WO2022082848A1 - 高光谱图像分类方法及相关设备 - Google Patents

高光谱图像分类方法及相关设备 Download PDF

Info

Publication number
WO2022082848A1
WO2022082848A1 PCT/CN2020/125243 CN2020125243W WO2022082848A1 WO 2022082848 A1 WO2022082848 A1 WO 2022082848A1 CN 2020125243 W CN2020125243 W CN 2020125243W WO 2022082848 A1 WO2022082848 A1 WO 2022082848A1
Authority
WO
WIPO (PCT)
Prior art keywords
sample set
classified
training
classification
model
Prior art date
Application number
PCT/CN2020/125243
Other languages
English (en)
French (fr)
Inventor
贾森
赵晴晴
徐萌
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 WO2022082848A1 publication Critical patent/WO2022082848A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00Scenes; Scene-specific elements
    • G06V20/10Terrestrial scenes
    • G06V20/13Satellite images
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/21Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214Generating training patterns; Bootstrap methods, e.g. bagging or boosting

Definitions

  • the present application relates to the technical field of image processing, and in particular, to a hyperspectral image classification method and related equipment.
  • hyperspectral images with spectral resolution in the order of magnitude of 10-2 ⁇ are called hyperspectral images.
  • hyperspectral sensors mounted on different space platforms namely imaging spectrometers, in the ultraviolet, visible, near-infrared and mid-infrared regions of the electromagnetic spectrum, the target area is simultaneously imaged with tens to hundreds of continuous and subdivided spectral bands . While obtaining the surface image information, it also obtains its spectral information, which achieves the combination of the spectrum and the image.
  • hyperspectral images are not only greatly improved in terms of information richness, but also provide a possibility for more reasonable and effective analysis and processing of such spectral data in terms of processing technology.
  • the embodiments of the present application provide a hyperspectral image classification method and related equipment, which can improve the accuracy of image classification while ensuring adaptation to a variety of hyperspectral image data.
  • an embodiment of the present application provides a hyperspectral image classification method, including:
  • the processing step is to perform the following processing on the training sample set and the to-be-classified sample set:
  • a second preset number of pixels that satisfy the preset selection strategy are selected from the sample set to be classified, and the second preset number of pixels adding points to the training sample set to update the training sample set, and deleting corresponding pixels from the to-be-classified sample set to update the to-be-classified sample set;
  • Classification step using the updated set of training samples for model training to obtain a first image classification model, and using the first image classification model to predict the updated set of samples to be classified to obtain each sample to be classified The first classification prediction information of .
  • K training subsets and a preset selection strategy to select a second preset number of pixels that satisfy the preset selection strategy from the sample set to be classified, including:
  • K training subsets to perform model training to obtain K second image classification models, respectively use the K second image classification models to predict the sample set to be classified to obtain the value of each sample to be classified.
  • K second classification prediction information
  • the second preset number of pixels are selected from the sample set to be classified according to the second classification prediction information.
  • performing model training by using the K training subsets to obtain K second image classification models includes:
  • the K training subsets and the second to-be-trained model are used for model training, respectively, to obtain the K second image classification models, and the second to-be-trained models include a collaborative representation classification model, a support vector machine, and extreme learning. any one of the machines.
  • the selecting the second preset number of pixels from the sample set to be classified according to the second classification prediction information includes:
  • the second preset number of pixels with the lowest classification confidence are selected from the sample set to be classified according to the classification confidence.
  • the updated training sample set and the updated sample set to be classified are regarded as a new training sample set and a new sample set to be classified, and are processed as described above. processing of steps;
  • the classifying step includes:
  • the updated training sample set and the updated sample set to be classified are regarded as the new training sample set and the new sample set to be classified, and the Process as described in the process steps.
  • performing model training using the updated training sample set to obtain a first image classification model including:
  • the first model to be trained includes support vector machines, sparse representation classification models, and multiple logistic regression classification any of the models.
  • a determination module configured to determine a training sample set and a sample set to be classified of a target hyperspectral image including K-type features; the target hyperspectral image is composed of the pixel points in the training sample set and the sample set to be classified
  • the set of training samples includes the first preset number of pixels of each type of features
  • a processing module configured to perform the following processing on the training sample set and the to-be-classified sample set:
  • K training subsets are generated according to the training sample set, wherein the training sample set is culled from the pixel points of a class of features to obtain a training subset;
  • a second preset number of pixels that satisfy the preset selection strategy are selected from the sample set to be classified, and the second preset number of pixels adding points to the training sample set to update the training sample set, and deleting corresponding pixels from the to-be-classified sample set to update the to-be-classified sample set;
  • a classification module configured to perform model training using the updated training sample set to obtain a first image classification model, and use the first image classification model to predict the updated sample set to be classified to obtain each to-be-classified sample set The first classification prediction information of the classified sample.
  • an embodiment of the present application provides a hyperspectral image classification device, including: a processor and a memory;
  • the processor is connected to a memory, wherein the memory is used to store program codes, and the processor is used to call the program codes to execute the hyperspectral image classification method according to the first aspect.
  • an embodiment of the present application provides a computer storage medium, where the computer storage medium stores a computer program, and the computer program includes program instructions, and when executed by a processor, the program instructions are executed as in the first aspect The described hyperspectral image classification method.
  • the first classification prediction information of each to-be-classified sample realizes the object classification of the target hyperspectral image.
  • the classification method of the embodiment of the present application adopts the class ablation strategy, can process the target hyperspectral image in multiple views, can effectively enhance the accuracy of small sample classification, and has high classification accuracy of the objects in the hyperspectral image; in addition, based on class ablation
  • the active learning method implements hyperspectral image classification, which can adapt to the input target hyperspectral image.
  • FIG. 1 is a schematic flowchart of a hyperspectral image classification method provided in an embodiment of the present application
  • FIG. 2 is a schematic flowchart of a hyperspectral image classification method provided by an embodiment of the present application
  • FIG. 3 is a schematic structural diagram of a hyperspectral image classification device provided by an embodiment of the present application.
  • FIG. 4 is a schematic structural diagram of a hyperspectral image classification device provided by an embodiment of the present application.
  • the existing hyperspectral image classification methods not only have low classification accuracy of the ground objects in the hyperspectral image, but also require different parameter settings when facing the hyperspectral image data collected by different resolutions and different sensors, namely, When applying different hyperspectral image data, manual intervention is required for parameter adjustment, and there is room for improvement.
  • the present application proposes a hyperspectral image classification method, which can effectively improve the classification accuracy of ground objects in the hyperspectral images while adapting to different hyperspectral images.
  • FIG. 1 is a schematic flowchart of a hyperspectral image classification method provided by an embodiment of the present application; the hyperspectral image classification method includes:
  • Determining step 101 determining a training sample set and a sample set to be classified of a target hyperspectral image including K-type features; the target hyperspectral image is composed of pixels in the training sample set and pixels in the sample set to be classified, and the training sample The set includes a first preset number of pixels of each type of features;
  • the target hyperspectral image includes K-type objects, and the objects can be any objects, such as animals, plants, vehicles, buildings, and so on.
  • the specific value of the first preset number can be set according to actual needs, for example, 3, 5, 10, 15, 20, 50, and so on.
  • a first preset number of pixels are randomly selected, and the selected pixels are manually marked with specific categories to form a training sample set.
  • the unselected pixels in the target hyperspectral image constitute the sample set to be classified. Assuming that the total number of pixels of the target hyperspectral image is X, and the first preset number is 10, the number of samples in the training sample set is 10K, and the number of samples in the sample set to be classified is X-10K.
  • processing step 102 the following processing is performed on the training sample set and the sample set to be classified:
  • K training subsets are generated according to the training sample set, wherein the training sample set is culled from the pixels of a class of ground objects to obtain a training subset;
  • the pixels of a class of ground objects in the training sample set are removed, and the remaining pixels in the training sample set constitute a training subset. Since the training sample set includes K types of objects, the pixels of one category of objects are eliminated each time, and finally K training subsets can be obtained.
  • K training subsets and a preset selection strategy select a second preset number of pixels from the sample set to be classified that meet the preset selection strategy, and add the second preset number of pixels to the training sample set to update the training sample set, and delete the corresponding pixel points from the sample set to be classified to update the sample set to be classified;
  • the preset selection strategy can be adjusted according to needs; similarly, the specific value of the second preset number can be set according to actual needs, for example, 10, 15, 20, 50 and so on.
  • a second preset number of pixels that satisfy the preset selection strategy can be selected from the sample set to be classified, and these pixels are manually labeled and classified, and then added to the In the training sample set, to update the training sample set; at the same time, these pixels are also deleted from the to-be-classified sample set to update the to-be-classified sample set.
  • Classification step 103 using the updated training sample set for model training to obtain a first image classification model, and using the first image classification model to predict the updated sample set to be classified to obtain the first classification prediction of each sample to be classified information.
  • the model is trained to obtain a trained first image classification model; the updated sample set to be classified is predicted by using the first image classification model to obtain each sample set to be classified.
  • the first classification prediction information of the sample to be classified since the target hyperspectral image is composed of pixels in the training sample set and pixels in the sample set to be classified, the categories of the pixels in the training sample set have been marked manually, therefore, when using the first image classification model After obtaining the first classification prediction information of the pixel points in the sample set to be classified, that is, the category of the pixel points, the classification of the target hyperspectral image can be completed.
  • the first image classification model may also be used to perform prediction processing on all the pixel points of the target hyperspectral image, so as to obtain the classification prediction result of the target hyperspectral image.
  • the classification method of the embodiment of the present application adopts a class ablation strategy to obtain multiple training subsets, can process target hyperspectral images in multiple views, effectively enhances the accuracy of small sample classification, and has high classification accuracy for objects in hyperspectral images;
  • the active learning method based on class ablation realizes hyperspectral image classification, which can adapt to the input target hyperspectral image.
  • a second preset number of pixels that satisfy the preset selection strategy are selected from the sample set to be classified, including:
