CN114187530A - Remote sensing image change detection method based on neural network structure search - Google Patents

Remote sensing image change detection method based on neural network structure search Download PDF

Info

Publication number
CN114187530A
CN114187530A CN202111515285.9A CN202111515285A CN114187530A CN 114187530 A CN114187530 A CN 114187530A CN 202111515285 A CN202111515285 A CN 202111515285A CN 114187530 A CN114187530 A CN 114187530A
Authority
CN
China
Prior art keywords
remote sensing
sensing image
detection
change detection
block
Prior art date
Legal status (The legal status 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 status listed.)
Pending
Application number
CN202111515285.9A
Other languages
Chinese (zh)
Inventor
李阳阳
郭宣威
吴彬
陈茜
焦李成
尚荣华
李玲玲
马文萍
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xidian University
Original Assignee
Xidian University
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 Xidian University filed Critical Xidian University
Priority to CN202111515285.9A priority Critical patent/CN114187530A/en
Publication of CN114187530A publication Critical patent/CN114187530A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/086—Learning methods using evolutionary algorithms, e.g. genetic algorithms or genetic programming

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Physiology (AREA)
  • Image Analysis (AREA)

Abstract

The invention provides a remote sensing image change detection method based on neural network structure search, which comprises the following steps: acquiring a training, verifying and testing sample set; constructing a super neural network model; carrying out iterative training on the super neural network model; searching the trained super neural network model by adopting a genetic algorithm to obtain structure search parameters; constructing a remote sensing image change detection model based on the structure search parameters; carrying out iterative training on the remote sensing image change detection model; and obtaining a change detection result of the remote sensing image. The invention adopts the genetic algorithm to search the trained hyper-network neural model, and determines the structure of the remote sensing image change detection model through the structure search parameters obtained by searching, the structure of the remote sensing image change detection model can be highly matched with the characteristics of the remote sensing image, the remote sensing image characteristics are more fully extracted, and the change detection precision of the remote sensing image is effectively improved.

Description

Remote sensing image change detection method based on neural network structure search
Technical Field
The invention belongs to the technical field of image processing, relates to a remote sensing image change detection method, and particularly relates to a remote sensing image change detection method based on neural network structure search, which can be used in the fields of geological disaster monitoring, land cover investigation, city planning and the like.
Background
The remote sensing image is an image obtained by collecting and processing electromagnetic radiation energy of a ground object target by a remote sensing imaging instrument on a remote sensing platform. Two remote sensing images obtained by shooting the same place at two different moments can reflect the change condition of the ground object target. The remote sensing image change detection means detecting a changed area between two remote sensing images shot at the same place and different time. How to improve the change detection precision is the key and difficult point of the remote sensing image change detection.
The existing remote sensing image change detection method is mainly divided into a traditional change detection method and a change detection method based on deep learning. Conventional change detection methods include methods based on direct pixel comparison, methods based on feature levels, methods based on object-oriented image analysis, and the like. The detection precision of the methods on the remote sensing image is low. With the wide application of deep learning in various fields, a variety of change detection methods based on deep learning, such as a change detection method based on a twin neural network, a change detection method based on an autoencoder, and the like, have appeared. The change detection method based on deep learning obtains higher detection precision on the remote sensing image and is widely applied.
The neural network used in the existing remote sensing image change detection method based on deep learning is often designed manually. The structure of the artificially designed neural network cannot be well matched with the characteristics of the remote sensing image, the characteristics of the remote sensing image are not fully extracted, and the detection precision is reduced. For example, a patent application with publication number CN112613352A entitled "a twin network-based remote sensing image change detection method" discloses a twin convolutional neural network-based remote sensing image change detection method. The method comprises the steps of firstly extracting features of two images at different moments by using a coding end of a twin convolutional neural network, then decoding the features extracted by the coding end by using a decoding end of the twin convolutional neural network to obtain a feature map, then predicting class probability of each pixel in the feature map by using a LogSoftmax classifier to obtain a prediction probability map, and then binarizing the prediction probability map to obtain a change detection result map. The method has the defects that the structure of the twin convolutional neural network is designed manually, the characteristics of the remote sensing image cannot be well matched, the characteristic extraction capability of the remote sensing image is weak, and further improvement of the detection precision is influenced.
