoding of each pixel, eight of 0, 1, 2, 3, 4, 5, 6, and 7 are counted in the nxn rectangular neighborhood of each pixel in the pixel.
  • the symbiosis frequency of the three-dimensional local binary mode coding is used to obtain the histogram features of the pixel.
  • the histogram feature represents the distribution of eight different three-dimensional local binary codes within the nxn rectangular neighborhood of the pixel.
  • step S5 the histogram features of each pixel in the pixel are sequentially connected in series to obtain a three-dimensional LBP feature corresponding to the pixel;
  • step S6 the obtained three-dimensional LBP features are sent to a classifier for classification.
  • the classification of hyperspectral remote sensing images can be realized by using common classifiers such as KNN (k-Nearest Neighbor), SVM (Support Vector Machine), and SRC (Sparse Representation Classification).
  • KNN k-Nearest Neighbor
  • SVM Small Vector Machine
  • SRC Small Representation Classification
  • the invention provides a feature extraction and classification method for hyperspectral remote sensing images, and extends the two-dimensional LBP to the three-dimensional LBP, fully utilizing the spatial-spectral context relationship in the hyperspectral remote sensing image, and introducing a relaxation threshold discriminating operation to the noise
  • the rotation-invariant three-dimensional LBP model proposed by the present invention considers the essential features of hyperspectral remote sensing images, and has the advantages of strong pertinence, simple operation, and high computational efficiency.
  • a hyperspectral remote sensing image feature extraction and classification system provided by the present invention will be described in detail below.
  • FIG. 5 a schematic structural diagram of a hyperspectral remote sensing image feature extraction and classification system 10 according to an embodiment of the present invention is shown.
  • the hyperspectral remote sensing image feature extraction and classification system 10 mainly includes a sampling module 11, a binarization module 12, an encoding module 13, a statistics module 14, a serial module 15, and a classification module 16.
  • a sampling module 11 is configured to use six vertices of a positive octahedral field of each pixel as sampling points of a three-dimensional local binary pattern, and describe a local spatial spectral structure of the pixel by using the gray distribution T of the six sampling points, where T ⁇ t(g 0 , g 1 , g 2 , g 3 , g 4 , g 5 ), representing the gray scale distribution of these six sampling points, g 0 , g 1 , g 2 , g 3 , g 4 , g 5 indicates the pixel values of the six sampling points, respectively.
  • the sampling method in the sampling module 11 is as described in the related description in step S1, and is not described again here.
  • the binarization module 12 is configured to compare the gray value of the six vertex pixels of the regular octahedron with the pixel value of the center point one by one, and if the absolute value of the difference between the two is less than the preset discriminant threshold, the corresponding The vertex pixel is marked as 1, otherwise marked as 0, to form a binary pattern T1 of the local spatial spectral structure of the central pixel, where T1 ⁇ t(s(g 0 -g c ,s(g 1 -g c ),... , s(g 5 -g c )), g c represents the pixel value of the center point.
  • the binarization method in the binarization module 12 is as described in the related description in step S2, and is not described again here.
  • the encoding module 13 is configured to uniquely mark the binary pattern with the same spatial topology by using the number of 1 of the binary patterns to obtain the three-dimensional local binary encoding corresponding to the binary patterns, namely: Where the spatial topology of the binary mode passes To be measured, the same binary value ⁇ their spatial patterns indicate the same topology, where, g i, g j represent the binary mode of the i, j th sampling point, g i and g j and i ⁇ adjacent j,i,j ⁇ 0,1,2,3,4,5 ⁇ .
  • the encoding method in the encoding module 13 is as described in the related description in step S3, and will not be described again here.
  • the statistic module 14 is configured to calculate 0, 1, 2, 3, 4, 5, 6, and 7 in the nxn rectangular neighborhood of each pixel in the pixel after obtaining the three-dimensional local binary pattern encoding of each pixel.
  • the co-occurrence frequencies of the three-dimensional local binary mode are encoded to obtain the histogram features of the pixels.
  • the serial module 15 is configured to serially connect the histogram features of each pixel in the pixel to obtain a three-dimensional local binary pattern feature corresponding to the pixel.
  • the classification module 16 is configured to send the obtained three-dimensional local binary pattern features to the classifier for classification.
  • the hyperspectral remote sensing image feature extraction and classification system 10 provided by the invention extends the two-dimensional LBP to the three-dimensional LBP, fully utilizes the spatial-spectral context relationship in the hyperspectral remote sensing image, and introduces a relaxation threshold discriminating operation, The noise has good robustness.
  • the rotation-invariant three-dimensional LBP model proposed by the present invention considers the essential features of hyperspectral remote sensing images, and has the advantages of strong pertinence, simple operation, and high computational efficiency.
  • the three-dimensional LBP feature extracted by the invention is stronger than the two-dimensional LBP feature, and is different in two
  • the classification results of resolutions and data sets of different sizes are described and verified.
  • the second data set was obtained by the National Science Foundation-funded center using airborne laser mapping on June 23, 2012 at the Houston campus and surrounding areas (Houston University, URL: http://www.grss-ieee. Org/community/technical-committees/data-fusion/2013-ieee-grss-data-fusion-contest/).
  • the data set has 144 bands, each band image size is 349x1905, spatial resolution is 2.5m / pixel, 15029 mark samples, a total of 15 types of features.
  • the classification results using sparse representations show that the classification results of 3D LBP features are much higher than the 2D LBP, especially on the small sample classification problem, and the classification results of 2D LBP features on the Pavia Centre dataset in 3 samples. It is 72.75%, and the classification result of 3D LBP features is as high as 92.87%. On the Houston University dataset, the classification result of 2D LBP features is 45.98%, and the classification result of 3D LBP features is as high as 68.61%.
  • the above results show that the classification accuracy of 3D LBP features is much higher than that of 2D LBP, and the effect is particularly obvious on small sample classification problems.
  • each unit included is only divided according to functional logic, but is not limited to the above division, as long as the corresponding function can be implemented; in addition, the specific name of each functional unit is also They are only used to facilitate mutual differentiation and are not intended to limit the scope of the present invention.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Multimedia (AREA)
  • Astronomy & Astrophysics (AREA)
  • Remote Sensing (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Image Analysis (AREA)