  • K training subsets are used for model training to obtain K second image classification models; then K second image classification models are used to perform category prediction on the sample set to be classified, so as to obtain each sample set in the sample set to be classified.
  • the second classification prediction information of each sample to be classified since there are K second image classification models, one sample to be classified can obtain K pieces of second classification prediction information.
  • the second classification prediction information is classification information describing the samples to be classified by the second image classification model.
  • K training subsets for model training to obtain K second image classification models includes:
  • the K training subsets and the second to-be-trained model are used for model training respectively to obtain K second image classification models, and the second to-be-trained model includes any one of a collaborative representation classification model, a support vector machine, and an extreme learning machine.
  • Step 1022 Select a second preset number of pixels from the sample set to be classified according to the second classification prediction information.
  • a second preset number of pixels are selected from the sample set to be classified according to the preset selection strategy and the second classification prediction information.
  • the second classification prediction information includes the prediction category and the corresponding residual value; then selecting the second preset number of pixels that meet the preset selection strategy specifically includes:
  • a second preset number of pixels with the lowest classification confidence are selected from the sample set to be classified.
  • 10 pixels with the lowest classification confidence are selected from the sample set to be classified and added to the training sample set.
  • model training is performed using the updated training sample set to obtain a first image classification model, including:
  • Model training is performed using the updated training sample set and the first model to be trained to obtain a first image classification model, where the first model to be trained includes any one of a support vector machine, a sparse representation classification model, and a multiple logistic regression classification model.
  • the classification step 103 includes:
  • Step 1031 determining the update times of the training sample set
  • the number of times of updating the training sample set is recorded.
  • Step 1032 when the number of updates is less than the number of times threshold, the training sample set after the update and the sample set to be classified after the update are used as the new training sample set and the new sample set to be classified, and it is processed as the processing step;
  • the specific value of the number of times threshold can be set according to actual needs.
  • the number of times threshold can be set to 5, 8, 10, 15, etc., or it can be set to 1% of the total number of pixels of the target hyperspectral image, or It can be 2% or other percentage values. Judging the number of updates obtained in step 1031, when the number of updates is less than the number of times threshold, the current updated training sample set and the current updated set of samples to be classified are used as the new training sample set and new to be classified.
  • the sample set perform the processing of the above-mentioned processing step 102, that is, repeat the processing step 102 once, continue to use the new training sample set to select a second preset number of pixels from the new sample set to be classified and add them to the training sample set, update
  • the training sample set is deleted, the corresponding pixel points are deleted from the to-be-classified sample set, and the to-be-classified sample set is updated.
  • Step 1033 when the number of updates is greater than or equal to the number of times threshold, use the updated training sample set to obtain the first classification prediction information of each to-be-classified sample in the updated to-be-classified sample set.
  • the processing steps are not repeatedly performed at this time, but the current updated training sample set is used to obtain the first image classification model, and the first image classification model is used to classify the current The set of samples to be classified is predicted, and the first classification prediction information of each sample to be classified is obtained.
  • the classifying step 103 includes:
  • Step 1034 when it is determined that the first image classification model does not meet the convergence condition, the updated training sample set and the updated sample set to be classified are regarded as the new training sample set and the new sample set to be classified, and are processed as follows. processing of steps.
  • the current updated training sample set can also be used to obtain a first image classification model, and the first image classification model can be used to process the current updated to-be-classified
  • the sample set obtains the first classification prediction information of each sample to be classified.
  • the convergence condition may be the accuracy of the first classification prediction information.
  • an accuracy threshold such as 98% or 99%
  • step numbers of the above steps 1031 , 1032 , 1033 , and 1034 are only for distinguishing the steps, and do not limit the execution order of the steps.
  • FIG. 2 is a schematic flowchart of a hyperspectral image classification method provided by an embodiment of the present application; the hyperspectral image classification method is specifically described below with an example:
  • the training sample set is divided into K training subsets by the method of class ablation, and each training subset excludes one type of ground objects and only contains K-1 types of ground objects, denoted as D 1 , D 2 , ... , D K .
  • each training subset train and predict the samples to be classified separately.
  • the specific operations are as follows:
  • C(y) records the number of times that the sample to be classified is classified into the kth class
  • ⁇ C(y) represents the difference between the number of times the sample to be classified is classified into the most frequent class and the number of times of the second multiple class:
  • ⁇ + represents the class that has been assigned the most times.
  • the C(y,k) function can find out the two categories to which the samples are most easily assigned, and calculate the frequency difference between the two categories by counting the prediction results of different views.
  • samples with high classification confidence are classified into the two most frequent categories, and the absolute value of the difference between the corresponding minimum residual values should be larger.
  • ⁇ Q(y) is used to represent the most frequent category and the second most frequent category.
  • the absolute value of the difference between the residual values corresponding to the subclass that is:
  • function is used to obtain the number of unique prediction categories.
  • the obtained training sample set D is input into the support vector machine for model training, and the sample set G to be classified uses the above model to perform final result prediction.
  • the classification method of the embodiment of the present application can more comprehensively analyze the effective information in the multi-view results (multiple training subsets), and further improve the active learning query strategy; Sample classification accuracy.
  • FIG. 3 is a schematic structural diagram of a hyperspectral image classification device provided by an embodiment of the present application.
  • the hyperspectral image classification device includes:
  • a determination module 301 is used to determine a training sample set and a sample set to be classified of a target hyperspectral image including K-type features; the target hyperspectral image is composed of pixels in the training sample set and pixels in the sample set to be classified,
  • the training sample set includes a first preset number of pixels of each type of features;
  • the processing module 302 is configured to perform the following processing on the training sample set and the sample set to be classified:
  • K training subsets and a preset selection strategy select a second preset number of pixels from the sample set to be classified that meet the preset selection strategy, and add the second preset number of pixels to the training sample set to update the training sample set, and delete the corresponding pixel points from the sample set to be classified to update the sample set to be classified;
  • the classification module 303 is used to perform model training using the updated training sample set to obtain a first image classification model, and use the first image classification model to predict the updated sample set to be classified to obtain the first image classification model of each sample to be classified. Classification prediction information.
  • the processing module 302 uses the K training subsets and the preset selection strategy to select a second preset number of pixel points from the sample set to be classified that satisfy the preset selection strategy, it specifically executes:
  • a second preset number of pixels are selected from the sample set to be classified according to the second classification prediction information.
  • K training subsets for model training to obtain K second image classification models includes:
  • the K training subsets and the second model to be trained are used for model training respectively to obtain K second image classification models, and the second model to be trained includes any one of a collaborative representation classification model, a support vector machine, and an extreme learning machine.
  • the second classification prediction information includes prediction categories and corresponding residual values
  • the processing module selects a second preset number of pixel points that satisfy the preset selection strategy from the sample set to be classified according to the second classification prediction information, it specifically executes:
  • a second preset number of pixels with the lowest classification confidence are selected from the sample set to be classified.
  • the classification module 303 performs model training using the updated training sample set to obtain the first image classification model, and specifically executes:
  • the updated training sample set and the updated to-be-classified sample set are used as the new training sample set and the new to-be-classified sample set, and are processed as in the processing steps;
  • the updated training sample set is used to obtain the first classification prediction information of each to-be-classified sample in the updated to-be-classified sample set.
  • the updated training sample set and the updated sample set to be classified are regarded as the new training sample set and the new sample set to be classified, and the processing steps are performed on them. .
  • the embodiments of the present application further provide a hyperspectral image classification device.
  • FIG. 4 is a schematic structural diagram of a hyperspectral image classification device provided by an embodiment of the present application.
  • the above-mentioned hyperspectral image classification apparatus can be applied to the hyperspectral image classification apparatus 400 , and the hyperspectral image classification apparatus 400 may include: a processor 401 , a network interface 404 and a memory 405 .
  • the hyperspectral image classification device 400 may further include: a user interface 403 , and at least one communication bus 402 . Among them, the communication bus 402 is used to realize the connection and communication between these components.
  • the user interface 403 may include a display screen (Display) and a keyboard (Keyboard), and the optional user interface 403 may also include a standard wired interface and a wireless interface.
  • the network interface 404 may include a standard wired interface and a wireless interface (eg, a WI-FI interface).
  • the memory 405 may be high-speed RAM memory or non-volatile memory, such as at least one disk memory.
  • the memory 405 can optionally also be at least one storage device located away from the aforementioned processor 401 .
  • the memory 405 as a computer storage medium may include an operating system, a network communication module, a user interface module, and a device control application program.
  • the network interface 404 can provide a network communication function; the user interface 403 is mainly used to provide an input interface for the user; and the processor 401 can be used to call the data stored in the memory 405
  • the device controls the application program to implement the steps of the hyperspectral image classification method described in the above embodiment.
  • the hyperspectral image classification apparatus 400 described in the embodiments of the present application can perform the hyperspectral image classification method described above, and can also perform the description of the hyperspectral image classification apparatus described above, which will not be repeated here. In addition, the description of the beneficial effects of using the same method will not be repeated.
  • an embodiment of the present application further provides a computer storage medium, and the computer storage medium stores a computer program executed by the aforementioned hyperspectral image classification device, and the computer program It includes program instructions.
  • the processor executes the program instructions, it can execute the description of the hyperspectral image classification method described above, and therefore will not be repeated here.
  • the description of the beneficial effects of using the same method will not be repeated.
  • the program can be stored in a computer-readable storage medium, and when the program is executed , which may include the processes of the above-mentioned method embodiments.
  • the storage medium may be a magnetic disk, an optical disk, a read-only memory (Read-Only Memory, ROM), or a random access memory (Random Access Memory, RAM) or the like.