The principle of Neural network structure Search is that on a given super Neural network model, a certain optimization algorithm is used to Search out structure Search parameters by using data characteristics, and then the structure of the Neural network model is determined according to the structure Search parameters, the structure can be highly matched with the data characteristics, and the Neural network model using the structure can fully extract the characteristics of data and has higher detection precision, so that the detection precision of the remote sensing image change detection method can be improved if the Neural network structure Search is applied to the remote sensing image change detection method.
Disclosure of Invention
The invention aims to provide a remote sensing image change detection method based on neural network structure search aiming at the defects of the prior art, and the method is used for solving the technical problem of low detection precision in the prior art.
In order to achieve the purpose, the technical scheme adopted by the invention comprises the following steps:
(1) acquiring a training sample set, a verification sample set and a test sample set:
(1a) obtaining A, B T remote sensing images HA (changed appearance) with the time of R multiplied by R from a remote sensing image change detection data set1,···,HAt,···,HAT}、HB＝{HB1,···,HBt,···,HBTAnd a label image HL ═ HL of a region where HB HAs changed with respect to HA { HL }1,···,HLt,···,HLTWherein R is more than or equal to 256, T is more than or equal to 200, HAt、HBtThe t-th remote sensing image HL representing the A moment and the B momenttRepresents HBtWith respect to HAtA label image of the changed region;
(1b) HA each remote sensing image at A and B momentst、HBtAnd label image HLtRespectively cutting the image blocks into Z image blocks with the size of p multiplied by p to obtain a remote sensing image block set HA' ═ { HA ] corresponding to HA1′,···,HAt′,···,HA′TRemote sensing image block set HB' ═ HB corresponding to the remote sensing image blocks1′,···,HBt′,···,HB′TAnd HL' is ═ HL for the set of labeled image blocks corresponding to HL1′,···,HLt′,···,HL′TAnd (c) the step of (c) in which,
Figure BDA0003406707370000021
HAt' means HAtA set of corresponding remote sensing image blocks,
Figure BDA0003406707370000022
HBt' represents HBtA set of corresponding remote sensing image blocks,
Figure BDA0003406707370000031
HLt' represents HLtA set of corresponding labeled image blocks that are,
Figure BDA0003406707370000032
respectively represent HAt′、HBt′、HLt' the z-th image block;
(1c) more than half of the image blocks are randomly extracted from the image block sets HA ', HB ' and HL ' to form a training sample set, and half of the image blocks in the rest image blocks form a verification sample set, and the other half of the image blocks form a test sample set;
(2) constructing a remote sensing image change detection super neural network model M:
constructing a remote sensing image change detection super neural network model M comprising a feature extraction super network, a feature fusion network and a detection network which are connected in sequence, wherein the feature extraction super network comprises a convolution layer and U block blocks connected with the convolution layer in sequence, and each block comprises v unit blocks which are arranged in parallel and are composed of x convolution layers connected in sequence; the feature fusion network comprises a plurality of fusion blocks which are connected in sequence, wherein each fusion block comprises an upsampling layer, a concatemate layer and a convolutional layer which are connected in sequence; the detection network comprises a cascade coding layer and an upper sampling layer;
(3) carrying out iterative training on the remote sensing image change detection super neural network model M:
(3a) the number of initialization iterations is N, the maximum number of iterations is N, N is more than or equal to 10000, and the current remote sensing image change detection super neural network model is MnAnd let n equal to 1, Mn＝M；
(3b) Super neural network model M for detecting changes of training sample set as remote sensing imagenThe input of (a) is propagated forward:
(3b1) the method comprises the steps that a feature extraction super network carries out feature extraction on B random A-moment remote sensing image blocks and B random B-moment remote sensing image blocks of a training sample set to obtain feature maps of the B A-moment remote sensing image blocks and feature maps of the B-moment remote sensing image blocks, wherein B is more than or equal to 16, the feature map of each A-moment remote sensing image block and the feature map of each B-moment remote sensing image block comprise d sub-feature maps with equal quantity, and U is more than d and more than 2;