Abstract

一种高光谱遥感图像特征提取和分类方法及***，其中，所述方法包括：采样步骤、二值化步骤、编码步骤、统计步骤、串联步骤以及分类步骤。将二维LBP扩展到三维LBP，充分利用了高光谱遥感图像中空间-光谱上下文关系，通过引入了松弛阈值判别操作，对噪声有很好的鲁棒性，而且，所涉及的三维LBP模型考虑了高光谱遥感图像的本质特征，具备针对性强、操作简单以及计算效率高的优点。

Description

一种高光谱遥感图像特征提取和分类方法及其*** 技术领域
本发明涉及图像处理领域，尤其涉及一种高光谱遥感图像特征提取和分类方法及其***。
背景技术
由遥感传感器在电磁波谱的可见光、近红外、中红外和热红外波段范围内对地面上的物质成像得到的高光谱遥感图像不仅可提供地物的空间几何信息，同时也包含丰富的反映地物特有物理性状的光谱信息。因此，从高光谱遥感图像中提取地物对象的光谱、纹理、形状等特征和知识，进行地物识别的高光谱分类技术应运而生。
最早的高光谱分类技术直接使用光谱特征进行分类。然而，受光照、气候变化、云层厚度以及混合像元等因素的影响，高光谱遥感图像中存在大量同物异谱和异物同谱的现象，导致误分类现象严重；同时，高光谱遥感图像包含了丰富的地物空间结构信息，因此，综合考虑高光谱数据的空间和光谱信息可以有效提升地物的分类精度，获得空间连续性较好的分类图。光谱-空间特征相结合的高光谱遥感图像分类研究已成为当前热点，其核心问题在于如何提取纹理、形状、对象、语义等空间结构信息，以及如何实现光谱信息与空间特征的有机结合。根据光谱特征和空间特征结合方式的不同，可以将光谱-空间分类大致分为组合式处理和融合式处理两种策略。
组合式处理策略将提取到的空间特征与光谱特征一起输入分类器得到分类结果(简称为空间信息预处理)，或利用图像分割的方法对原始分类结果进行对象的正则化处理，以得到空间同质度较高的分类图(简称为空间信息后处理)。前者主要包括形态学分析、基于边缘保持和稀疏表示的空间滤波等空间特征提 取方法，后者主要包括多元逻辑回归、超图生成等方法。通过空间信息的引入，组合式光谱-空间特征分类方法的分类效果较好，且计算复杂度较低。但高光谱遥感影像本身是三维结构，组合式空间信息处理方法获得的空间特征与光谱特征是分离的，忽略了光谱和空间结构的上下文关系；同时空间信息后处理方式受分类结果的影响大，即若某一类地物的大部分被错分，利用后处理方法会加剧这种错误。
融合式处理策略通过一组预先定义的多尺度核或三维结构滤波器与原始高光谱数据直接进行数学运算获得空谱一体的特征描述。由于该类方法将三维高光谱遥感影像作为整体进行处理，能够充分挖掘光谱域和空间域的上下文关系，近年来获得了越来越多的关注。但是，由于地物的光谱和空间分布结构通常是未知的，需要定义足够多的尺度或三维结构滤波器才能得到充足的空谱一体化表示特征，导致产生的特征维度极高，特征冗余度大使得分类过程非常耗时，降低了算法的实用性。
目前，现有技术中忽略了空间-光谱的上下文关系，在刻画和提取数据光谱和空间信息的内在结构和统计关系上是不充分和不精确的，而且包含有大量的冗余信息。
发明内容
有鉴于此，本发明的目的在于提供一种高光谱遥感图像特征提取和分类方法及其***，旨在解决现有技术中忽略了空间-光谱的上下文关系，对噪声敏感且模型特征冗余，以及计算方法复杂度较高的问题。
本发明提出一种高光谱遥感图像特征提取和分类方法，主要包括：
采样步骤、将每个像素的正八面体领域的六个顶点作为三维局部二值模式的采样点，并利用这六个采样点的灰度分布T描述像素的局部空谱结构，其中，T≈t(g0，g1，g2，g3，g4，g5)，表示这六个采样点的灰度分布，g0、g1、g2、g3、g4、g5分别表示这六个采样点的像素值；
二值化步骤、将正八面体的六个顶点像素的灰度值与中心点的像素值一一进行比较，如果二者差值的绝对值小于预设的判别阈值，则将对应的顶点像素标记为1，否则标记为0，以形成中心像素的局部空谱结构的二值模式T1，其中，T1≈t(s(g0-gc，s(g1-gc)，…，s(g5-gc))，gc表示中心点的像素值；
编码步骤、对于具有相同空间拓扑结构的二值模式，使用这些二值模式中1的数量进行唯一标记，以得到这些二值模式对应的三维局部二值编码，即：
Figure PCTCN2016104268-appb-000001
其中，二值模式的空间拓扑结构通过
Figure PCTCN2016104268-appb-000002
来进行衡量，Γ值相同的二值模式表明它们的空间拓扑结构相同，这里，gi、gj分别表示二值模式中第i、j个采样点，gi与gj相邻且i≠j，i，j∈{0，1，2，3，4，5}；
统计步骤、在得到每个像素的三维局部二值模式编码后，在像元中每个像素的nxn矩形邻域内统计0、1、2、3、4、5、6、7这8个三维局部二值模式编码的共生频率，来得到像素的直方图特征；
串联步骤、将像元中每个像素的直方图特征依次进行串联，得到像元对应的三维局部二值模式特征；
分类步骤、将得到的三维局部二值模式特征送入分类器进行分类。
另一方面，本发明还提供一种高光谱遥感图像特征提取和分类***，所述***包括：
采样模块，用于将每个像素的正八面体领域的六个顶点作为三维局部二值模式的采样点，并利用这六个采样点的灰度分布T描述像素的局部空谱结构，其中，T≈t(g0，g1，g2，g3，g4，g5)，表示这六个采样点的灰度分布，g0、g1、g2、g3、g4、g5分别表示这六个采样点的像素值；
二值化模块，用于将正八面体的六个顶点像素的灰度值与中心点的像素值一一进行比较，如果二者差值的绝对值小于预设的判别阈值，则将对应的顶点 像素标记为1，否则标记为0，以形成中心像素的局部空谱结构的二值模式T1，其中，T1≈t(s(g0-gc，s(g1-gc)，…，s(g5-gc))，gc表示中心点的像素值；
编码模块，用于对于具有相同空间拓扑结构的二值模式，使用这些二值模式中1的数量进行唯一标记，以得到这些二值模式对应的三维局部二值编码，即：
Figure PCTCN2016104268-appb-000003
其中，二值模式的空间拓扑结构通过
Figure PCTCN2016104268-appb-000004