Landscapes

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

Abstract

本申请实施例公开一种高光谱图像分类方法及相关设备,先确定目标高光谱图像的训练样本集合和待分类样本集合,再采用类消融策略,由训练样本集合生成K个训练子集;再利用K个训练子集和预设选择策略,从待分类样本集合中选取第二预设数目的像素点加入训练样本集合中,以更新训练样本集合,并更新待分类样本集合;最后,利用更新后的训练样本集合进行模型训练以得到第一图像分类模型,并利用第一图像分类模型对更新后的待分类样本集合进行预测得到每个待分类样本的第一分类预测信息,实现对目标高光谱图像的地物分类。通过多视图处理目标高光谱图像,可以有效增强小样本分类的准确性;基于类消融的主动学习方法可以自适应输入的目标高光谱图像。

Description

高光谱图像分类方法及相关设备 技术领域
本申请涉及图像处理技术领域,尤其涉及一种高光谱图像分类方法及相关设备。
背景技术
光谱分辨率在10-2λ数量级范围内的光谱图像称为高光谱图像(Hyperspectral Image)。通过搭载在不同空间平台上的高光谱传感器,即成像光谱仪,在电磁波谱的紫外、可见光、近红外和中红外区域,以数十至数百个连续且细分的光谱波段对目标区域同时成像。在获得地表图像信息的同时,也获得其光谱信息,做到了光谱与图像的结合。与多光谱遥感影像相比,高光谱影像不仅在信息丰富程度方面有了极大的提高,在处理技术上,对该类光谱数据进行更为合理、有效的分析处理提供了可能。
由于高光谱图像较为复杂的空间-光谱特性,使用一种智能的方式构造尽量小的训练集合进行分类处理尤为重要。现有的高光谱图像分类方法,在使用小样本训练方法时对地物的分类准确度低,而且,面对不同分辨率、不同传感器采集的高光谱图像数据需要不同的参数设置,即在适用不同的高光谱图像数据时,需要人工干预进行参数调整,存在改进空间。
发明内容
本申请实施例提供了一种高光谱图像分类方法及相关设备,在保证自适应多种高光谱图像数据的同时,提升图像分类准确性。
第一方面,本申请实施例提供了一种高光谱图像分类方法,包括:
确定步骤,确定包括K类地物的目标高光谱图像的训练样本集合和待分类样本集合;所述目标高光谱图像由所述训练样本集合中的像素点和所述待分类样本集合中的像素点组成,所述训练样本集合包括每类地物的第一预设数目的像素点;
处理步骤,对所述训练样本集合和所述待分类样本集合进行如下处理:
根据所述训练样本集合生成K个训练子集,其中,将所述训练样本集合剔除一类地物的像素点,以得到一个训练子集;
利用所述K个训练子集和预设选择策略,从所述待分类样本集合中选取第二预设数目的满足所述预设选择策略的像素点,将所述第二预设数目的像素点加入所述训练样本集合以更新所述训练样本集合,并从所述待分类样本集合中删除对应的像素点以更新所述待分类样本集合;
分类步骤,利用更新后的所述训练样本集合进行模型训练以得到第一图像分类模型,并利用所述第一图像分类模型对更新后的所述待分类样本集合进行预测得到每个待分类样本的第一分类预测信息。
可选地,所述利用所述K个训练子集和预设选择策略,从所述待分类样本集合中选取第二预设数目的满足所述预设选择策略的像素点,包括:
分别利用所述K个训练子集进行模型训练以得到K个第二图像分类模型,分别利用所述K个第二图像分类模型对所述待分类样本集合进行预测以得到每个待分类样本的K个第二分类预测信息;
根据所述第二分类预测信息从所述待分类样本集合中选取所述第二预设数目的像素点。
可选地,所述分别利用所述K个训练子集进行模型训练以得到K个第二图像分类模型包括:
分别利用所述K个训练子集和第二待训练模型进行模型训练,以得到所述K个第二图像分类模型,所述第二待训练模型包括协同表示分类模型、支持向量机、极限学习机中的任意一个。
可选地,所述第二分类预测信息包括预测类别以及对应的残差值;
所述根据所述第二分类预测信息从所述待分类样本集合中选取所述第二预设数目的像素点,包括:
根据每个所述待分类样本的K个所述预测类别和K个所述残差值计算对应的分类置信度;
根据所述分类置信度从所述待分类样本集合中选取所述第二预设数目的分类置信度最低的像素点。
可选地,所述分类步骤包括:
确定所述训练样本集合的更新次数;
所述更新次数小于次数阈值时,将更新后的所述训练样本集合和更新后的所述待分类样本集合作为新的训练样本集合和新的待分类样本集合,并对其进行如所述处理步骤的处理;
所述更新次数大于或等于所述次数阈值时,利用更新后的所述训练样本集合获取更新后的所述待分类样本集合中每个待分类样本的第一分类预测信息。
可选地,所述分类步骤包括:
确定所述第一图像分类模型不满足收敛条件时,将更新后的所述训练样本集合和更新后的所述待分类样本集合作为新的训练样本集合和新的待分类样本集合,并对其进行如所述处理步骤的处理。
可选地,所述利用更新后的所述训练样本集合进行模型训练以得到第一图像分类模型,包括:
利用更新后的所述训练样本集合和第一待训练模型进行模型训练,以得到所述第一图像分类模型,所述第一待训练模型包括支持向量机、稀疏表示分类模型、多元逻辑回归分类模型中的任意一个。
第二方面,本申请实施例提供了一种高光谱图像分类装置,包括:
确定模块,用于确定包括K类地物的目标高光谱图像的训练样本集合和待分类样本集合;所述目标高光谱图像由所述训练样本集合中的像素点和所述待分类样本集合中的像素点组成,所述训练样本集合包括每类地物的第一预设数目的像素点;
处理模块,用于对所述训练样本集合和所述待分类样本集合进行如下处理:
根据所述训练样本集合生成K个训练子集,其中,将所述训练样本集合剔除一类地物的像素点,以得到一个训练子集;
利用所述K个训练子集和预设选择策略,从所述待分类样本集合中选取第二预设数目的满足所述预设选择策略的像素点,将所述第二预设数目的像素点加入所述训练样本集合以更新所述训练样本集合,并从所述待分类样本集合中删除对应的像素点以更新所述待分类样本集合;
分类模块,用于利用更新后的所述训练样本集合进行模型训练以得到第一图像分类模型,并利用所述第一图像分类模型对更新后的所述待分类样本集合进行预测得到每个待分类样本的第一分类预测信息。
第三方面,本申请实施例提供了一种高光谱图像分类设备,包括:处理器和存储器;
所述处理器和存储器相连,其中,所述存储器用于存储程序代码,所述处理器用于调用所述程序代码,以执行如第一方面所述的高光谱图像分类方法。
第四方面,本申请实施例提供了一种计算机存储介质,所述计算机存储介质存储有计算机程序,所述计算机程序包括程序指令,所述程序指令当被处理器执行时,执行如第一方面所述的高光谱图像分类方法。