(3b2) the feature fusion network fuses the feature maps of the remote sensing image blocks at each A moment to obtain B fused feature maps at the A moments, and simultaneously fuses the feature maps of the remote sensing image blocks at each B moment to obtain B fused feature maps at the B moments;
(3c) the detection network encodes the fusion characteristic graph at each A moment and the corresponding fusion characteristic graph at each B moment, and performs up-sampling on the detection characteristic graph obtained by encoding to obtain B detection graphs;
(3d) calculating M by using a contrast loss function and through the b detection graphs and the corresponding b label image blocksnDetected loss value l ofnThen using a gradient descent method and detecting the loss value lnTo MnWeight parameter w ofnUpdating is carried out;
(3e) judging whether N is true or not, if so, obtaining a trained remote sensing image change detection super neural network model M', otherwise, making N be N +1, and executing the step (3 b);
(4) searching the trained remote sensing image change detection hyper-network neural model M' by adopting a genetic algorithm to obtain a structure search parameter y:
searching the trained remote sensing image change detection hyper-neural network model M' by adopting a genetic algorithm to obtain a structure search parameter
Figure BDA0003406707370000041
Wherein the content of the first and second substances,
Figure BDA0003406707370000042
indicates a value of
Figure BDA0003406707370000043
The gene position corresponding to the u block of (1),
Figure BDA0003406707370000044
is an integer;
(5) constructing a remote sensing image change detection model Y based on a structure search parameter Y:
constructing a remote sensing image change detection model Y comprising a feature extraction network, a feature fusion network and a detection network which are connected in sequence; the feature extraction network comprises a convolution layer and U ceil blocks connected in sequence, each ceil block consists of 3 convolution layers connected in sequence, and the convolution kernel size of the 3 convolution layers contained in the U ceil block
Figure BDA0003406707370000045
Is determined by the structure search parameter y:
Figure BDA0003406707370000046
Figure BDA0003406707370000047
Figure BDA0003406707370000048
the feature fusion network and the detection network have the same structures as those of the feature fusion network and the detection network in the step (2);
(6) carrying out iterative training on the remote sensing image change detection model Y:
(6a) the number of initialization iterations is O, the maximum number of iterations is O, O is more than or equal to 10000, and the current remote sensing image change detection model is YoAnd let o be 1, Yo＝Y；
(6b) Taking a training sample set as a remote sensing image change detection model YoThe input of (a) is propagated forward:
(6b1) the method comprises the steps that a feature extraction network carries out feature extraction on r A-moment remote sensing image blocks and r B-moment remote sensing image blocks of a training sample set at random to obtain feature maps of the r A-moment remote sensing image blocks and feature maps of the B-moment remote sensing image blocks, wherein r is more than or equal to 16, the feature map of each A-moment remote sensing image block and the feature map of each B-moment remote sensing image block comprise d sub-feature maps with equal quantity, and U is more than d and more than 2;
(6b2) the feature fusion network fuses the feature maps of the remote sensing image blocks at each A moment to obtain r fused feature maps at the A moment, and simultaneously fuses the feature maps of the remote sensing image blocks at each B moment to obtain r fused feature maps at the B moment;
(6c) the detection network encodes the fusion characteristic graph at each A moment and the corresponding fusion characteristic graph at each B moment, and performs up-sampling on the detection characteristic graph obtained by encoding to obtain r detection graphs;
(6d) calculating Y by using a contrast loss function and through r detection graphs and r label image blocks corresponding to the detection graphsoIs detected by the loss value qoThen using a gradient descent method and detecting the loss value qoFor YoWeight parameter theta ofoUpdating is carried out;
(6e) judging whether O is true or not, if so, obtaining a trained remote sensing image change detection model Y', otherwise, making O be O +1, and executing the step (6 b);
(7) obtaining a change detection result of the remote sensing image:
and carrying out forward propagation by taking the test sample set as the input of the remote sensing image change detection model Y' to obtain detection graphs of all test samples, assigning the pixel points with the pixel values smaller than 2 in the detection graphs to be 0, and assigning the rest pixel points to be 255 to obtain the detection results of all the test samples.