来进行衡量，Γ值相同的二值模式表明它们的空间拓扑结构相同，这里，gi、gj分别表示二值模式中第i、j个采样点，gi与gj相邻且i≠j，i，j∈{0，1，2，3，4，5}；
统计模块，用于在得到每个像素的三维局部二值模式编码后，在像元中每个像素的nxn矩形邻域内统计0、1、2、3、4、5、6、7这8个三维局部二值模式编码的共生频率，来得到像素的直方图特征；
串联模块，用于将像元中每个像素的直方图特征依次进行串联，得到像元对应的三维局部二值模式特征；
分类模块，用于将得到的三维局部二值模式特征送入分类器进行分类。
本发明提供的技术方案，将二维局部二值模式(Local Binary Patterns，LBP)扩展到三维LBP，充分利用了高光谱遥感图中空间-光谱的上下文关系，通过引入了松弛阈值判别操作，对噪声有很好的鲁棒性，本发明提出的旋转不变三维LBP模型考虑高光谱遥感图像的本质特征，具备针对性强、操作简单以及计算效率高的优点。
附图说明
图1为本发明一实施方式中高光谱遥感图像特征提取和分类方法流程图；
图2为本发明一实施方式中在三维LBP中使用正八面体近似球面的采样模型示意图；
图3为本发明一实施方式中局部空谱结构被二值化的过程示意图；
图4为本发明一实施方式中七种不同空间拓扑结构的二值模式示意图；
图5为本发明一实施方式中高光谱遥感图像特征提取和分类***10的结构示意图。
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。
以下将对本发明所提供的一种高光谱遥感图像特征提取和分类方法进行详细说明。
请参阅图1，为本发明一实施方式中高光谱遥感图像特征提取和分类方法流程图。
在本实施方式中，该方法中的采样步骤、二值化步骤、编码步骤、统计步骤、串联步骤以及分类步骤分别对应图1中的步骤S1-S6。
在步骤S1中，将每个像素的正八面体领域的六个顶点作为三维局部二值模式的采样点，并利用这六个采样点的灰度分布T描述像素的局部空谱结构，其中，T≈t(g0，g1，g2，g3，g4，g5)，表示这六个采样点的灰度分布，g0、g1、g2、g3、g4、g5分别表示这六个采样点的像素值。
在本实施方式中，利用三维局部二值模式(Local Binary Patterns，LBP)模型在中心像素的半径为R的球面邻域上进行采样，使用正八面体近似球面，即采样点是正八面体的6个顶点，如果半径R为1，则这6个采样点分别是中心像素左、右、上、下、前、后的6个像素。6个采样点的灰度分布表征中心像素的局部空谱结构。图2展示了三维LBP中使用正八面体近似球面的采样模型。
在步骤S2中，将正八面体的六个顶点像素的灰度值与中心点的像素值一 一进行比较，如果二者差值的绝对值小于预设的判别阈值，则将对应的顶点像素标记为1，否则标记为0，以形成中心像素的局部空谱结构的二值模式T1，其中，T1≈t(s(g0-gc，s(g1-gc)，…，s(g5-gc))，gc表示中心点的像素值。
在本实施方式中，该二值化为局部空谱结构的二值化，为了计算6个采样点的灰度分布T，对于每个像素，将其外接圆半径为R的正八面体邻域的顶点像素的灰度值与中心点的像素值进行比较，该中心点为所述正八面体的中心点，若顶点像素的灰度值与中心像素值差值的绝对值小于预设的判别阈值δ，则该顶点像素被标记为1，否则标记0。标记后的正八面体邻域的顶点称为中心像素的局部空谱结构的二值模式T1，该二值模式T1可以表示为
T1≈t(s(g0-gc，s(g1-gc)，…，s(g5-gc))；
其中，gc表示中心点的像素值，δ表示预设的判别阈值，
Figure PCTCN2016104268-appb-000005
在本实施方式中，将中心像素的局部空谱结构使用中心像素的灰度值二值化。若采样点灰度值与中心像素灰度值的差值绝对值大于判别阈值，则该采样点被标记为1，否则标记为0。中心像素的局部空谱结构二值化后称为其二值模式。图3展示了局部空谱结构被二值化的过程。其中，图3的左边展示灰度值为145的中心点像素的局部空谱结构，图3的右边展示了其二值化后对应的二值模式。
在步骤S3中，对于具有相同空间拓扑结构的二值模式，使用这些二值模式中1的数量进行唯一标记，以得到这些二值模式对应的三维局部二值编码，即：
Figure PCTCN2016104268-appb-000006
其中，二值模式的空间拓扑结构通过
Figure PCTCN2016104268-appb-000007
来进行衡量，Γ值相同的二值模式表明它们的空间拓扑结构相同，这里，gi、gj分别表示二值模式中第i、j个采样点，gi与gj相邻且i≠j，i，j∈{0，1，2，3，4，5}。
在本实施方式中，由于不同的二值模式可能对应着相同的局部空谱结构，为了识别这种现象，根据二值模式的空间拓扑结构可将64种二值模式规约为8种。假设Γ表示二值模式的空间拓扑结构，不同的Γ值对应着不同局部空谱结构的二值模式，那么二值模式的空间拓补结构使用如下表达式计算：
Figure PCTCN2016104268-appb-000008
其中，具有不同空间拓扑结构的二值模式中1的个数不同，对于具有相同空间拓扑结构的二值模式，使用这些二值模式中1的数量进行唯一标记，即得到中心像素的邻域的二值模式的三维局部二值模式(即3DLBP)编码。这个过程的数学表达如下：
Figure PCTCN2016104268-appb-000009
在本实施方式中，该编码为三维局部二值模式编码，为了区分各像素周围的局部空谱结构，对其二值模式使用不同的编码进行标记。由于不同的二值模式可能对应着相同的局部空谱结构，为了区分这种现象，根据二值模式的空间拓补结构对二值模式进行编码。二值模式的空间拓补结构指的是1在二值模式对应的三维空间上的分布。6个采样点产生64种不同的二值模式，这64种二值模式根据其空间拓扑结构被规约为8种，这8种模式分别对应着不同的局部空谱结构，依次使用0、1、2、3、4、5、6、7进行标记。图4中展示了0、1、2、3、4、5、6七种不同空间拓扑结构的二值模式示意图，分别包含了1、6、12、8、12、6、1种二值模式，图中实心点代表1，空心点代表0。在这7中模式中，1在正八面体邻域中是紧挨在一起，且不成线，区别在于1的数量不同。
