WO2017166156A1 - 一种高光谱图像的分类方法及其系统 - Google Patents

一种高光谱图像的分类方法及其系统 Download PDF

Info

Publication number
WO2017166156A1
WO2017166156A1 PCT/CN2016/077976 CN2016077976W WO2017166156A1 WO 2017166156 A1 WO2017166156 A1 WO 2017166156A1 CN 2016077976 W CN2016077976 W CN 2016077976W WO 2017166156 A1 WO2017166156 A1 WO 2017166156A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
codebook
spectral domain
hyperspectral image
interest
Prior art date
Application number
PCT/CN2016/077976
Other languages
English (en)
French (fr)
Inventor
李岩山
石伟
夏荣杰
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
Priority to PCT/CN2016/077976 priority Critical patent/WO2017166156A1/zh
Publication of WO2017166156A1 publication Critical patent/WO2017166156A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G06V10/765—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects using rules for classification or partitioning the feature space
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition

Definitions

  • the present invention relates to the field of computer vision, and in particular to a method for classifying hyperspectral images and a system thereof.
  • Image classification is an important research direction in the field of computer vision. It is also a typical pattern recognition problem. It is an image processing method that separates different categories of objects according to different characteristics reflected in image information. Hyperspectral image classification is one of the hotspot technologies in hyperspectral applications, and it is also a prerequisite for realizing various practical applications, such as target recognition and tracking, geological exploration and environmental monitoring.
  • the original hyperspectral image classification method is mainly for hyperspectral remote sensing images with low spatial resolution, while the existing hyperspectral images have high spatial resolution.
  • image classification is performed by pixel-by-pixel spectral curve, on the one hand, the efficiency is low, and on the other hand, the recognition rate is easily reduced by the influence of "homologous” and "homogeneous foreign matter".
  • the spectral information of the pixel and the spatial information of the image are used to classify the hyperspectral image, the spatial and spectral domains are separated for consideration.
  • the algorithm has high complexity, long time and low recognition rate, but the existing algorithm is difficult to apply to high resolution. Spectral image classification.
  • an object of the present invention is to provide a method for classifying hyperspectral images and a system thereof, which aim to solve the problem of low efficiency and low accuracy of recognition of hyperspectral images in the prior art.
  • the invention provides a classification method for hyperspectral images, and the classification method comprises:
  • the hyperspectral image is classified by the codebook and the support vector machine classifier, and the classification result is output.
  • the step of acquiring the codebook by using the codebook learning method specifically includes:
  • the m spectral domain interest points of all the image blocks in each hyperspectral image are extracted in parallel by m threads, and all the empty spectral domain interest points are characterized to obtain a description vector of each empty spectral domain interest point;
  • the description vectors of all the acquired empty spectral domain points of interest are clustered by using a preset clustering method to obtain a codebook.
  • the step of training the support vector machine classifier by the training method specifically includes:
  • the m spectral domain interest points of all the image blocks in each hyperspectral image are extracted in parallel by m threads, and all the empty spectral domain interest points are characterized to obtain a description vector of each empty spectral domain interest point;
  • the support vector machine classifier is trained using the obtained description vector.
  • the step of performing classification of the hyperspectral image by using the codebook and the support vector machine classifier, and outputting the classification result specifically includes:
  • the hyperspectral image to be classified is subjected to image segmentation
  • the m spectral domain interest points of all the image blocks in the hyperspectral image to be classified are extracted in parallel by m threads, and all the empty spectral domain interest points are characterized to obtain a description vector of each empty spectral domain interest point;
  • the obtained description vector is classified by the trained support vector machine classifier, and the classification result is output.
  • the present invention also provides a classification system for hyperspectral images, the classification system comprising:
  • a learning module configured to obtain a codebook by using a codebook learning method
  • a training module for training a support vector machine classifier by training
  • a classification module configured to perform classification of the hyperspectral image by using the codebook and the support vector machine classifier, and output the classification result.
  • the learning module comprises:
  • a first blocking sub-module for performing image segmentation on a hyperspectral image of a learning codebook
  • a first description sub-module for extracting, in m threads, the spatial spectral domain points of all image blocks in each hyperspectral image in parallel, and characterizing all the empty spectral domain points of interest to obtain each of the null spectral domains a description vector of the point of interest;
  • the first clustering sub-module is configured to cluster the description vectors of all the acquired empty spectral domain points of interest by using a preset clustering method to obtain a codebook.
  • the training module comprises:
  • a second block sub-module for performing image segmentation on a hyperspectral image of a support vector machine classifier to be trained
  • a second description sub-module for extracting, in m threads, the spatial spectral domain points of all image blocks in each hyperspectral image in parallel, and characterizing all the empty spectral domain points of interest to obtain each of the null spectral domains a description vector of the point of interest;
  • a first encoding sub-module configured to encode, according to the obtained codebook, a set of spatial spectral domain features of each hyperspectral image by using a preset model to obtain a description vector of the image;
  • the first training submodule is configured to train the support vector machine classifier by using the obtained description vector.
  • the classification module comprises:
  • a third block sub-module configured to perform image segmentation on the hyperspectral image to be classified
  • a third description sub-module for extracting, in m threads, the spatial spectral domain points of all image blocks in the hyperspectral image to be classified in parallel, and characterizing all the empty spectral domain points of interest to obtain each of the spatial spectra a description vector of the domain interest point;
  • a second coding submodule configured to use the preset model to be classified according to the obtained codebook
  • the set of null spectral domain points of interest in the hyperspectral image is encoded to obtain a description vector of the image
  • the first classification sub-module is configured to classify the obtained description vector by using the trained support vector machine classifier, and output the classification result.
  • the technical solution provided by the invention proposes a high spatial resolution hyperspectral image classification method based on the spatial spectral domain interest points, and the spatial spectral domain interest points can effectively utilize the spatial and spectral domain information for image classification.
  • the invention extracts the optical spectral domain interest points of the hyperspectral image through the space spectrum domain interest point detection algorithm, and divides the hyperspectral image into small blocks for image learning, training and classification through image segmentation, and proposes a new hyperspectral image classification.
  • the framework can effectively represent and classify hyperspectral images, which significantly improves the accuracy and efficiency of high spatial resolution hyperspectral image classification.