本申请实施例中,先确定包括K类地物的目标高光谱图像的训练样本集合和待分类样本集合,再采用类消融策略,由训练样本集合生成K个训练子集;再利用K个训练子集和预设选择策略,从待分类样本集合中选取第二预设数目的满足预设选择策略的像素点加入训练样本集合中,以更新训练样本集合,并将相应的像素点从待分类样本集合中删除,以更新待分类样本集合;最后,利用更新后的训练样本集合进行模型训练以得到第一图像分类模型,并利用第一图像分类模型对更新后的待分类样本集合进行预测得到每个待分类样本的第一分类预测信息,实现对目标高光谱图像的地物分类。本申请实施例的分类方法,采用类消融策略,可以多视图处理目标高光谱图像,可以有效增强小样本分类的准确性,对高光谱图像中地物的分类准确度高;另外,基于类消融的主动学习方法实现高光谱图像分类,可以自适应输入的目标高光谱图像。
附图说明
为了更清楚地说明本申请实施例技术方案,下面将对实施例描述中所需要使用的附图作简单地介绍,显而易见地,下面描述中的附图是本申请的一些实施例,对于本领域普通技术人员来讲,在不付出创造性劳动的前提下,还可以根据这些附图获得其他的附图。
图1是本申请实施例提供的一种高光谱图像分类方法的流程示意图;
图2是本申请实施例提供的一种高光谱图像分类方法的流程示意图;
图3是本申请实施例提供的一种高光谱图像分类装置的结构示意图;
图4是本申请实施例提供的一种高光谱图像分类设备的结构示意图。
具体实施方式
下面将结合本申请实施例中的附图,对本申请实施例中的技术方案进行清楚、完整地描述。
应当理解,本申请的说明书和权利要求书及附图中的术语“第一”、“第二”等是用于区别不同对象,而不是用于描述特定顺序。此外,术语“包括”和“具有”以及它们任何变形,意图在于覆盖不排他的包含。例如包含了一系列步骤或单元的过程、方法、系统、产品或设备没有限定于已列出的步骤或单元,而是可选地还包括没有列出的步骤或单元,或可选地还包括对于这些过程、方法、产品或设备固有的其它步骤或单元。
在本申请中提及“实施例”意味着,结合实施例描述的特定特征、结构或特性可以包含在本申请的至少一个实施例中。在说明书中的各个位置出现该短语并不一定均是指相同的实施例,也不是与其它实施例互斥的独立的或备选的实施例。本领域技术人员显式地和隐式地理解的是,本申请所描述的实施例可以与其它实施例相结合。
现有的高光谱图像分类方法,不仅对高光谱图像中的地物的分类准确度低,而且,面对不同分辨率、不同传感器采集的高光谱图像数据时,需要进行不同的参数设置,即在适用不同的高光谱图像数据时,需要人工干预进行参数调整,存在改进空间。基于上述技术问题,本申请提出一种高光谱图像分类方法,在自适应不同的高光谱图像的同时,还可以有效提高对高光谱图像中的地物的分类准确度。
请参见图1,图1是本申请实施例提供的一种高光谱图像分类方法的流程示意图;高光谱图像分类方法包括:
确定步骤101,确定包括K类地物的目标高光谱图像的训练样本集合和待分类样本集合;目标高光谱图像由训练样本集合中的像素点和待分类样本集合 中的像素点组成,训练样本集合包括每类地物的第一预设数目的像素点;
具体地,目标高光谱图像中包括K类地物,地物可以为任意物体,例如动物、植物、交通工具、建筑等等。第一预设数目的具体数值可以根据实际需要进行设置,例如设置为3个、5个、10个、15个、20个、50个等等。对于目标高光谱图像中的每一类地物,随机选取第一预设数目的像素点,并对选取的像素点进行人工标记具体类别后,组成训练样本集合。目标高光谱图像中未被选择的像素点组成待分类样本集合。假设目标高光谱图像的像素点总数为X个,第一预设数目为10个,则训练样本集合的样本个数为10K,待分类样本集合的样本个数为X-10K个。
处理步骤102,对训练样本集合和待分类样本集合进行如下处理:
根据训练样本集合生成K个训练子集,其中,将训练样本集合剔除一类地物的像素点,以得到一个训练子集;
具体地,基于类消融原理,在生成一个训练子集的时候,将训练样本集合中一类地物的像素点剔除掉,训练样本集合中剩余的像素点构成一个训练子集。由于训练样本集合包括K类地物,每次剔除一类地物的像素点,最终可以得到K个训练子集。
利用K个训练子集和预设选择策略,从待分类样本集合中选取第二预设数目的满足预设选择策略的像素点,将第二预设数目的像素点加入训练样本集合以更新训练样本集合,并从待分类样本集合中删除对应的像素点以更新待分类样本集合;
具体地,预设选择策略可以根据需要进行调整;同样地,第二预设数目的具体数值可以根据实际需要进行设置,例如设置为10个、15个、20个、50个等等。利用K个训练子集和预设选择策略,可以从待分类样本集合中选取出第二预设数目的满足预设选择策略的像素点,并对这些像素点进行人工标记类别后,将其加入训练样本集合中,以更新训练样本集合;同时,还会将这些像素点从待分类样本集合中删除,以更新待分类样本集合。
分类步骤103,利用更新后的训练样本集合进行模型训练以得到第一图像分类模型,并利用第一图像分类模型对更新后的待分类样本集合进行预测得到 每个待分类样本的第一分类预测信息。
具体地,基于更新后的训练样本集合,对模型进行训练以得到训练后的第一图像分类模型;利用第一图像分类模型对更新后的待分类样本集合进行预测得到待分类样本集合中每个待分类样本的第一分类预测信息。特别地,由于目标高光谱图像由训练样本集合中的像素点和待分类样本集合中的像素点组成,训练样本集合中像素点的类别已经由人工进行标记,因此,在利用第一图像分类模型得到待分类样本集合中的像素点的第一分类预测信息,即像素点的类别后,即可以完成对目标高光谱图像的分类。特别地,也可以利用第一图像分类模型对目标高光谱图像所有的像素点进行预测处理,以得到目标高光谱图像的分类预测结果。
本申请实施例的分类方法,采用类消融策略得到多个训练子集,可以多视图处理目标高光谱图像,有效增强小样本分类的准确性,对高光谱图像中地物的分类准确度高;另外,基于类消融的主动学习方法实现高光谱图像分类,可以自适应输入的目标高光谱图像。
在一个可能的实施例中,处理步骤102中,利用K个训练子集和预设选择策略,从待分类样本集合中选取第二预设数目的满足预设选择策略的像素点,包括:
步骤1021,分别利用K个训练子集进行模型训练以得到K个第二图像分类模型,分别利用K个第二图像分类模型对待分类样本集合进行预测以得到每个待分类样本的K个第二分类预测信息;
具体地,先分别利用K个训练子集进行模型训练以得到K个第二图像分类模型;再分别利用K个第二图像分类模型对待分类样本集合进行类别预测,以得到待分类样本集合中每个待分类样本的第二分类预测信息,由于有K个第二图像分类模型,因此,一个待分类样本可以得到K个第二分类预测信息。其中,第二分类预测信息为描述第二图像分类模型对待分类样本的分类信息。
进一步地,分别利用K个训练子集进行模型训练以得到K个第二图像分类模型包括:
分别利用K个训练子集和第二待训练模型进行模型训练,以得到K个第 二图像分类模型,第二待训练模型包括协同表示分类模型、支持向量机、极限学习机中的任意一个。
步骤1022,根据第二分类预测信息从待分类样本集合中选取第二预设数目的像素点。