在步骤S4中，在得到每个像素的三维局部二值模式编码后，在像元中每个像素的nxn矩形邻域内统计0、1、2、3、4、5、6、7这8个三维局部二值模式编码的共生频率，来得到像素的直方图特征。
在本实施方式中，直方图特征表示了在像素的nxn矩形邻域内8种不同三维局部二值编码的分布。
在步骤S5中，将像元中每个像素的直方图特征依次进行串联，得到像元对应的三维LBP特征；
在步骤S6中，将得到的三维LBP特征送入分类器进行分类。
在本实施方式中，采用常见的分类器如KNN(k-Nearest Neighbor)、SVM(Support Vector Machine)、SRC(Sparse Representation Classification)均可实现高光谱遥感图像的分类。
本发明提供的一种高光谱遥感图像特征提取和分类方法，将二维LBP扩展到三维LBP，充分利用了高光谱遥感图中空间-光谱的上下文关系，通过引入了松弛阈值判别操作，对噪声有很好的鲁棒性，本发明提出的旋转不变三维LBP模型考虑高光谱遥感图像的本质特征，具备针对性强、操作简单以及计算效率高的优点。
以下将对本发明所提供的一种高光谱遥感图像特征提取和分类***进行详细说明。
请参阅图5，所示为本发明一实施方式中高光谱遥感图像特征提取和分类***10的结构示意图。
在本实施方式中，高光谱遥感图像特征提取和分类***10，主要包括采样模块11、二值化模块12、编码模块13、统计模块14、串联模块15以及分类模块16。
采样模块11，用于将每个像素的正八面体领域的六个顶点作为三维局部二值模式的采样点，并利用这六个采样点的灰度分布T描述像素的局部空谱结构，其中，T≈t(g0，g1，g2，g3，g4，g5)，表示这六个采样点的灰度分布，g0、g1、g2、g3、g4、g5分别表示这六个采样点的像素值。
在本实施方式中，采样模块11中的采样方法如步骤S1中的相关记载所述，在此不重新描述。
二值化模块12，用于将正八面体的六个顶点像素的灰度值与中心点的像素值一一进行比较，如果二者差值的绝对值小于预设的判别阈值，则将对应的顶 点像素标记为1，否则标记为0，以形成中心像素的局部空谱结构的二值模式T1，其中，T1≈t(s(g0-gc，s(g1-gc)，…，s(g5-gc))，gc表示中心点的像素值。
在本实施方式中，二值化模块12中的二值化方法如步骤S2中的相关记载所述，在此不重新描述。
编码模块13，用于对于具有相同空间拓扑结构的二值模式，使用这些二值模式中1的数量进行唯一标记，以得到这些二值模式对应的三维局部二值编码，即：
Figure PCTCN2016104268-appb-000010
其中，二值模式的空间拓扑结构通过
Figure PCTCN2016104268-appb-000011
来进行衡量，Γ值相同的二值模式表明它们的空间拓扑结构相同，这里，gi、gj分别表示二值模式中第i、j个采样点，gi与gj相邻且i≠j，i，j∈{0，1，2，3，4，5}。
在本实施方式中，编码模块13中的编码方法如步骤S3中的相关记载所述，在此不重新描述。
统计模块14，用于在得到每个像素的三维局部二值模式编码后，在像元中每个像素的nxn矩形邻域内统计0、1、2、3、4、5、6、7这8个三维局部二值模式编码的共生频率，来得到像素的直方图特征。
串联模块15，用于将像元中每个像素的直方图特征依次进行串联，得到像元对应的三维局部二值模式特征。
分类模块16，用于将得到的三维局部二值模式特征送入分类器进行分类。
本发明提供的一种高光谱遥感图像特征提取和分类***10，将二维LBP扩展到三维LBP，充分利用了高光谱遥感图中空间-光谱的上下文关系，通过引入了松弛阈值判别操作，对噪声有很好的鲁棒性，本发明提出的旋转不变三维LBP模型考虑高光谱遥感图像的本质特征，具备针对性强、操作简单以及计算效率高的优点。
本发明提取的三维LBP特征比二维LBP特征的判别力更强，以2个不同 分辨率、不同大小的数据集的分类结果进行说明和验证。第一个数据集是由ROSIS-03传感器在意大利帕维亚中心拍摄获得(Pavia Centre，URL：http：//www.ehu.eus/ccwintco/index.php？title＝Hyperspectral_Remote_Sensing_Scenes)，该数据集总共102个波段，每个波段大小为为610x340，空间分辨率1.3米/像素，42776个标记样本，共9类地物。第二个数据集是由美国国家科学基金会资助中心使用机载激光映射于2012年6月23日在休斯敦校园及周边地区拍摄获得(Houston University，URL：http：//www.grss-ieee.org/community/technical-committees/data-fusion/2013-ieee-grss-data-fusion-contest/)。该数据集有144个波段，每个波段图像尺寸为349x1905，空间分辨率为2.5m/像素，15029个标记样本，共15类地物。使用稀疏表示的分类结果表明，三维LBP特征的分类结果远远高于二维LBP，特别是在小样本分类问题上，3个样本时，在Pavia Centre数据集上，二维LBP特征的分类结果为72.75％，而三维LBP特征的分类结果高达92.87％；在Houston University数据集上，二维LBP特征的分类结果为45.98％，而三维LBP特征的分类结果高达68.61％。上述结果表明，三维LBP特征的分类准确率远远高于二维LBP，且在小样本分类问题上效果特别明显。
值得注意的是，上述实施例中，所包括的各个单元只是按照功能逻辑进行划分的，但并不局限于上述的划分，只要能够实现相应的功能即可；另外，各功能单元的具体名称也只是为了便于相互区分，并不用于限制本发明的保护范围。
另外，本领域普通技术人员可以理解实现上述各实施例方法中的全部或部分步骤是可以通过程序来指令相关的硬件来完成，相应的程序可以存储于一计算机可读取存储介质中，所述的存储介质，如ROM/RAM、磁盘或光盘等。
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 (2)