  • FIG. 1 is a flow chart of a method for classifying hyperspectral images according to an embodiment of the present invention
  • step S11 shown in FIG. 1 according to an embodiment of the present invention
  • step S12 shown in FIG. 1 according to an embodiment of the present invention
  • step S13 shown in FIG. 1 according to an embodiment of the present invention
  • FIG. 5 is a chart showing classification results of high spatial resolution hyperspectral images in an embodiment of the present invention.
  • FIG. 6 is a chart showing a result of classification of a pseudo color image of a hyperspectral image according to an embodiment of the present invention
  • FIG. 7 is an overall frame diagram of a method for classifying hyperspectral images according to an embodiment of the present invention.
  • FIG. 8 is a schematic diagram showing the internal structure of a classification system 10 for a hyperspectral image according to an embodiment of the present invention.
  • FIG. 9 is a schematic diagram showing the internal structure of the learning module 11 shown in FIG. 8 according to an embodiment of the present invention.
  • FIG. 10 is a schematic diagram showing the internal structure of the training module 12 shown in FIG. 8 according to an embodiment of the present invention.
  • FIG. 11 is a schematic diagram showing the internal structure of the classification module 13 shown in FIG. 8 according to an embodiment of the present invention.
  • a specific embodiment of the present invention provides a method for classifying hyperspectral images, and the method mainly includes the following steps:
  • the invention proposes a high spatial resolution hyperspectral image classification method based on the spatial spectral domain interest points, and the spatial spectral domain interest points can effectively utilize the spatial and spectral domain information for image classification.
  • the invention extracts the optical spectral domain interest points of the hyperspectral image through the space spectrum domain interest point detection algorithm, and divides the hyperspectral image into small blocks for image learning, training and classification through image segmentation, and proposes a new hyperspectral image classification.
  • the framework can effectively represent and classify hyperspectral images, which significantly improves the accuracy and efficiency of high spatial resolution hyperspectral image classification.
  • FIG. 1 is a flowchart of a method for classifying hyperspectral images according to an embodiment of the present invention.
  • step S11 the codebook is acquired by the codebook learning method.
  • the step S12 of acquiring the codebook by the codebook learning method specifically includes S111-S113, as shown in FIG. 2.
  • FIG. 2 is a detailed flowchart of step S11 shown in FIG. 1 according to an embodiment of the present invention.
  • step S111 the hyperspectral image of the codebook to be learned is subjected to image segmentation.
  • the image blocks are respectively represented, where n i represents the number of image blocks.
  • step S112 the m spectral domain interest points of all image blocks in each hyperspectral image are extracted in parallel by m threads, and all the empty spectral domain interest points are characterized to obtain each of the empty spectral domain interest points. Description vector.
  • the null spectral domain interest point is a feature point extracted on the spatial domain composed of the spatial and spectral domains of the hyperspectral image, reflecting the characteristics in the spatial domain and the spectral domain, and the use of the spatial spectral domain interest point can be effective. Image classification using spatial and spectral domain information.
  • the m spectral domain interest points of all the image blocks in each hyperspectral image in step S111 are extracted in parallel by m threads, and the 3D SIFT descriptor is used to describe all the empty spectral domain interest points.
  • Each of the null spectral domain points of interest is described as a 640-dimensional description vector.
  • step S113 the description vectors of all the acquired empty spectral domain points of interest are clustered by using a preset clustering method to obtain a codebook.
  • the preset clustering method is a K-means clustering method, and all the description vectors obtained in step S112 are clustered by using the K-means clustering method, and k clustering centers obtained after clustering are clustered. , called code words or visual words, expressed by v i .
  • step S12 the support vector machine classifier is trained by the training method.
  • the step S12 of training the Support Vector Machine (SVM) classifier by the training method specifically includes S121-S124, as shown in FIG.
  • FIG. 3 is a detailed flowchart of step S12 shown in FIG. 1 according to an embodiment of the present invention.
  • step S121 the hyperspectral image of the support vector machine classifier to be trained is subjected to image segmentation.
  • the image blocks are respectively represented, where n j represents the number of image blocks.
  • step S122 the m spectral domain interest points of all the image blocks in each hyperspectral image are extracted in parallel by m threads, and all the empty spectral domain interest points are characterized to obtain each of the empty spectral domain interest points. Description vector.
  • the m spectral domain interest points of all the image blocks in each hyperspectral image in step S121 are extracted in parallel by m threads, and the 3D SIFT descriptors are used to perform all the spatial spectral domain interest points.
  • the description describes each of the null spectral domain points of interest as a 640-dimensional description vector.
  • step S123 based on the acquired codebook, the spatial spectral feature set of each hyperspectral image is encoded by using a preset model to obtain a description vector of the image.
  • the preset model is a BoW model, and based on the codebook V obtained in the learning step shown in FIG. 2, the high spatial resolution hyperspectral image T j of each SVM classifier to be trained is utilized by the BoW model.
  • step S124 the support vector machine classifier is trained using the acquired description vector.
  • the SVM classifier is trained using the description vector set S of T obtained in step S123.
  • step S13 the hyperspectral image is classified by the codebook and the support vector machine classifier, and the classification result is output.
  • the step S13 of performing classification of the hyperspectral image by using the codebook and the support vector machine classifier, and outputting the classification result specifically includes S131-S134, as shown in FIG.
  • FIG. 4 is a detailed flowchart of step S13 shown in FIG. 1 according to an embodiment of the present invention.
  • step S131 the hyperspectral image to be classified is subjected to image segmentation.
  • step S132 the m spectral domain interest points of all the image blocks in the hyperspectral image to be classified are extracted in parallel by m threads, and all the empty spectral domain interest points are characterized to obtain each empty spectral domain interest point. Description vector.
  • the null spectral domain points of the image blocks corresponding to all the images I test in the hyperspectral image to be classified in step S131 are also extracted in parallel by m threads, and each spatial domain is described by 3D SIFT descriptors.
  • the points of interest are described, and each of the empty spectral domain points of interest is described as a 640-dimensional description vector, and the spatial spectral domain points extracted from the N hyperspectral image blocks are merged to form an empty spectral domain feature set of the hyperspectral image I test .
  • F test ⁇ f 1 ,f 2 ,...,f M ⁇ .
  • step S133 based on the acquired codebook, the set of the null spectral domain points of interest in the hyperspectral image to be classified is encoded by the preset model to obtain a description vector of the image.
  • the preset model is a BoW model, and based on the codebook V obtained in the learning step shown in FIG. 2, the spatial spectral domain interest of the image I test in the high spatial resolution hyperspectral image to be classified by the BoW model is used.
  • the point set F test is encoded to obtain a description vector D of the image I test .