具体地,根据预设选择策略和第二分类预测信息从待分类样本集合中选取第二预设数目的像素点。进一步地,本申请实施例中,第二分类预测信息包括预测类别以及对应的残差值;则选取第二预设数目的满足预设选择策略的像素点具体包括:
根据每个待分类样本的K个预测类别和K个残差值计算对应的分类置信度;
根据分类置信度从待分类样本集合中选取第二预设数目的分类置信度最低的像素点。
以第二预设数目为10个为例,从待分类样本集合中选取分类置信度最低的10个像素点加入训练样本集合中。
本申请实施例中,利用分类置信度选取第二预设数目的像素点加入训练样本集合中。其中,一个待分类样本的分类置信度低,表示多个第二图像分类模型对其预测的类别的准确度低,第二图像分类模型不能很好预测它的类别,此时需要将其作为训练样本,以提升第一图像分类模型的分类准确度。
在一个可能的实施例中,分类步骤103中,利用更新后的训练样本集合进行模型训练以得到第一图像分类模型,包括:
利用更新后的训练样本集合和第一待训练模型进行模型训练,以得到第一图像分类模型,第一待训练模型包括支持向量机、稀疏表示分类模型、多元逻辑回归分类模型中的任意一个。
在一个可能的实施例中,分类步骤103包括:
步骤1031,确定训练样本集合的更新次数;
具体地,在对训练样本集合进行更新后,记录训练样本集合的更新次数。
步骤1032,更新次数小于次数阈值时,将更新后的训练样本集合和更新后的待分类样本集合作为新的训练样本集合和新的待分类样本集合,并对其进 行如处理步骤的处理;
具体地,次数阈值的具体数值可以根据实际需要进行设置,例如,次数阈值可以设置为5、8、10、15等等,也可以是设置为目标高光谱图像的像素点总数的1%,也可以是2%,还可以是其他百分比数值。对步骤1031获得的更新次数进行判断,当更新次数小于次数阈值时,此时将当前的更新后的训练样本集合和当前的更新后的待分类样本集合作为新的训练样本集合和新的待分类样本集合,对其进行上述处理步骤102的处理,即重复执行一次处理步骤102,继续利用新的训练样本集合从新的待分类样本集合选取第二预设数目的像素点加入训练样本集合中,更新训练样本集合,并将对应的像素点从待分类样本集合中删除,更新待分类样本集合。
步骤1033,更新次数大于或等于次数阈值时,利用更新后的训练样本集合获取更新后的待分类样本集合中每个待分类样本的第一分类预测信息。
具体地,当判断更新次数大于或等于次数阈值时,此时不重复执行处理步骤,而是利用当前的更新后的训练样本集合获得第一图像分类模型,并利用第一图像分类模型对当前的待分类样本集合进行预测,得到每个待分类样本的第一分类预测信息。
在另一个可能的实施例中,分类步骤103包括:
步骤1034,确定第一图像分类模型不满足收敛条件时,将更新后的训练样本集合和更新后的待分类样本集合作为新的训练样本集合和新的待分类样本集合,并对其进行如处理步骤的处理。
具体地,除了设置次数阈值作为终止重复执行处理步骤的条件之外,还可以利用当前的更新后的训练样本集合得到第一图像分类模型,利用第一图像分类模型处理当前的更新后的待分类样本集合得到每个待分类样本的第一分类预测信息。接着,可以根据第一分类预测信息确定当前的第一图像分类模型是否满足收敛条件,当判断当前的第一图像分类模型不满足收敛条件时,将当前的更新后的训练样本集合和当前的更新后的待分类样本集合作为新的训练样本集合和新的待分类样本集合,并对其进行如处理步骤的处理,直到第一图像分类模型满足收敛条件。特别地,收敛条件可以为第一分类预测信息的准确度, 当第一分类预测信息的分类准确度达到准确度阈值时,例如98%或99%,此时可以认为第一图像分类模型满足收敛条件。
特别指出的是,上述步骤1031、步骤1032、步骤1033、步骤1034的步骤编号仅为对步骤进行区别,不对步骤的执行顺序造成限定。
参考图2,图2是本申请实施例提供的一种高光谱图像分类方法的流程示意图;下面以一个实例对高光谱图像分类方法进行具体说明:
S1、训练样本集合初始化。对于输入的高光谱图像H∈R X×Y×Z,其中X,Y表示图像的空间维度,Z代表图像的光谱维数,R表示实数。H中共有K类地物,人工从每个地物类别中随机选取3个样本(同时进行人工类别标记),共n个像素点作为训练集D∈R Z×n。高光谱图像剩余的像素点作为待分类样本集合G∈R Z×m,其中m为待分类样本的个数。令迭代次数iter=0。
S2、采用类消融的方法将训练样本集合划分为K个训练子集,每个训练子集剔除一类地物,仅包含K-1类地物,记为D 1,D 2,...,D K。对每一个训练子集分别训练并预测待分类样本,具体操作如下:
(1)、对于训练子集D k,k∈(1,K),使用协同表示方法训练模型,根据训练后的模型计算出待分类样本每一类的残差值,选择其中残差值最小的类作为待分类样本的预测类别,预测类别以及对应的残差值即为预测结果L k
(2)、保留所有待分类样本的预测类别
Figure PCTCN2020125243-appb-000001
和其对应的残差值
Figure PCTCN2020125243-appb-000002
(3)、每个训练子集经过(1)和(2)操作后最终得到整体的待分类样本的预测类别P C∈R K×m和对应的残差值P Q∈R K×m
S3、根据查询策略选择候选数据,具体操作如下:
(1)、对于待分类样本y,其分类置信度(Classification Confidence,CC)为:
Figure PCTCN2020125243-appb-000003
其中C(y)记录了待分类样本被分到第k类的次数,△C(y)表示被分到最多次类的次数和第二多次类次数之差:
Figure PCTCN2020125243-appb-000004
Figure PCTCN2020125243-appb-000005
其中,ω +代表了被分到次数最多的类。C(y,k)函数可以找出样本最容易分配的两个类别,通过对不同视图的预测结果进行计数,计算出这两个类别的频率差。
同样,分类置信度高的样本,被分到次数最多的两个类别,其对应的最小残差值之差的绝对值应该更大,用△Q(y)来表示最多次类和第二多次类对应的残差值之差的绝对值,即:
ΔQ(y)=|Q first-Q second|
Figure PCTCN2020125243-appb-000006
Figure PCTCN2020125243-appb-000007
分类置信度较低的样本更有可能根据不同视图改变预测结果,因此也要考虑预测不同类的数量值:
N(y)=|Unique(P C(y))|
其中|Unique(·)|函数用于获取不重复预测类别数目。
(2)、从待分类样本集合中选取15个分类置信度最小,即
Figure PCTCN2020125243-appb-000008
的样本集合S,并对集合S进行人工标记。
S4、将S中的样本添加到训练样本集合D中,并从待分类样本集合G中删去。更新训练样本数n,待分类样本数m。