  1. 一种高光谱遥感图像特征提取和分类方法，其特征在于，所述方法包括：
    采样步骤、将每个像素的正八面体领域的六个顶点作为三维局部二值模式的采样点，并利用这六个采样点的灰度分布T描述像素的局部空谱结构，其中，T≈t(g0，g1，g2，g3，g4，g5)，表示这六个采样点的灰度分布，g0、g1、g2、g3、g4、g5分别表示这六个采样点的像素值；
    二值化步骤、将正八面体的六个顶点像素的灰度值与中心点的像素值一一进行比较，如果二者差值的绝对值小于预设的判别阈值，则将对应的顶点像素标记为1，否则标记为0，以形成中心像素的局部空谱结构的二值模式T1，其中，T1≈t(s(g0-gc，s(g1-gc)，…，s(g5-gc))，gc表示中心点的像素值；
    编码步骤、对于具有相同空间拓扑结构的二值模式，使用这些二值模式中1的数量进行唯一标记，以得到这些二值模式对应的三维局部二值编码，即：
    Figure PCTCN2016104268-appb-100001
    其中，二值模式的空间拓扑结构通过
    Figure PCTCN2016104268-appb-100002
    来进行衡量，Γ值相同的二值模式表明它们的空间拓扑结构相同，这里，gi、gj分别表示二值模式中第i、j个采样点，gi与gj相邻且i≠j，i，j∈{0，1，2，3，4，5}；
    统计步骤、在得到每个像素的三维局部二值模式编码后，在像元中每个像素的nxn矩形邻域内统计0、1、2、3、4、5、6、7这8个三维局部二值模式编码的共生频率，来得到像素的直方图特征；
    串联步骤、将像元中每个像素的直方图特征依次进行串联，得到像元对应的三维局部二值模式特征；
    分类步骤、将得到的三维局部二值模式特征送入分类器进行分类。
  2. 一种高光谱遥感图像特征提取和分类***，其特征在于，所述***包括：
    采样模块，用于将每个像素的正八面体领域的六个顶点作为三维局部二值模式的采样点，并利用这六个采样点的灰度分布T描述像素的局部空谱结构，其中，T≈t(g0，g1，g2，g3，g4，g5)，表示这六个采样点的灰度分布，g0、g1、g2、g3、g4、g5分别表示这六个采样点的像素值；
    二值化模块，用于将正八面体的六个顶点像素的灰度值与中心点的像素值一一进行比较，如果二者差值的绝对值小于预设的判别阈值，则将对应的顶点像素标记为1，否则标记为0，以形成中心像素的局部空谱结构的二值模式T1，其中，T1≈t(s(g0-gc，s(g1-gc)，…，s(g5-gc))，gc表示中心点的像素值；
    编码模块，用于对于具有相同空间拓扑结构的二值模式，使用这些二值模式中1的数量进行唯一标记，以得到这些二值模式对应的三维局部二值编码，即：
    Figure PCTCN2016104268-appb-100003
    其中，二值模式的空间拓扑结构通过
    Figure PCTCN2016104268-appb-100004
    来进行衡量，Γ值相同的二值模式表明它们的空间拓扑结构相同，这里，gi、gj分别表示二值模式中第i、j个采样点，gi与gj相邻且i≠j，i，j∈{0，1，2，3，4，5}；
    统计模块，用于在得到每个像素的三维局部二值模式编码后，在像元中每个像素的nxn矩形邻域内统计0、1、2、3、4、5、6、7这8个三维局部二值模式编码的共生频率，来得到像素的直方图特征；
    串联模块，用于将像元中每个像素的直方图特征依次进行串联，得到像元对应的三维局部二值模式特征；
    分类模块，用于将得到的三维局部二值模式特征送入分类器进行分类。
PCT/CN2016/104268 2016-11-01 2016-11-01 一种高光谱遥感图像特征提取和分类方法及其*** WO2018081929A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
PCT/CN2016/104268 WO2018081929A1 (zh) 2016-11-01 2016-11-01 一种高光谱遥感图像特征提取和分类方法及其***
US15/978,189 US10509984B2 (en) 2016-11-01 2018-05-13 Method and system for extracting and classifying features of hyperspectral remote sensing image