  • step S134 the obtained description vector is classified by the trained support vector machine classifier, and the classification result is output.
  • the description vector D obtained in step S133 is classified by the SVM classifier trained by the training step shown in FIG. 3, and the classification result is output.
  • step S11 whether it is step S11, step S12, or step S13, it is for a large number of images, including a plurality of consecutive images, not just for one image.
  • a high spatial resolution hyperspectral image classification method based on the spatial spectral domain interest point is applied to classify a set of high spatial resolution hyperspectral images.
  • the codebook to be learned is learned.
  • Each image is extracted from 3500 spatial spectral points of interest for clustering, and the number of vocabulary obtained after clustering is set to 2500.
  • the classification experiment 10 times we obtained the classification results as shown in Fig. 5. Calculating the mean and standard deviation of the classification results of these 10 experiments, it can be seen that the accuracy of this group of hyperspectral image classification is 93.22 ⁇ 0.51 (%).
  • the accuracy of the pseudo-color image classification of the hyperspectral image shown in FIG. 6 is much lower than the accuracy of the hyperspectral image classification, which fully demonstrates the high space based on the optical spectral domain interest point proposed by the present invention. Resolution method and system validity and accuracy of resolution hyperspectral imagery.
  • FIG. 7 is an overall frame diagram of a method for classifying hyperspectral images according to an embodiment of the present invention.
  • the method for classifying hyperspectral images provided by the invention can effectively utilize the information of the spatial domain and the spectral domain to perform image classification, and can effectively represent and classify the hyperspectral image, thereby significantly improving the high spatial resolution. Accuracy and efficiency of spectral image classification.
  • the classification system 10 of the hyperspectral image mainly includes a learning module 11, a training module 12, and a classification module 13.
  • the classification system 10 of the hyperspectral image can effectively utilize the information of the spatial domain and the spectral domain to perform image classification, and can effectively represent and classify the hyperspectral image, and significantly improve the high spatial resolution hyperspectral image classification. Accuracy and efficiency.
  • the learning module 11 is configured to obtain a codebook by using a codebook learning method.
  • the learning module 11 specifically includes a first blocking sub-module 111, a first descriptive sub-module 112, and a first clustering sub-module 113, as shown in FIG.
  • FIG. 9 is a schematic diagram showing the internal structure of the learning module 11 shown in FIG. 8 according to an embodiment of the present invention.
  • the first blocking sub-module 111 is configured to perform image segmentation on the hyperspectral image of the codebook to be learned.
  • the specific image blocking method of the first blocking sub-module 111 is described in detail in the foregoing step S111, and the description thereof will not be repeated here.
  • the first description sub-module 112 is configured to extract, in m threads, the spatial spectral domain points of all the image blocks in each hyperspectral image in parallel, and characterize all the empty spectral domain points of interest to obtain each of the spatial spectra. Description vector of the domain interest point.
  • the first clustering sub-module 113 is configured to cluster the description vectors of all the acquired empty spectral domain points of interest by using a preset clustering method to obtain a codebook.
  • the specific clustering method of the first clustering sub-module 113 is as described in the foregoing step S113.
  • the relevant records in the description are not repeated here.
  • the training module 12 is configured to train the support vector machine classifier by training.
  • the training module 12 specifically includes a second blocking sub-module 121, a second description sub-module 122, a first encoding sub-module 123, and a first training sub-module 124, as shown in FIG.
  • FIG. 10 is a schematic diagram showing the internal structure of the training module 12 shown in FIG. 8 according to an embodiment of the present invention.
  • the second blocking sub-module 121 performs image segmentation on the hyperspectral image of the support vector machine classifier to be trained.
  • the specific image blocking method of the second blocking sub-module 121 is described in detail in the foregoing step S121, and is not described repeatedly.
  • the second description sub-module 122 is configured to extract, in m threads, the spatial spectral domain points of all the image blocks in each hyperspectral image in parallel, and characterize all the empty spectral domain points of interest to obtain each of the spatial spectra. Description vector of the domain interest point.
  • the first encoding sub-module 123 is configured to encode the spatial spectral domain feature set of each hyperspectral image by using a preset model to obtain a description vector of the image based on the acquired codebook.
  • the specific coding method of the first coding sub-module 123 is described in detail in the foregoing step S123, and is not described repeatedly.
  • the first training sub-module 124 is configured to train the support vector machine classifier by using the obtained description vector.
  • the specific training method of the first training sub-module 124 is described in detail in the foregoing step S124, and the description is not repeated here.
  • the classification module 13 is configured to perform classification of the hyperspectral image by using the codebook and the support vector machine classifier, and output the classification result.
  • the classification module 13 specifically includes a third sub-module 131 and a third descriptor.
  • the module 132, the second encoding sub-module 133, and the first sorting sub-module 134 are as shown in FIG.
  • FIG. 11 is a schematic diagram showing the internal structure of the classification module 13 shown in FIG. 8 according to an embodiment of the present invention.
  • the third blocking sub-module 131 is configured to perform image segmentation on the hyperspectral image to be classified.
  • the specific image blocking method of the third blocking sub-module 131 is described in detail in the foregoing step S131, and the description is not repeated here.
  • the third description sub-module 132 is configured to extract, in m threads, the spatial spectral domain interest points of all the image blocks in the hyperspectral image to be classified, and characterize all the empty spectral domain interest points to obtain each empty Description vector of the spectral domain interest point.
  • the second encoding sub-module 133 is configured to encode, according to the obtained codebook, a set of null spectral domain points of interest in the hyperspectral image to be classified by using a preset model to obtain a description vector of the image.
  • the first classification sub-module 134 is configured to classify the obtained description vector by using the trained support vector machine classifier, and output the classification result.
  • the specific classification method of the first classification sub-module 134 is described in detail in the foregoing step S134, and the repeated description is not repeated here.
  • the classification system 10 for hyperspectral image provided by the invention can effectively utilize the information of the spatial domain and the spectral domain to perform image classification by using the spatial spectral interest points, can effectively represent and classify the hyperspectral image, and significantly improve the high spatial resolution. Rate the accuracy and efficiency of hyperspectral image classification.
  • 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)
  • Theoretical Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Health & Medical Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Databases & Information Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Software Systems (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