S5、令iter=iter+1,判断iter是否满足iter>=10,若不满足转到步骤S2,若满足,则转到步骤S6。
S6、获得的训练样本集合D输入支持向量机中进行模型训练,待分类样本集合G使用上述模型进行最终的结果预测。
本申请实施例的分类方法,可以更为全面分析多视图结果(多个训练子集)中的有效信息,进一步完善了主动学习查询策略;另外,多视图处理目标高光谱图像,有效增强了小样本分类准确性。
基于上述高光谱图像分类方法实施例的描述,本申请实施例还公开了一种高光谱图像分类装置,参考图3,图3是本申请实施例提供的一种高光谱图像分类装置的结构示意图,高光谱图像分类装置包括:
确定模块301,用于确定包括K类地物的目标高光谱图像的训练样本集合和待分类样本集合;目标高光谱图像由训练样本集合中的像素点和待分类样本集合中的像素点组成,训练样本集合包括每类地物的第一预设数目的像素点;
处理模块302,用于对训练样本集合和待分类样本集合进行如下处理:
根据训练样本集合生成K个训练子集,其中,将训练样本集合剔除一类地物的像素点,以得到一个训练子集;
利用K个训练子集和预设选择策略,从待分类样本集合中选取第二预设数目的满足预设选择策略的像素点,将第二预设数目的像素点加入训练样本集合以更新训练样本集合,并从待分类样本集合中删除对应的像素点以更新待分类样本集合;
分类模块303,用于利用更新后的训练样本集合进行模型训练以得到第一图像分类模型,并利用第一图像分类模型对更新后的待分类样本集合进行预测得到每个待分类样本的第一分类预测信息。
在一个可能的实施例中,处理模块302在利用K个训练子集和预设选择策略,从待分类样本集合中选取第二预设数目的满足预设选择策略的像素点时,具体执行:
分别利用K个训练子集进行模型训练以得到K个第二图像分类模型,分别利用K个第二图像分类模型对待分类样本集合进行预测以得到每个待分类样本的K个第二分类预测信息;
根据第二分类预测信息从待分类样本集合中选取第二预设数目的像素点。
在一个可能的实施例中,上述分别利用K个训练子集进行模型训练以得到K个第二图像分类模型包括:
分别利用K个训练子集和第二待训练模型进行模型训练,以得到K个第二图像分类模型,第二待训练模型包括协同表示分类模型、支持向量机、极限 学习机中的任意一个。
在一个可能的实施例中,第二分类预测信息包括预测类别以及对应的残差值;
处理模块在根据第二分类预测信息从待分类样本集合中选取第二预设数目的满足预设选择策略的像素点时,具体执行:
根据每个待分类样本的K个预测类别和K个残差值计算对应的分类置信度;
根据分类置信度从待分类样本集合中选取第二预设数目的分类置信度最低的像素点。
在一个可能的实施例中,分类模块303在利用更新后的训练样本集合进行模型训练以得到第一图像分类模型,具体执行:
利用更新后的训练样本集合和第一待训练模型进行模型训练,以得到第一图像分类模型,第一待训练模型包括支持向量机、稀疏表示分类模型、多元逻辑回归分类模型中的任意一个。
在一个可能的实施例中,分类模块303具体用于:
确定训练样本集合的更新次数;
更新次数小于次数阈值时,将更新后的训练样本集合和更新后的待分类样本集合作为新的训练样本集合和新的待分类样本集合,并对其进行如处理步骤的处理;
更新次数大于或等于次数阈值时,利用更新后的训练样本集合获取更新后的待分类样本集合中每个待分类样本的第一分类预测信息。
在一个可能的实施例中,分类模块303具体用于:
确定第一图像分类模型不满足收敛条件时,将更新后的训练样本集合和更新后的待分类样本集合作为新的训练样本集合和新的待分类样本集合,并对其进行如处理步骤的处理。
值得指出的是,其中,高光谱图像分类装置的具体功能实现方式可以参见上述高光谱图像分类方法的描述,这里不再进行赘述。高光谱图像分类装置中的各个单元或模块可以分别或全部合并为一个或若干个另外的单元或模块来 构成,或者其中的某个(些)单元或模块还可以再拆分为功能上更小的多个单元或模块来构成,这可以实现同样的操作,而不影响本申请的实施例的技术效果的实现。上述单元或模块是基于逻辑功能划分的,在实际应用中,一个单元(或模块)的功能也可以由多个单元(或模块)来实现,或者多个单元(或模块)的功能由一个单元(或模块)实现。
基于上述方法实施例以及装置实施例的描述,本申请实施例还提供一种高光谱图像分类设备。
请参见图4,是本申请实施例提供的一种高光谱图像分类设备的结构示意图。如图4所示,上述的高光谱图像分类装置可以应用于所述高光谱图像分类设备400,所述高光谱图像分类设备400可以包括:处理器401,网络接口404和存储器405,此外,所述高光谱图像分类设备400还可以包括:用户接口403,和至少一个通信总线402。其中,通信总线402用于实现这些组件之间的连接通信。其中,用户接口403可以包括显示屏(Display)、键盘(Keyboard),可选用户接口403还可以包括标准的有线接口、无线接口。网络接口404可选的可以包括标准的有线接口、无线接口(如WI-FI接口)。存储器405可以是高速RAM存储器,也可以是非不稳定的存储器(non-volatile memory),例如至少一个磁盘存储器。存储器405可选的还可以是至少一个位于远离前述处理器401的存储装置。如图4所示,作为一种计算机存储介质的存储器405中可以包括操作系统、网络通信模块、用户接口模块以及设备控制应用程序。
在图4所示的高光谱图像分类设备400中,网络接口404可提供网络通讯功能;而用户接口403主要用于为用户提供输入的接口;而处理器401可以用于调用存储器405中存储的设备控制应用程序,以实现上述实施例所述的高光谱图像分类方法的步骤。
应当理解,本申请实施例中所描述的高光谱图像分类设备400可执行前文所述高光谱图像分类方法,也可执行前文所述高光谱图像分类装置的描述,在此不再赘述。另外,对采用相同方法的有益效果描述,也不再进行赘述。
此外,这里需要指出的是:本申请实施例还提供了一种计算机存储介质,且所述计算机存储介质中存储有前文提及的高光谱图像分类装置所执行的计算机程序,且所述计算机程序包括程序指令,当处理器执行所述程序指令时,能够执行前文所述高光谱图像分类方法的描述,因此,这里将不再进行赘述。另外,对采用相同方法的有益效果描述,也不再进行赘述。对于本申请所涉及的计算机存储介质实施例中未披露的技术细节,请参照本申请方法实施例的描述。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程,是可以通过计算机程序来指令相关的硬件来完成,程序可存储于一计算机可读取存储介质中,该程序在执行时,可包括如上述各方法的实施例的流程。其中,存储介质可为磁碟、光盘、只读存储记忆体(Read-Only Memory,ROM)或随机存储记忆体(Random Access Memory,RAM)等。
以上所揭露的仅为本申请较佳实施例而已,当然不能以此来限定本申请之权利范围,因此依本申请权利要求所作的等同变化,仍属本申请所涵盖的范围。