Compared with the prior art, the invention has the following advantages:
the invention adopts the genetic algorithm to search the trained remote sensing image change detection super-network neural model, and determines the structure of the remote sensing image change detection model through the structure search parameters obtained by searching, the structure of the remote sensing image change detection model can be highly matched with the characteristics of the remote sensing image, the defect of insufficient characteristics of the remote sensing image extracted by adopting a detection model designed manually in the prior art is avoided, and the change detection precision of the remote sensing image is effectively improved.
Drawings
FIG. 1 is a flow chart of an implementation of the present invention.
FIG. 2 is a schematic diagram of a characteristic extraction super-network of a remote sensing image change detection super-neural network model for extracting characteristics of a remote sensing image block and obtaining a characteristic diagram of the remote sensing image block.
FIG. 3 is a schematic diagram of a feature fusion network of a remote sensing image change detection hyper-neural network model fusing feature maps of remote sensing image blocks to obtain a fusion feature map.
Fig. 4 is a schematic diagram of a detection graph obtained by encoding A, B fusion characteristic graphs and then up-sampling the encoded detection characteristic graphs by a detection network of a remote sensing image change detection super-neural network model.
FIG. 5 is a schematic diagram of a feature extraction network of a remote sensing image change detection model performing feature extraction on a remote sensing image block to obtain a feature map of the remote sensing image block.
Detailed Description
The invention is described in further detail below with reference to the figures and specific examples.
Referring to fig. 1, the present invention includes the steps of:
step 1) obtaining a training sample set, a verification sample set and a test sample set:
(1a) obtaining A, B T remote sensing images HA (changed appearance) with the time of R multiplied by R from a remote sensing image change detection data set1,···,HAt,···,HAT}、HB＝{HB1,···,HBt,···,HBTAnd a label image HL ═ HL of a region where HB HAs changed with respect to HA { HL }1,···,HLt,···,HLTWherein R is more than or equal to 256, T is more than or equal to 200, HAt、HBtThe t-th remote sensing image HL representing the A moment and the B momenttRepresents HBtWith respect to HAtA label image of the changed region;
in this embodiment, the A, B time remote sensing image and the label image are acquired from the LEVIR-CD data set, where R is 1024 and T is 637;
(1b) HA each remote sensing image at A and B momentst、HBtAnd anLabel image HLtRespectively cutting the image blocks into Z image blocks with the size of p multiplied by p to obtain a remote sensing image block set HA' ═ { HA ] corresponding to HA1′,···,HAt′,···,HA′TRemote sensing image block set HB' ═ HB corresponding to the remote sensing image blocks1′,···,HBt′,···,HB′TAnd HL' is ═ HL for the set of labeled image blocks corresponding to HL1′,···,HLt′,···,HL′TAnd (c) the step of (c) in which,
Figure BDA0003406707370000061
HAt' means HAtA set of corresponding remote sensing image blocks,
Figure BDA0003406707370000071
HBt' represents HBtA set of corresponding remote sensing image blocks,
Figure BDA0003406707370000072
HLt' represents HLtA set of corresponding labeled image blocks that are,
Figure BDA0003406707370000073
respectively represent HAt′、HBt′、HLt' the z-th image block;
in the present embodiment, p is 256, Z is 16;
(1c) more than half of the image blocks are randomly extracted from the image block sets HA ', HB ' and HL ' to form a training sample set, and half of the image blocks in the rest image blocks form a verification sample set, and the other half of the image blocks form a test sample set;
step 2), constructing a remote sensing image change detection super neural network model M:
constructing a remote sensing image change detection super neural network model M comprising a feature extraction super network, a feature fusion network and a detection network which are connected in sequence, wherein:
the feature extraction super network comprises a convolution layer and U block blocks sequentially connected with the convolution layer, wherein each block comprises v unit blocks which are arranged in parallel and are formed by x convolution layers sequentially connected with each other; the feature fusion network comprises a plurality of fusion blocks which are connected in sequence, wherein each fusion block comprises an upsampling layer, a concatemate layer and a convolutional layer which are connected in sequence; the detection network comprises a cascade coding layer and an upper sampling layer;
in the present embodiment, U is 20, v is 4, and x is 3;
the convolution kernel size, the number of convolution kernels and the convolution step size of convolution layers contained in the feature extraction super network are respectively 3 x 3, 32 and 2;
the feature extraction super network comprises 20 block blocks, each block comprises 4 unit blocks, each unit block comprises 3 convolutional layers, the sizes of convolution kernels of the 3 convolutional layers contained in the first unit block are 1 multiplied by 1, 3 multiplied by 3 and 1 multiplied by 1 respectively, the sizes of convolution kernels of the 3 convolutional layers contained in the second unit block are 1 multiplied by 1, 5 multiplied by 5 and 1 multiplied by 1 respectively, the sizes of convolution kernels of the 3 convolutional layers contained in the third unit block are 1 multiplied by 1, 7 multiplied by 7 and 1 multiplied by 1 respectively, and the sizes of convolution kernels of the 3 convolutional layers contained in the fourth unit block are 3 multiplied by 3 respectively;
except that the convolution step length of the second convolution layer in 3 convolution layers contained in 4 unit blocks contained in the first, fifth, ninth and fifteenth blocks in the feature extraction super network is 2, the convolution step length of 3 convolution layers contained in 4 unit blocks contained in each block is 1;
in the feature extraction super network, the number of convolution kernels of 3 convolution layers contained in 4 unit blocks contained in first to fourth blocks is 64, the number of convolution kernels of 3 convolution layers contained in 4 unit blocks contained in fifth to eighth blocks is 128, the number of convolution kernels of 3 convolution layers contained in 4 unit blocks contained in ninth to sixteenth blocks is 256, and the number of convolution kernels of 3 convolution layers contained in 4 unit blocks contained in sixteenth to twenty blocks is 512;
the feature fusion network comprises 4 fusion blocks which are connected in sequence, the convolution kernel size, the number of convolution kernels and the convolution step length of a convolution layer contained in a first fusion block are respectively 3 x 3, 768 and 1, the convolution kernel size, the number of convolution kernels and the convolution step length of a convolution layer contained in a second fusion block are respectively 3 x 3, 384 and 1, the convolution kernel size, the number of convolution kernels and the convolution step length of a convolution layer contained in a third fusion block are respectively 3 x 3, 192 and 1, the convolution kernel size, the number of convolution kernels and the convolution step length of a convolution layer contained in a fourth fusion block are respectively 3 x 3, 96 and 1, and an upper sampling layer contained in each fusion block adopts a bilinear interpolation method;
the detection network, the up-sampling layer included in it adopts bilinear interpolation method, the calculation formula of coding included in the coding layer is:
Figure BDA0003406707370000081
FA and FB represent two characteristic diagrams input to the coding layer, the height, width and channel number of the two characteristic diagrams are respectively H, W and C, FO represents a characteristic diagram output by the coding layer, the height, width and channel number of the two characteristic diagrams are respectively H, W and 1, FA and FB represent two characteristic diagrams output by the coding layerh,w,cDenotes the value, FB, on the c channel of the h row and w column in the FAh,w,cRepresents the value on the w-th channel of the h-th row and w-th column in FB, FOh,w,1Represents the value on the 1 st channel of the w column of the h row in FB;