一种高光谱图像的分类方法，包括：通过码本学习方式获取码本（S11）；通过训练方式训练支持向量机分类器（S12）；利用所述码本和所述支持向量机分类器进行高光谱图像的分类，并输出分类结果（S13）。还提供一种高光谱图像的分类系统，通过空谱域兴趣点检测算法提取高光谱图像的空谱域兴趣点，通过图像分块将高光谱图像分割成小块进行学习、训练和分类，提出了一种新的高光谱图像分类框架，能有效对高光谱图像进行表示及分类，提高了高空间分辨率高光谱图像分类的准确性和效率。

Description

一种高光谱图像的分类方法及其系统 技术领域
本发明涉及计算机视觉领域，尤其涉及一种高光谱图像的分类方法及其系统。
背景技术
图像分类是计算机视觉领域的重要研究方向，也是一个典型的模式识别问题，它是根据图像信息中反映的不同特征，把不同类别的目标区分开来的图像处理方法。高光谱图像分类是高光谱应用中的热点技术之一，也是实现各种实际应用的前提，如目标识别及跟踪、地质勘探以及环境监测等方面。
原有的高光谱图像分类方法主要针对空间分辨率低的高光谱遥感图像，而现有的高光谱图像具有很高的空间分辨率。采用逐个像素的光谱曲线进行图像分类时，一方面效率低，另一方面识别率容易受“同物异谱”和“同谱异物”的影响而降低。采用像素的光谱信息以及图像的空间信息进行高光谱图像分类时，将空域和光谱域分开来考虑，算法复杂性高，耗时长，识别率低，但是现有算法难以适用于高分辨率的高光谱图像分类上。
因此，如何实现对高空间分辨率高光谱图像进行快速准确的分类一直是业界亟待实现的目标。
发明内容
有鉴于此，本发明的目的在于提供一种高光谱图像的分类方法及其系统，旨在解决现有技术中对高光谱图像识别效率低且准确度不高的问题。
本发明提出一种高光谱图像的分类方法，所述分类方法包括：
通过码本学习方式获取码本；
通过训练方式训练支持向量机分类器；
利用所述码本和所述支持向量机分类器进行高光谱图像的分类，并输出分类结果。
优选的，所述通过码本学习方式获取码本的步骤具体包括：
对待学习码本的高光谱图像进行图像分块；
用m个线程并行地提取每幅高光谱图像中所有图像块的空谱域兴趣点，并对所有空谱域兴趣点进行特征描述，以获取每一个空谱域兴趣点的描述向量；
利用预设聚类方法对获取的所有空谱域兴趣点的描述向量进行聚类以获取码本。
优选的，所述通过训练方式训练支持向量机分类器的步骤具体包括：
对待训练的支持向量机分类器的高光谱图像进行图像分块；
用m个线程并行地提取每幅高光谱图像中所有图像块的空谱域兴趣点，并对所有空谱域兴趣点进行特征描述，以获取每一个空谱域兴趣点的描述向量；
基于获取到的所述码本，利用预设模型对每幅高光谱图像的空谱域特征集进行编码，以获取该图像的描述矢量；
利用获取到的所述描述矢量训练支持向量机分类器。
优选的，所述利用所述码本和所述支持向量机分类器进行高光谱图像的分类，并输出分类结果的步骤具体包括：
将待分类的高光谱图像进行图像分块；
用m个线程并行地提取待分类的高光谱图像中所有图像块的空谱域兴趣点，并对所有空谱域兴趣点进行特征描述，以获取每一个空谱域兴趣点的描述向量；
基于获取到的所述码本，利用预设模型对待分类的高光谱图像中的空谱域兴趣点集进行编码，以获取该图像的描述矢量；
利用训练好的所述支持向量机分类器对获取到的所述描述矢量进行分类，并输出分类结果。
另一方面，本发明还提供一种高光谱图像的分类系统，所述分类系统包括：
学习模块，用于通过码本学习方式获取码本；
训练模块，用于通过训练方式训练支持向量机分类器；
分类模块，用于利用所述码本和所述支持向量机分类器进行高光谱图像的分类，并输出分类结果。
优选的，所述学习模块包括：
第一分块子模块，用于对待学习码本的高光谱图像进行图像分块；
第一描述子模块，用于用m个线程并行地提取每幅高光谱图像中所有图像块的空谱域兴趣点，并对所有空谱域兴趣点进行特征描述，以获取每一个空谱域兴趣点的描述向量；
第一聚类子模块，用于利用预设聚类方法对获取的所有空谱域兴趣点的描述向量进行聚类以获取码本。
优选的，所述训练模块包括：
第二分块子模块，用于对待训练的支持向量机分类器的高光谱图像进行图像分块；
第二描述子模块，用于用m个线程并行地提取每幅高光谱图像中所有图像块的空谱域兴趣点，并对所有空谱域兴趣点进行特征描述，以获取每一个空谱域兴趣点的描述向量；
第一编码子模块，用于基于获取到的所述码本，利用预设模型对每幅高光谱图像的空谱域特征集进行编码，以获取该图像的描述矢量；
第一训练子模块，用于利用获取到的所述描述矢量训练支持向量机分类器。
优选的，所述分类模块包括：
第三分块子模块，用于将待分类的高光谱图像进行图像分块；
第三描述子模块，用于用m个线程并行地提取待分类的高光谱图像中所有图像块的空谱域兴趣点，并对所有空谱域兴趣点进行特征描述，以获取每一个空谱域兴趣点的描述向量；
第二编码子模块，用于基于获取到的所述码本，利用预设模型对待分类的 高光谱图像中的空谱域兴趣点集进行编码，以获取该图像的描述矢量；
第一分类子模块，用于利用训练好的所述支持向量机分类器对获取到的所述描述矢量进行分类，并输出分类结果。
本发明提供的技术方案提出了一种基于空谱域兴趣点的高空间分辨率高光谱图像分类方法，利用空谱域兴趣点可以有效利用空域和光谱域的信息进行图像分类。本发明通过空谱域兴趣点检测算法提取高光谱图像的空谱域兴趣点，通过图像分块将高光谱图像分割成小块进行学习、训练和分类，提出了一种新的高光谱图像分类框架，能有效对高光谱图像进行表示及分类，显著提高了高空间分辨率高光谱图像分类的准确性和效率。
附图说明
图1为本发明一实施方式中高光谱图像的分类方法流程图；
图2为本发明一实施方式中图1所示的步骤S11的详细流程图；
图3为本发明一实施方式中图1所示的步骤S12的详细流程图；
图4为本发明一实施方式中图1所示的步骤S13的详细流程图；
图5为本发明一实施方式中高空间分辨率高光谱图像分类结果的图表；
图6为本发明一实施方式中高光谱图像的伪彩色图像分类结果的图表；
图7为本发明一实施方式中高光谱图像的分类方法的整体框架图；
图8为本发明一实施方式中高光谱图像的分类系统10的内部结构示意图；
图9为本发明一实施方式中图8所示的学习模块11的内部结构示意图；
图10为本发明一实施方式中图8所示的训练模块12的内部结构示意图；
图11为本发明一实施方式中图8所示的分类模块13的内部结构示意图。