Claims (10)

  1. 一种高光谱图像分类方法,其特征在于,包括:
    确定步骤,确定包括K类地物的目标高光谱图像的训练样本集合和待分类样本集合;所述目标高光谱图像由所述训练样本集合中的像素点和所述待分类样本集合中的像素点组成,所述训练样本集合包括每类地物的第一预设数目的像素点;
    处理步骤,对所述训练样本集合和所述待分类样本集合进行如下处理:
    根据所述训练样本集合生成K个训练子集,其中,将所述训练样本集合剔除一类地物的像素点,以得到一个训练子集;
    利用所述K个训练子集和预设选择策略,从所述待分类样本集合中选取第二预设数目的满足所述预设选择策略的像素点,将所述第二预设数目的像素点加入所述训练样本集合以更新所述训练样本集合,并从所述待分类样本集合中删除对应的像素点以更新所述待分类样本集合;
    分类步骤,利用更新后的所述训练样本集合进行模型训练以得到第一图像分类模型,并利用所述第一图像分类模型对更新后的所述待分类样本集合进行预测得到每个待分类样本的第一分类预测信息。
  2. 根据权利要求1所述的方法,其特征在于,所述利用所述K个训练子集和预设选择策略,从所述待分类样本集合中选取第二预设数目的满足所述预设选择策略的像素点,包括:
    分别利用所述K个训练子集进行模型训练以得到K个第二图像分类模型,分别利用所述K个第二图像分类模型对所述待分类样本集合进行预测以得到每个待分类样本的K个第二分类预测信息;
    根据所述第二分类预测信息从所述待分类样本集合中选取所述第二预设数目的像素点。
  3. 根据权利要求2所述的方法,其特征在于,所述分别利用所述K个训练子集进行模型训练以得到K个第二图像分类模型包括:
    分别利用所述K个训练子集和第二待训练模型进行模型训练,以得到所述K个第二图像分类模型,所述第二待训练模型包括协同表示分类模型、支 持向量机、极限学习机中的任意一个。
  4. 根据权利要求2或3所述的方法,其特征在于,所述第二分类预测信息包括预测类别以及对应的残差值;
    所述根据所述第二分类预测信息从所述待分类样本集合中选取所述第二预设数目的像素点,包括:
    根据每个所述待分类样本的K个所述预测类别和K个所述残差值计算对应的分类置信度;
    根据所述分类置信度从所述待分类样本集合中选取所述第二预设数目的分类置信度最低的像素点。
  5. 根据权利要求1至3任一项所述的方法,其特征在于,所述分类步骤包括:
    确定所述训练样本集合的更新次数;
    所述更新次数小于次数阈值时,将更新后的所述训练样本集合和更新后的所述待分类样本集合作为新的训练样本集合和新的待分类样本集合,并对其进行如所述处理步骤的处理;
    所述更新次数大于或等于所述次数阈值时,利用更新后的所述训练样本集合获取更新后的所述待分类样本集合中每个待分类样本的第一分类预测信息。
  6. 根据权利要求1至3任一项所述的方法,其特征在于,所述分类步骤包括:
    确定所述第一图像分类模型不满足收敛条件时,将更新后的所述训练样本集合和更新后的所述待分类样本集合作为新的训练样本集合和新的待分类样本集合,并对其进行如所述处理步骤的处理。
  7. 根据权利要求1至3任一项所述的方法,其特征在于,所述利用更新后的所述训练样本集合进行模型训练以得到第一图像分类模型,包括:
    利用更新后的所述训练样本集合和第一待训练模型进行模型训练,以得到所述第一图像分类模型,所述第一待训练模型包括支持向量机、稀疏表示分类模型、多元逻辑回归分类模型中的任意一个。
  8. 一种高光谱图像分类装置,其特征在于,包括:
    确定模块,用于确定包括K类地物的目标高光谱图像的训练样本集合和待分类样本集合;所述目标高光谱图像由所述训练样本集合中的像素点和所述待分类样本集合中的像素点组成,所述训练样本集合包括每类地物的第一预设数目的像素点;
    处理模块,用于对所述训练样本集合和所述待分类样本集合进行如下处理:
    根据所述训练样本集合生成K个训练子集,其中,将所述训练样本集合剔除一类地物的像素点,以得到一个训练子集;
    利用所述K个训练子集和预设选择策略,从所述待分类样本集合中选取第二预设数目的满足所述预设选择策略的像素点,将所述第二预设数目的像素点加入所述训练样本集合以更新所述训练样本集合,并从所述待分类样本集合中删除对应的像素点以更新所述待分类样本集合;
    分类模块,用于利用更新后的所述训练样本集合进行模型训练以得到第一图像分类模型,并利用所述第一图像分类模型对更新后的所述待分类样本集合进行预测得到每个待分类样本的第一分类预测信息。
  9. 一种高光谱图像分类设备,其特征在于,包括:处理器和存储器;
    所述处理器和存储器相连,其中,所述存储器用于存储程序代码,所述处理器用于调用所述程序代码,以执行如权利要求1-7任一项所述的高光谱图像分类方法。
  10. 一种计算机存储介质,其特征在于,所述计算机存储介质存储有计算机程序,所述计算机程序包括程序指令,所述程序指令当被处理器执行时,执行如权利要求1-7任一项所述的高光谱图像分类方法。
PCT/CN2020/125243 2020-10-23 2020-10-30 高光谱图像分类方法及相关设备 WO2022082848A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202011144440.6 2020-10-23
CN202011144440.6A CN112257603B (zh) 2020-10-23 2020-10-23 高光谱图像分类方法及相关设备