the function of the remote sensing image change detection super-neural network model is to provide a plurality of candidate remote sensing image change detection sub-network models for the subsequent search process, and when only one unit block is reserved in each block of the super-neural network extracted by the characteristics of the super-neural network model, a candidate remote sensing image change detection sub-network model can be formed; the super neural network model comprises 420A candidate subnetwork model, the subsequent search process being from 420Finding out a sub-network model with high detection precision from the candidate sub-network models;
step 3) carrying out iterative training on the remote sensing image change detection super neural network model M:
(3a) the number of initialization iterations is N, the maximum number of iterations is N, N is more than or equal to 10000, and the current remote sensing image change detection super neural network model is MnAnd let n equal to 1, Mn＝M；
In this embodiment, N is 30000, and N is 30000 for more sufficient model training;
(3b) super neural network model M for detecting changes of training sample set as remote sensing imagenThe input of (a) is propagated forward:
(3b1) the method comprises the steps that a feature extraction super network carries out feature extraction on B optional A-moment remote sensing image blocks and B B-moment remote sensing image blocks of a training sample set to obtain feature maps of the B A-moment remote sensing image blocks and feature maps of the B-moment remote sensing image blocks, wherein B is more than or equal to 16, and the number d of sub-feature maps contained in the feature map of each A-moment remote sensing image block and the feature map of each B-moment remote sensing image block is 5;
in this embodiment, b is 32, and 5 sub-feature maps included in the feature map of the remote sensing image block at each A, B time point are respectively from a convolutional layer, a fourth block, an eighth block, a sixteenth block, and a twentieth block of the feature extraction super network.
(3b2) The feature fusion network fuses the feature maps of the remote sensing image blocks at each A moment to obtain B fused feature maps at the A moments, and simultaneously fuses the feature maps of the remote sensing image blocks at each B moment to obtain B fused feature maps at the B moments;
(3c) the detection network encodes the fusion characteristic graph at each A moment and the corresponding fusion characteristic graph at each B moment, and performs up-sampling on the detection characteristic graph obtained by encoding to obtain B detection graphs;
(3d) calculating M by using a contrast loss function and through the b detection graphs and the corresponding b label image blocksnDetected loss value l ofnThen using a gradient descent method and detecting the loss value lnTo MnWeight parameter w ofnUpdating is carried out;
detecting the loss value l in step (3d)nAnd a weight parameter wnThe update formulas of (a) and (b) are respectively:
Figure BDA0003406707370000091
Figure BDA0003406707370000092
wherein b represents the number of the detection maps and the label image blocks, E represents the number of pixel points contained in each detection map and each label image block,
Figure BDA0003406707370000093
the pixel value of the e-th pixel point in the lambda detection graph is represented,
Figure BDA0003406707370000094
b represents the pixel value of the e-th pixel point in the lambda-th label image blockuRepresenting the number of pixel points with the pixel value of 0 in the b label image blocks, bcExpressing the number of pixel points with the pixel value of 1 in the b label image blocks, wherein Max expresses a maximum function; eta is learning rate, 1e-6 is not less than 0.1, lnRepresents the detection loss value, w' represents wnAs a result of the update, the result of the update,
Figure BDA0003406707370000101
representing the partial derivative calculation.