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。
本发明具体实施方式提供了一种高光谱图像的分类方法，所述方法主要包括如下步骤：
S11、通过码本学习方式获取码本；
S12、通过训练方式训练支持向量机分类器；
S13、利用所述码本和所述支持向量机分类器进行高光谱图像的分类，并输出分类结果。
本发明提出了一种基于空谱域兴趣点的高空间分辨率高光谱图像分类方法，利用空谱域兴趣点可以有效利用空域和光谱域的信息进行图像分类。本发明通过空谱域兴趣点检测算法提取高光谱图像的空谱域兴趣点，通过图像分块将高光谱图像分割成小块进行学习、训练和分类，提出了一种新的高光谱图像分类框架，能有效对高光谱图像进行表示及分类，显著提高了高空间分辨率高光谱图像分类的准确性和效率。
以下将对本发明所提供的一种高光谱图像的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图像的分类方法流程图。
在步骤S11中，通过码本学习方式获取码本。
在本实施方式中，所述通过码本学习方式获取码本的步骤S12具体包括S111—S113，如图2所示。
请参阅图2，为本发明一实施方式中图1所示的步骤S11的详细流程图。
在步骤S111中，对待学习码本的高光谱图像进行图像分块。
在本实施方式中，对待学习码本的高空间分辨率高光谱图像集I＝{I1,I2,I3,…,In}中的图像Ii进行图像分块，可以表示为
Figure PCTCN2016077976-appb-000001
其中
Figure PCTCN2016077976-appb-000002
分别表示图像块，其中ni表示图像块的数量。
在步骤S112中，用m个线程并行地提取每幅高光谱图像中所有图像块的空谱域兴趣点，并对所有空谱域兴趣点进行特征描述，以获取每一个空谱域兴趣点的描述向量。
在本实施方式中，空谱域兴趣点是在高光谱图像的空域和光谱域组成的空谱域上提取的特征点，反映了空域和光谱域上的特性，利用空谱域兴趣点可以有效利用空域和光谱域的信息进行图像分类。
在本实施方式中，用m个线程并行地提取步骤S111中每幅高光谱图像中所有图像块的空谱域兴趣点，并选用3D SIFT描述子对所有空谱域兴趣点进行特征描述，将每一个空谱域兴趣点描述为640维的描述向量。
在步骤S113中，利用预设聚类方法对获取的所有空谱域兴趣点的描述向量进行聚类以获取码本。
在本实施方式中，预设聚类方法为K-means聚类方法，利用K-means聚类方法对步骤S112中得到的所有的描述向量进行聚类，聚类后得到的k个聚类中心，称为码字或视觉单词，用vi来表示。而所有视觉单词的组合即为码本V＝{v1，v2，v3，…，vk}。
请继续参阅图1，在步骤S12中，通过训练方式训练支持向量机分类器。
在本实施方式中，所述通过训练方式训练支持向量机(Support Vector Machine，SVM)分类器的步骤S12具体包括S121—S124，如图3所示。
请参阅图3，为本发明一实施方式中图1所示的步骤S12的详细流程图。
在步骤S121中，对待训练的支持向量机分类器的高光谱图像进行图像分块。
在本实施方式中，对待训练的SVM分类器的高空间分辨率高光谱图像集T＝{T1,T2,T3,…,TL}中的图像Tj进行图像分块，可以表示为
Figure PCTCN2016077976-appb-000003
其中
Figure PCTCN2016077976-appb-000004
分别表示图像块，其中nj表示图像块的数量。
在步骤S122中，用m个线程并行地提取每幅高光谱图像中所有图像块的空谱域兴趣点，并对所有空谱域兴趣点进行特征描述，以获取每一个空谱域兴趣点的描述向量。
在本实施方式中，用m个线程并行地提取步骤S121中每幅高光谱图像中所有图像块的空谱域兴趣点，并用3D SIFT描述子对所有空谱域兴趣点进行特 征描述，将每一个空谱域兴趣点描述为640维的描述向量。
在步骤S123中，基于获取到的所述码本，利用预设模型对每幅高光谱图像的空谱域特征集进行编码，以获取该图像的描述矢量。
在本实施方式中，预设模型为BoW模型，基于图2所示的学习步骤中得到的码本V，利用BoW模型对每幅待训练的SVM分类器的高空间分辨率高光谱图像Tj的空谱域特征集进行编码，得到所有图像的描述矢量，记为：S＝{S1,S2,S3,…,SL}，其中Sj∈S为Tj∈T的图像描述矢量。
在步骤S124中，利用获取到的所述描述矢量训练支持向量机分类器。
在本实施方式中，利用步骤S123得到的T的描述矢量集S训练SVM分类器。
请继续参阅图1，在步骤S13中，利用所述码本和所述支持向量机分类器进行高光谱图像的分类，并输出分类结果。
在本实施方式中，所述利用所述码本和所述支持向量机分类器进行高光谱图像的分类，并输出分类结果的步骤S13具体包括S131—S134，如图4所示。
请参阅图4，为本发明一实施方式中图1所示的步骤S13的详细流程图。
在步骤S131中，将待分类的高光谱图像进行图像分块。
在本实施方式中，将待分类的高空间分辨率高光谱图像Itest进行图像分块，可以表示为Itest＝{b1，b2，b3，…，bN}，其中b1，b2，b3，…，bN分别表示图像块，N表示图像块的数量。
在步骤S132中，用m个线程并行地提取待分类的高光谱图像中所有图像块的空谱域兴趣点，并对所有空谱域兴趣点进行特征描述，以获取每一个空谱域兴趣点的描述向量。
在本实施方式中，同样用m个线程并行地提取步骤S131中待分类的高光谱图像中所有图像Itest对应的图像块的空谱域兴趣点，并用3D SIFT描述子对每一个空谱域兴趣点进行描述，将每一个空谱域兴趣点描述为640维的描述向量，将N块高光谱图像块中提取的空谱域兴趣点融合形成高光谱图像Itest的空谱域 特征集，记为Ftest＝{f1,f2,…,fM}。
在步骤S133中，基于获取到的所述码本，利用预设模型对待分类的高光谱图像中的空谱域兴趣点集进行编码，以获取该图像的描述矢量。
在本实施方式中，预设模型为BoW模型，基于图2所示的学习步骤中得到的码本V，利用BoW模型对待分类的高空间分辨率高光谱图像中图像Itest的空谱域兴趣点集Ftest进行编码，得到图像Itest的描述矢量D。