Publications (1)

Publication Number Publication Date
WO2022082848A1 true WO2022082848A1 (zh) 2022-04-28

Family

ID=74264832

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/125243 WO2022082848A1 (zh) 2020-10-23 2020-10-30 高光谱图像分类方法及相关设备

Country Status (2)

Country Link
CN (1) CN112257603B (zh)
WO (1) WO2022082848A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115497010A (zh) * 2022-09-30 2022-12-20 北京恒歌科技有限公司 一种基于深度学习的地理信息的识别方法及系统
CN115984559A (zh) * 2022-12-27 2023-04-18 二十一世纪空间技术应用股份有限公司 智能样本精选方法及相关装置
CN116912201A (zh) * 2023-07-13 2023-10-20 上海频准激光科技有限公司 一种光纤熔接质量预测系统

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113674142B (zh) * 2021-08-30 2023-10-17 国家计算机网络与信息安全管理中心 图像中目标物的消融方法、装置、计算机设备及介质
CN113933248A (zh) * 2021-10-12 2022-01-14 广东省科学院智能制造研究所 一种高光谱在线检测系统及方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180260657A1 (en) * 2016-11-01 2018-09-13 Shenzhen University Method and system for extracting and classifying features of hyperspectral remote sensing image
CN110033039A (zh) * 2019-04-10 2019-07-19 大连海事大学 一种基于目标约束检测的高光谱实时分类方法
CN110738171A (zh) * 2019-10-15 2020-01-31 大连海事大学 基于类特征迭代随机采样的高光谱图像谱空分类方法
CN111191736A (zh) * 2020-01-05 2020-05-22 西安电子科技大学 基于深度特征交叉融合的高光谱图像分类方法

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102208037B (zh) * 2011-06-10 2012-10-24 西安电子科技大学 基于高斯过程分类器协同训练算法的高光谱图像分类方法
CN104463247B (zh) * 2014-12-09 2016-08-31 山东大学 一种高光谱图像分类中光谱向量互相关特征的抽取方法
CN105809205B (zh) * 2016-03-31 2019-07-02 深圳大学 一种高光谱图像的分类方法及其系统
CN106096663B (zh) * 2016-06-24 2019-03-26 长春工程学院 一种基于稀疏异构分组的高光谱遥感影像的分类器构建方法
CN111222576B (zh) * 2020-01-08 2023-03-24 西安理工大学 一种高分辨率遥感图像分类方法
CN111414942B (zh) * 2020-03-06 2022-05-03 重庆邮电大学 一种基于主动学习和卷积神经网络的遥感图像分类方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180260657A1 (en) * 2016-11-01 2018-09-13 Shenzhen University Method and system for extracting and classifying features of hyperspectral remote sensing image
CN110033039A (zh) * 2019-04-10 2019-07-19 大连海事大学 一种基于目标约束检测的高光谱实时分类方法
CN110738171A (zh) * 2019-10-15 2020-01-31 大连海事大学 基于类特征迭代随机采样的高光谱图像谱空分类方法
CN111191736A (zh) * 2020-01-05 2020-05-22 西安电子科技大学 基于深度特征交叉融合的高光谱图像分类方法

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115497010A (zh) * 2022-09-30 2022-12-20 北京恒歌科技有限公司 一种基于深度学习的地理信息的识别方法及系统
CN115984559A (zh) * 2022-12-27 2023-04-18 二十一世纪空间技术应用股份有限公司 智能样本精选方法及相关装置
CN115984559B (zh) * 2022-12-27 2024-01-12 二十一世纪空间技术应用股份有限公司 智能样本精选方法及相关装置
CN116912201A (zh) * 2023-07-13 2023-10-20 上海频准激光科技有限公司 一种光纤熔接质量预测系统
CN116912201B (zh) * 2023-07-13 2024-03-08 上海频准激光科技有限公司 一种光纤熔接质量预测系统

Also Published As

Publication number Publication date
CN112257603B (zh) 2022-06-17
CN112257603A (zh) 2021-01-22

Similar Documents

Publication Publication Date Title
WO2022082848A1 (zh) 高光谱图像分类方法及相关设备
JP6813055B2 (ja) 画像処理システム、画像処理方法及びプログラム記憶媒体
JP6994588B2 (ja) 顔特徴抽出モデル訓練方法、顔特徴抽出方法、装置、機器および記憶媒体
JP2020501238A (ja) 顔検出トレーニング方法、装置及び電子機器
US10956785B2 (en) Methods, systems, and media for selecting candidates for annotation for use in training classifiers
CN110717953B (zh) 基于cnn-lstm组合模型的黑白图片的着色方法和系统
US9183450B2 (en) Inspection apparatus
CN110197716B (zh) 医学影像的处理方法、装置及计算机可读存储介质
US20150110386A1 (en) Tree-based Linear Regression for Denoising
CN107203558B (zh) 对象推荐方法和装置、推荐信息处理方法和装置
CN110910445A (zh) 一种物件尺寸检测方法、装置、检测设备及存储介质
CN111222557A (zh) 图像分类方法、装置、存储介质及电子设备
JP5244438B2 (ja) データ分類装置、データ分類方法、データ分類プログラムおよび電子機器
CN110366029B (zh) 一种视频间插入图像帧的方法、系统及电子设备
CN116524296A (zh) 设备缺陷检测模型的训练方法、装置和设备缺陷检测方法
JP2021527859A (ja) 深層領域拡張を使用した画像内の不規則形状のセグメント化
CN115760775A (zh) 番茄病害检测方法、装置、设备及存储介质
CN113408374B (zh) 基于人工智能的产量预估方法、装置、设备及存储介质
CN114897290A (zh) 业务流程的演化识别方法、装置、终端设备以及存储介质
JP6349477B1 (ja) 類似度判定プログラム
US20210183027A1 (en) Systems and methods for recognition of user-provided images
CN112070162A (zh) 多类别处理任务训练样本构建方法、设备及介质
CN112906805A (zh) 图像训练样本筛选和任务模型训练方法、装置及电子设备
CN106841054B (zh) 种子品种识别方法及装置
CN113469244B (zh) 小众app分类系统

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: 20958428

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20958428

Country of ref document: EP

Kind code of ref document: A1