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2016/104268 WO2018081929A1 (zh) 2016-11-01 2016-11-01 一种高光谱遥感图像特征提取和分类方法及其***

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US15/978,189 Continuation US10509984B2 (en) 2016-11-01 2018-05-13 Method and system for extracting and classifying features of hyperspectral remote sensing image

Publications (1)

Publication Number Publication Date
WO2018081929A1 true WO2018081929A1 (zh) 2018-05-11

Family

ID=62075392

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2016/104268 WO2018081929A1 (zh) 2016-11-01 2016-11-01 一种高光谱遥感图像特征提取和分类方法及其***

Country Status (2)

Country Link
US (1) US10509984B2 (zh)
WO (1) WO2018081929A1 (zh)

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109740504A (zh) * 2018-12-28 2019-05-10 四创科技有限公司 一种基于遥感影像提取海域资源的方法
CN110232317A (zh) * 2019-05-05 2019-09-13 五邑大学 基于超像素分割及两阶段分类策略的超光谱图像分类方法
CN110348458A (zh) * 2019-06-27 2019-10-18 河南科技大学 一种基于局部方向极值模式的图像纹理分类方法
CN110363053A (zh) * 2018-08-09 2019-10-22 中国人民解放军战略支援部队信息工程大学 一种遥感影像居民地提取方法及装置
CN111061897A (zh) * 2020-01-15 2020-04-24 国网江西省电力有限公司电力科学研究院 一种适于电网高分辨遥感图像多尺度自动截取与分类方法
CN111126256A (zh) * 2019-12-23 2020-05-08 武汉大学 一种基于自适应空谱多尺度网络的高光谱图像分类方法
CN111612099A (zh) * 2020-06-03 2020-09-01 江苏科技大学 基于局部排序差值细化模式的纹理图像分类方法及***
CN112070098A (zh) * 2020-08-20 2020-12-11 西安理工大学 基于频率调整模型的高光谱图像显著目标检测方法
CN112364876A (zh) * 2020-11-25 2021-02-12 北京紫光青藤微***有限公司 一种高效的条码二值化方法及***
CN113762389A (zh) * 2021-09-08 2021-12-07 大连海事大学 一种基于动态诊断的高光谱影像小样本分类方法
CN114140644A (zh) * 2021-11-17 2022-03-04 武汉理工大学 一种结合mrlbp和gnp的纹理分类方法
CN115346009A (zh) * 2022-05-18 2022-11-15 上海航遥信息技术有限公司 基于高光谱数据和倾斜三维数据的地理实体语义建模方法
CN116582195A (zh) * 2023-06-12 2023-08-11 浙江瑞通电子科技有限公司 一种基于人工智能的无人机信号频谱识别算法