在步骤S134中，利用训练好的所述支持向量机分类器对获取到的所述描述矢量进行分类，并输出分类结果。
在本实施方式中，利用图3所示的训练步骤所训练好的SVM分类器对步骤S133中得到的描述矢量D进行分类，输出分类结果。
在本实施方式中，不管是步骤S11，步骤S12，还是步骤S13，都是针对大量图像而言，包括连续的多幅图像，而不仅仅只是针对一幅图像而言。
在本实施方式中，应用上述的基于空谱域兴趣点的高空间分辨率高光谱图像的分类方法对一组高空间分辨率高光谱图像进行分类实验，在学习码本时，待学习码本的每幅图像提取3500个空谱域兴趣点用于聚类，聚类后得到的词汇数设为2500。重复进行10次分类实验我们得到如图5所示的分类结果。计算这10次实验的分类结果的均值和标准差，可知，这组高光谱图像分类的准确率为93.22±0.51(％)。
为了说明本发明提出的基于空谱域兴趣点的高空间分辨率高光谱图像的分类方法的有效性，还提取了上述所有待分类高光谱图像的伪彩色图进行了二维图像的分类实验，并计算其准确率。重复进行10次实验，得到如图6所示的分类结果。计算这10次实验的分类结果的均值和标准差，可得到分类的准确率为84.42±0.70(％)。
与图5相比，图6中显示的高光谱图像的伪彩色图像分类的准确率远低于高光谱图像分类的准确率，这充分说明了本发明提出的基于空谱域兴趣点的高空间分辨率高光谱图像的分类方法与系统的有效性以及准确性。
请参阅图7，为本发明一实施方式中高光谱图像的分类方法的整体框架图。
本发明提供的一种高光谱图像的分类方法利用空谱域兴趣点可以有效利用空域和光谱域的信息进行图像分类，能有效对高光谱图像进行表示及分类，显著提高了高空间分辨率高光谱图像分类的准确性和效率。
请参阅图8，所示为本发明一实施方式中高光谱图像的分类系统10的结构示意图。在本实施方式中，高光谱图像的分类系统10主要包括学习模块11、训练模块12以及分类模块13。该高光谱图像的分类系统10，利用空谱域兴趣点可以有效利用空域和光谱域的信息进行图像分类，能有效对高光谱图像进行表示及分类，显著提高了高空间分辨率高光谱图像分类的准确性和效率。
其中，学习模块11，用于通过码本学习方式获取码本。
在本实施方式中，学习模块11具体包括第一分块子模块111、第一描述子模块112以及第一聚类子模块113，如图7所示。
请参阅图9，所示为本发明一实施方式中图8所示的学习模块11的内部结构示意图。
第一分块子模块111，用于对待学习码本的高光谱图像进行图像分块。
在本实施方式中，第一分块子模块111具体的图像分块方法详见前述步骤S111中的相关记载，在此不做重复描述。
第一描述子模块112，用于用m个线程并行地提取每幅高光谱图像中所有图像块的空谱域兴趣点，并对所有空谱域兴趣点进行特征描述，以获取每一个空谱域兴趣点的描述向量。
在本实施方式中，第一描述子模块112具体的特征描述方法详见前述步骤S112中的相关记载，在此不做重复描述。
第一聚类子模块113，用于利用预设聚类方法对获取的所有空谱域兴趣点的描述向量进行聚类以获取码本。
在本实施方式中，第一聚类子模块113具体的聚类方法详见前述步骤S113 中的相关记载，在此不做重复描述。
请继续参阅图8，训练模块12，用于通过训练方式训练支持向量机分类器。
在本实施方式中，训练模块12具体包括第二分块子模块121、第二描述子模块122、第一编码子模块123以及第一训练子模块124，如图10所示。
请参阅图10，所示为本发明一实施方式中图8所示的训练模块12的内部结构示意图。
第二分块子模块121，用于对待训练的支持向量机分类器的高光谱图像进行图像分块。
在本实施方式中，第二分块子模块121具体的图像分块方法详见前述步骤S121中的相关记载，在此不做重复描述。
第二描述子模块122，用于用m个线程并行地提取每幅高光谱图像中所有图像块的空谱域兴趣点，并对所有空谱域兴趣点进行特征描述，以获取每一个空谱域兴趣点的描述向量。
在本实施方式中，第二描述子模块122具体的特征描述方法详见前述步骤S122中的相关记载，在此不做重复描述。
第一编码子模块123，用于基于获取到的所述码本，利用预设模型对每幅高光谱图像的空谱域特征集进行编码，以获取该图像的描述矢量。
在本实施方式中，第一编码子模块123具体的编码方法详见前述步骤S123中的相关记载，在此不做重复描述。
第一训练子模块124，用于利用获取到的所述描述矢量训练支持向量机分类器。
在本实施方式中，第一训练子模块124具体的训练方法详见前述步骤S124中的相关记载，在此不做重复描述。
请继续参阅图8，分类模块13，用于利用所述码本和所述支持向量机分类器进行高光谱图像的分类，并输出分类结果。
在本实施方式中，分类模块13具体包括第三分块子模块131、第三描述子 模块132、第二编码子模块133以及第一分类子模块134，如图11所示。
请参阅图11，所示为本发明一实施方式中图8所示的分类模块13的内部结构示意图。
第三分块子模块131，用于将待分类的高光谱图像进行图像分块。
在本实施方式中，第三分块子模块131具体的图像分块方法详见前述步骤S131中的相关记载，在此不做重复描述。
第三描述子模块132，用于用m个线程并行地提取待分类的高光谱图像中所有图像块的空谱域兴趣点，并对所有空谱域兴趣点进行特征描述，以获取每一个空谱域兴趣点的描述向量。
在本实施方式中，第三描述子模块132具体的特征描述方法详见前述步骤S132中的相关记载，在此不做重复描述。
第二编码子模块133，用于基于获取到的所述码本，利用预设模型对待分类的高光谱图像中的空谱域兴趣点集进行编码，以获取该图像的描述矢量。
在本实施方式中，第二编码子模块133具体的编码方法详见前述步骤S133中的相关记载，在此不做重复描述。
第一分类子模块134，用于利用训练好的所述支持向量机分类器对获取到的所述描述矢量进行分类，并输出分类结果。
在本实施方式中，第一分类子模块134具体的分类方法详见前述步骤S134中的相关记载，在此不做重复描述。
本发明提供的一种高光谱图像的分类系统10，利用空谱域兴趣点可以有效利用空域和光谱域的信息进行图像分类，能有效对高光谱图像进行表示及分类，显著提高了高空间分辨率高光谱图像分类的准确性和效率。
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 (8)

  1. 一种高光谱图像的分类方法，其特征在于，所述分类方法包括：
    通过码本学习方式获取码本；
    通过训练方式训练支持向量机分类器；
    利用所述码本和所述支持向量机分类器进行高光谱图像的分类，并输出分类结果。
  2. 如权利要求1所述的高光谱图像的分类方法，其特征在于，所述通过码本学习方式获取码本的步骤具体包括：
    对待学习码本的高光谱图像进行图像分块；
    用m个线程并行地提取每幅高光谱图像中所有图像块的空谱域兴趣点，并对所有空谱域兴趣点进行特征描述，以获取每一个空谱域兴趣点的描述向量；
    利用预设聚类方法对获取的所有空谱域兴趣点的描述向量进行聚类以获取码本。
  3. 如权利要求2所述的高光谱图像的分类方法，其特征在于，所述通过训练方式训练支持向量机分类器的步骤具体包括：
    对待训练的支持向量机分类器的高光谱图像进行图像分块；
    用m个线程并行地提取每幅高光谱图像中所有图像块的空谱域兴趣点，并对所有空谱域兴趣点进行特征描述，以获取每一个空谱域兴趣点的描述向量；
    基于获取到的所述码本，利用预设模型对每幅高光谱图像的空谱域特征集进行编码，以获取该图像的描述矢量；
    利用获取到的所述描述矢量训练支持向量机分类器。
  4. 如权利要求3所述的高光谱图像的分类方法，其特征在于，所述利用所述码本和所述支持向量机分类器进行高光谱图像的分类，并输出分类结果的步骤具体包括：
    将待分类的高光谱图像进行图像分块；
    用m个线程并行地提取待分类的高光谱图像中所有图像块的空谱域兴趣 点，并对所有空谱域兴趣点进行特征描述，以获取每一个空谱域兴趣点的描述向量；
    基于获取到的所述码本，利用预设模型对待分类的高光谱图像中的空谱域兴趣点集进行编码，以获取该图像的描述矢量；
    利用训练好的所述支持向量机分类器对获取到的所述描述矢量进行分类，并输出分类结果。
  5. 一种高光谱图像的分类系统，其特征在于，所述分类系统包括：
    学习模块，用于通过码本学习方式获取码本；
    训练模块，用于通过训练方式训练支持向量机分类器；
    分类模块，用于利用所述码本和所述支持向量机分类器进行高光谱图像的分类，并输出分类结果。
  6. 如权利要求5所述的高光谱图像的分类系统，其特征在于，所述学习模块包括：
    第一分块子模块，用于对待学习码本的高光谱图像进行图像分块；
    第一描述子模块，用于用m个线程并行地提取每幅高光谱图像中所有图像块的空谱域兴趣点，并对所有空谱域兴趣点进行特征描述，以获取每一个空谱域兴趣点的描述向量；
    第一聚类子模块，用于利用预设聚类方法对获取的所有空谱域兴趣点的描述向量进行聚类以获取码本。
  7. 如权利要求6所述的高光谱图像的分类系统，其特征在于，所述训练模块包括：
    第二分块子模块，用于对待训练的支持向量机分类器的高光谱图像进行图像分块；
    第二描述子模块，用于用m个线程并行地提取每幅高光谱图像中所有图像块的空谱域兴趣点，并对所有空谱域兴趣点进行特征描述，以获取每一个空谱域兴趣点的描述向量；
    第一编码子模块，用于基于获取到的所述码本，利用预设模型对每幅高光 谱图像的空谱域特征集进行编码，以获取该图像的描述矢量；
    第一训练子模块，用于利用获取到的所述描述矢量训练支持向量机分类器。
  8. 如权利要求7所述的高光谱图像的分类系统，其特征在于，所述分类模块包括：
    第三分块子模块，用于将待分类的高光谱图像进行图像分块；
    第三描述子模块，用于用m个线程并行地提取待分类的高光谱图像中所有图像块的空谱域兴趣点，并对所有空谱域兴趣点进行特征描述，以获取每一个空谱域兴趣点的描述向量；
    第二编码子模块，用于基于获取到的所述码本，利用预设模型对待分类的高光谱图像中的空谱域兴趣点集进行编码，以获取该图像的描述矢量；
    第一分类子模块，用于利用训练好的所述支持向量机分类器对获取到的所述描述矢量进行分类，并输出分类结果。
PCT/CN2016/077976 2016-03-31 2016-03-31 一种高光谱图像的分类方法及其系统 WO2017166156A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2016/077976 WO2017166156A1 (zh) 2016-03-31 2016-03-31 一种高光谱图像的分类方法及其系统