Families Citing this family (20)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109658380A (zh) * 2018-11-09 2019-04-19 广西壮族自治区遥感信息测绘院 基于前期林地矢量数据的林地变化检测方法
CN109583380B (zh) * 2018-11-30 2020-01-17 广东工业大学 一种基于注意力约束非负矩阵分解的高光谱分类方法
CN109784402A (zh) * 2019-01-15 2019-05-21 中国第一汽车股份有限公司 基于多层次特征融合的快速无人车驾驶场景分割方法
CN110458192B (zh) * 2019-07-05 2022-06-14 中国地质大学（武汉） 基于视觉显著性的高光谱遥感图像分类方法及***
CN110866493B (zh) * 2019-11-14 2022-03-22 生态环境部南京环境科学研究所 基于遥感影像的规模化生猪养殖场识别方法
CN111476287A (zh) * 2020-04-02 2020-07-31 中国人民解放军战略支援部队信息工程大学 一种高光谱影像小样本分类方法及装置
CN111898423B (zh) * 2020-06-19 2022-12-23 北京理工大学 一种基于形态学的多源遥感图像地物精细分类方法
CN111881953B (zh) * 2020-07-14 2022-04-22 安徽大学 基于局部二值模式和knn分类器的遥感高光谱图像分类方法
CN111914922B (zh) * 2020-07-27 2022-07-29 河海大学 一种基于局部卷积和空洞卷积的高光谱图像分类方法
CN112287978B (zh) * 2020-10-07 2022-04-15 武汉大学 一种基于自注意力上下文网络的高光谱遥感图像分类方法
CN112257603B (zh) * 2020-10-23 2022-06-17 深圳大学 高光谱图像分类方法及相关设备
CN112580431B (zh) * 2020-11-20 2022-06-24 北京航空航天大学 一种适用于星上在轨处理的高带宽遥感图像目标提取方法
CN112990313B (zh) * 2021-03-16 2023-04-07 中国科学院长春光学精密机械与物理研究所 高光谱图像异常检测方法、装置、计算机设备及存储介质
CN112926532B (zh) * 2021-04-01 2024-05-10 深圳前海微众银行股份有限公司 信息处理方法、装置、设备、存储介质及计算机程序产品
CN113435265B (zh) * 2021-06-09 2024-07-09 深圳大学 高光谱图像分类方法、装置、电子设备及存储介质
CN113436289B (zh) * 2021-08-04 2023-08-22 成都工业学院 一种基于方向向量相关系数的图像编码及分类方法
CN113723255B (zh) * 2021-08-24 2023-09-01 中国地质大学（武汉） 一种高光谱影像分类方法和存储介质
CN115631372B (zh) * 2022-10-18 2023-06-09 菏泽市土地储备中心 基于土壤遥感数据的土地信息分类管理方法
CN116863327B (zh) * 2023-06-05 2023-12-15 中国石油大学(华东) 一种基于双域分类器协同对抗的跨域小样本分类方法
CN117975269A (zh) * 2024-01-31 2024-05-03 重庆千信新能源有限公司 通过机器视觉进行光伏电力设备故障监督方法

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP2187339A1 (en) * 2008-11-12 2010-05-19 Fundación Robotiker Method for integrating spectral and spatial features for classifying materials
CN103559500A (zh) * 2013-10-15 2014-02-05 北京航空航天大学 一种基于光谱与纹理特征的多光谱遥感图像地物分类方法
US8649607B1 (en) * 2010-01-25 2014-02-11 Stc.Unm Spectral ratio contrast for edge detection in spectral images
CN105608433A (zh) * 2015-12-23 2016-05-25 北京化工大学 一种基于核协同表达的高光谱图像分类方法
CN105654098A (zh) * 2016-03-23 2016-06-08 深圳大学 一种高光谱遥感图像的分类方法及其***
CN105740799A (zh) * 2016-01-27 2016-07-06 深圳大学 基于三维Gabor特征选择的高光谱遥感图像分类方法及***
CN106485238A (zh) * 2016-11-01 2017-03-08 深圳大学 一种高光谱遥感图像特征提取和分类方法及其***

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP2187339A1 (en) * 2008-11-12 2010-05-19 Fundación Robotiker Method for integrating spectral and spatial features for classifying materials
US8649607B1 (en) * 2010-01-25 2014-02-11 Stc.Unm Spectral ratio contrast for edge detection in spectral images
CN103559500A (zh) * 2013-10-15 2014-02-05 北京航空航天大学 一种基于光谱与纹理特征的多光谱遥感图像地物分类方法
CN105608433A (zh) * 2015-12-23 2016-05-25 北京化工大学 一种基于核协同表达的高光谱图像分类方法
CN105740799A (zh) * 2016-01-27 2016-07-06 深圳大学 基于三维Gabor特征选择的高光谱遥感图像分类方法及***
CN105654098A (zh) * 2016-03-23 2016-06-08 深圳大学 一种高光谱遥感图像的分类方法及其***
CN106485238A (zh) * 2016-11-01 2017-03-08 深圳大学 一种高光谱遥感图像特征提取和分类方法及其***