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2016/077976 WO2017166156A1 (zh) 2016-03-31 2016-03-31 一种高光谱图像的分类方法及其系统

Publications (1)

Publication Number Publication Date
WO2017166156A1 true WO2017166156A1 (zh) 2017-10-05

Family

ID=59963323

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2016/077976 WO2017166156A1 (zh) 2016-03-31 2016-03-31 一种高光谱图像的分类方法及其系统

Country Status (1)

Country Link
WO (1) WO2017166156A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111428787A (zh) * 2020-03-24 2020-07-17 上海海洋大学 一种基于gpu的高光谱图像并行分类方法
CN112149758A (zh) * 2020-10-24 2020-12-29 中国人民解放军国防科技大学 一种基于欧式距离和深度学习的高光谱开放集分类方法
CN115240074A (zh) * 2022-09-22 2022-10-25 山东锋士信息技术有限公司 一种基于协方差表示的高光谱图像分类方法及设备

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102156885A (zh) * 2010-02-12 2011-08-17 中国科学院自动化研究所 基于级联式码本生成的图像分类方法
CN102156871A (zh) * 2010-02-12 2011-08-17 中国科学院自动化研究所 基于类别相关的码本和分类器投票策略的图像分类方法
CN103208011A (zh) * 2013-05-05 2013-07-17 西安电子科技大学 基于均值漂移和组稀疏编码的高光谱图像空谱域分类方法
US8761510B2 (en) * 2011-11-19 2014-06-24 Nec Laboratories America, Inc. Object-centric spatial pooling for image classification
CN104021396A (zh) * 2014-06-23 2014-09-03 哈尔滨工业大学 基于集成学习的高光谱遥感数据分类方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102156885A (zh) * 2010-02-12 2011-08-17 中国科学院自动化研究所 基于级联式码本生成的图像分类方法
CN102156871A (zh) * 2010-02-12 2011-08-17 中国科学院自动化研究所 基于类别相关的码本和分类器投票策略的图像分类方法
US8761510B2 (en) * 2011-11-19 2014-06-24 Nec Laboratories America, Inc. Object-centric spatial pooling for image classification
CN103208011A (zh) * 2013-05-05 2013-07-17 西安电子科技大学 基于均值漂移和组稀疏编码的高光谱图像空谱域分类方法
CN104021396A (zh) * 2014-06-23 2014-09-03 哈尔滨工业大学 基于集成学习的高光谱遥感数据分类方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
LI, YANSHAN ET AL.: "A Novel Visual Codebook Model Based on Fuzzy Geometry for Large-Scale Image Classification", PATTERN RECOGNITION, vol. 48, no. 10, 4 March 2015 (2015-03-04), pages 3125 - 3134, XP029177380 *

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111428787A (zh) * 2020-03-24 2020-07-17 上海海洋大学 一种基于gpu的高光谱图像并行分类方法
CN112149758A (zh) * 2020-10-24 2020-12-29 中国人民解放军国防科技大学 一种基于欧式距离和深度学习的高光谱开放集分类方法
CN112149758B (zh) * 2020-10-24 2022-03-04 中国人民解放军国防科技大学 一种基于欧式距离和深度学习的高光谱开放集分类方法
CN115240074A (zh) * 2022-09-22 2022-10-25 山东锋士信息技术有限公司 一种基于协方差表示的高光谱图像分类方法及设备
CN115240074B (zh) * 2022-09-22 2023-08-11 山东锋士信息技术有限公司 一种基于协方差表示的高光谱图像分类方法及设备

Similar Documents

Publication Publication Date Title
KR101917000B1 (ko) 화물을 검사하는 방법 및 그 시스템
Fiaschi et al. Learning to count with regression forest and structured labels
EP2808827B1 (en) System and method for OCR output verification
JP6687364B2 (ja) 画像署名をフィッシャーベクトルとマッチングするための適応型語彙
CN105809205B (zh) 一种高光谱图像的分类方法及其系统
CN106845341B (zh) 一种基于虚拟号牌的无牌车辆识别方法
US11055538B2 (en) Object re-identification with temporal context
KR102024867B1 (ko) 예제 피라미드에 기초하여 입력 영상의 특징을 추출하는 방법 및 얼굴 인식 장치
Tuia et al. Recent advances in domain adaptation for the classification of remote sensing data
AU2011207120B8 (en) Identifying matching images
Li et al. Robust vehicle detection in high-resolution aerial images with imbalanced data
US20130259371A1 (en) Appearance and Context Based Object Classification in Images
WO2017166156A1 (zh) 一种高光谱图像的分类方法及其系统
Awang et al. Vehicle counting system based on vehicle type classification using deep learning method
WO2007026951A1 (ja) 画像検索方法及び装置
Yasmeen et al. Text detection and classification from low quality natural images
Sanin et al. K-tangent spaces on Riemannian manifolds for improved pedestrian detection
Vrtagić et al. Video Data Extraction and Processing for Investigation of Vehicles' Impact on the Asphalt Deformation Through the Prism of Computational Algorithms.
Xi et al. Beyond context: Exploring semantic similarity for tiny face detection
CN115049889A (zh) 存储介质和推理方法
Han et al. Image recognition by learned linear subspace of combined bag-of-features and low-level features
Zhi-Jie Image classification method based on visual saliency and bag of words model
WO2015178001A1 (ja) 画像照合システム、画像照合方法、およびプログラムを記憶する記録媒体
Lin et al. Bayesian information criterion based feature filtering for the fusion of multiple features in high-spatial-resolution satellite scene classification
Chaib et al. A VHR scene classification method integrating sparse PCA and saliency computing

Legal Events

Date Code Title Description
NENP Non-entry into the national phase

Ref country code: DE

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

Ref document number: 16895937

Country of ref document: EP

Kind code of ref document: A1

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

Free format text: NOTING OF LOSS OF RIGHTS (EPO FORM 1205A DATED 04.02.2019)

122 Ep: pct application non-entry in european phase

Ref document number: 16895937

Country of ref document: EP

Kind code of ref document: A1