Cited By (24)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110363053A (zh) * 2018-08-09 2019-10-22 中国人民解放军战略支援部队信息工程大学 一种遥感影像居民地提取方法及装置
CN110363053B (zh) * 2018-08-09 2021-09-14 中国人民解放军战略支援部队信息工程大学 一种遥感影像居民地提取方法及装置
CN109740504A (zh) * 2018-12-28 2019-05-10 四创科技有限公司 一种基于遥感影像提取海域资源的方法
CN109740504B (zh) * 2018-12-28 2022-10-14 四创科技有限公司 一种基于遥感影像提取海域资源的方法
CN110232317A (zh) * 2019-05-05 2019-09-13 五邑大学 基于超像素分割及两阶段分类策略的超光谱图像分类方法
CN110232317B (zh) * 2019-05-05 2023-01-03 五邑大学 基于超像素分割及两阶段分类策略的超光谱图像分类方法、装置及介质
CN110348458B (zh) * 2019-06-27 2022-09-30 河南科技大学 一种基于局部方向极值模式的图像纹理分类方法
CN110348458A (zh) * 2019-06-27 2019-10-18 河南科技大学 一种基于局部方向极值模式的图像纹理分类方法
CN111126256A (zh) * 2019-12-23 2020-05-08 武汉大学 一种基于自适应空谱多尺度网络的高光谱图像分类方法
CN111126256B (zh) * 2019-12-23 2022-02-15 武汉大学 一种基于自适应空谱多尺度网络的高光谱图像分类方法
CN111061897A (zh) * 2020-01-15 2020-04-24 国网江西省电力有限公司电力科学研究院 一种适于电网高分辨遥感图像多尺度自动截取与分类方法
CN111061897B (zh) * 2020-01-15 2023-06-16 国网江西省电力有限公司电力科学研究院 一种适于电网高分辨遥感图像多尺度自动截取与分类方法
CN111612099A (zh) * 2020-06-03 2020-09-01 江苏科技大学 基于局部排序差值细化模式的纹理图像分类方法及***
CN111612099B (zh) * 2020-06-03 2022-11-29 江苏科技大学 基于局部排序差值细化模式的纹理图像分类方法及***
CN112070098A (zh) * 2020-08-20 2020-12-11 西安理工大学 基于频率调整模型的高光谱图像显著目标检测方法
CN112070098B (zh) * 2020-08-20 2024-02-09 西安理工大学 基于频率调整模型的高光谱图像显著目标检测方法
CN112364876A (zh) * 2020-11-25 2021-02-12 北京紫光青藤微***有限公司 一种高效的条码二值化方法及***
CN112364876B (zh) * 2020-11-25 2024-05-14 北京紫光青藤微***有限公司 一种高效的条码二值化方法及***
CN113762389A (zh) * 2021-09-08 2021-12-07 大连海事大学 一种基于动态诊断的高光谱影像小样本分类方法
CN113762389B (zh) * 2021-09-08 2024-04-12 大连海事大学 一种基于动态诊断的高光谱影像小样本分类方法
CN114140644A (zh) * 2021-11-17 2022-03-04 武汉理工大学 一种结合mrlbp和gnp的纹理分类方法
CN115346009A (zh) * 2022-05-18 2022-11-15 上海航遥信息技术有限公司 基于高光谱数据和倾斜三维数据的地理实体语义建模方法
CN116582195A (zh) * 2023-06-12 2023-08-11 浙江瑞通电子科技有限公司 一种基于人工智能的无人机信号频谱识别算法
CN116582195B (zh) * 2023-06-12 2023-12-26 浙江瑞通电子科技有限公司 一种基于人工智能的无人机信号频谱识别方法

Also Published As

Publication number Publication date
US10509984B2 (en) 2019-12-17
US20180260657A1 (en) 2018-09-13

Similar Documents

Publication Publication Date Title
WO2018081929A1 (zh) 一种高光谱遥感图像特征提取和分类方法及其***
Gao et al. Automatic change detection in synthetic aperture radar images based on PCANet
Ai et al. Automatic pixel-level pavement crack detection using information of multi-scale neighborhoods
Lu et al. Joint dictionary learning for multispectral change detection
İlsever et al. Two-dimensional change detection methods: remote sensing applications
Sirmacek et al. Urban-area and building detection using SIFT keypoints and graph theory
CN102663391B (zh) 一种图像的多特征提取与融合方法及***
Sirmacek et al. Urban area detection using local feature points and spatial voting
He et al. A method of green litchi recognition in natural environment based on improved LDA classifier
Tao et al. Unsupervised detection of built-up areas from multiple high-resolution remote sensing images
CN106485238B (zh) 一种高光谱遥感图像特征提取和分类方法及其***
CN112734734A (zh) 一种基于改进残差网络的铁路隧道裂缝检测方法
CN103679192A (zh) 基于协方差特征的图像场景类型判别方法
CN103226616A (zh) 一种图像内容检索***及其图像内容稀疏学习方法
CN108073940B (zh) 一种非结构化环境中的3d目标实例物体检测的方法
CN103473545A (zh) 一种基于多特征的文本图像相似度度量方法
Su et al. Hyperspectral image classification based on volumetric texture and dimensionality reduction
Forczmański et al. Stamps detection and classification using simple features ensemble
Zhu et al. Multiple disjoint dictionaries for representation of histopathology images
Pan et al. Improving scene text detection by scale-adaptive segmentation and weighted CRF verification
Suruliandi et al. Local texture patterns-a univariate texture model for classification of images
Fernández-Carrobles et al. Bagging tree classifier and texture features for tumor identification in histological images
Liu et al. Automatic grape bunch detection in vineyards for precise yield estimation
Auleria et al. A review on KN earest neighbour based classification for object recognition
Yan et al. A modified three-dimensional gray-level co-occurrence matrix for image classification with digital surface model

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: 16920697

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 04/09/2019)

122 Ep: pct application non-entry in european phase

Ref document number: 16920697

Country of ref document: EP

Kind code of ref document: A1