In this embodiment, the initial learning rate η is 0.001, when the network iterates to the 1 st ten thousand, the learning rate η is 0.0001, when the network iterates to the 2 nd ten thousand, the learning rate η is 0.00001, the optimizer function uses a random gradient to decrease the SGD, and the learning rate is attenuated when the network iterates to a certain number of times so as to prevent the loss function from falling into a local minimum;
(3e) judging whether N is true or not, if so, obtaining a trained remote sensing image change detection super neural network model M', otherwise, making N be N +1, and executing the step (3 b);
the iterative training of the remote sensing image change detection super neural network model is equivalent to the 4 contained in the remote sensing image change detection super neural network model20Carrying out iterative training on the candidate remote sensing image change detection sub-network model, wherein the trained remote sensing image change detection super neural network model comprises 420Trained candidate remote sensing image change detection sub-networkA model;
step 4) searching the trained remote sensing image change detection super-network neural model M' by adopting a genetic algorithm to obtain a structure search parameter y, wherein the implementation steps are as follows:
(4a) initializing genetic algorithm parameters: the iteration number is G, the maximum iteration number is G, the number of elite reservations is K, and the father population is P ═ P1,···,pi,···,pNPAnd the sub-population is S { }, wherein NP represents the number of the parent population individuals, and piThe represented ith parent population individual,
Figure BDA0003406707370000102
indicates that the u block corresponds to a value of
Figure BDA0003406707370000103
The gene position of (a) is determined,
Figure BDA0003406707370000104
is an integer, and let g be 1;
in the embodiment, G is 20, NP is 50, K is 2, G is 20 to reduce the search time, NP is 50, K is 2 to prevent premature phenomenon;
(4b) taking PC as cross probability and PM as variation probability to each father population individual piPerforming crossover operation and mutation operation, and adding new individuals generated by crossover and mutation into the sub-population S to obtain the sub-population S ═ { S }1,···,sα,···,sNSWhere NS denotes the number of sub-population individuals, sαRepresents the alpha sub-population individuals;
in this embodiment, PC is 0.4, PM is 0.1, the crossover operation adopts a single-point crossover mode, and the mutation operation adopts a basic mutation mode;
(4c) combining NP father population individuals in the father population P and NS sub-population individuals in the sub-population S into a temporary population T ═ { T }1,···,tβ,···,tNTWhere NT denotes the number of temporary population individuals, NT ═ NP + NS, tβIndicates the beta-th temporary population of individuals,
Figure BDA0003406707370000111
indicates that the u block corresponds to a value of
Figure BDA0003406707370000112
The gene location of (a);
(4d) obtaining individual t of each temporary populationβCorresponding remote sensing image change detection sub-network model Mβ：
Removing the first block of the characteristic extraction hyper-network in the trained remote sensing image change detection hyper-neural network model M
Figure BDA0003406707370000113
The unit blocks except the unit block are obtained to obtain tβCorresponding remote sensing image change detection sub-network model Mβ；
(4e) Using the verification sample set as a sub-network model MβThe input of the verification sample is transmitted forward to obtain a detection image of each verification sample, the pixel points with the pixel value less than 2 in each detection image are assigned to be 0, meanwhile, the rest pixel points are assigned to be 1 to obtain detection result images of all the verification samples, and M is calculated according to all the detection result images and the corresponding label image blocks thereofβF1 score, and then the F1 score as tβIs a fitness value fβWherein, the F1 score is a change detection result evaluation index;
the larger the F1 score is, the better the change detection result is, and the higher the detection precision of the remote sensing image change detection sub-network model is;
(4f) selecting K temporary population individuals with the maximum fitness value in the temporary population T, and selecting
Figure BDA0003406707370000114
Selecting NP-K of the rest NT-K temporary population individuals in the T for probability to form a new father population P' ═ { P1′,···,pi′,···,p′NP}；
(4g) Judging whether G is true or not, if so, judging the fitness value in PHighest individual U Gene locus
Figure BDA0003406707370000115
As the structure search parameter y, among others,
Figure BDA0003406707370000116
indicates that the u block corresponds to a value of
Figure BDA0003406707370000117
The gene position of (a) is determined,
Figure BDA0003406707370000118
otherwise, let g be g +1, and perform step (4 b).
The searching of the trained remote sensing image change detection super-network neural model by adopting a genetic algorithm is to find out a remote sensing image change detection sub-network model with high detection precision: individuals with low fitness can be adaptively eliminated in the iterative process of the genetic algorithm, and individuals with higher fitness are generated, and when the iteration is terminated, the individuals with highest fitness in the father population are the individuals with highest fitness in all the individuals generated in the whole iterative process; the fitness of the individual is equal to the F1 score of the corresponding sub-network model, and the higher the F1 score is, the higher the detection accuracy of the sub-network model is, the detection accuracy of the sub-network model corresponding to the individual with the highest fitness in the parent population is;
the reason why the U gene positions of the individuals with the highest fitness value in the father population are used as the structure search parameter y when the iteration is terminated is as follows: the structure of the remote sensing image change detection sub-network model with high detection precision can be highly matched with the characteristics of the remote sensing image, and the characteristics of the remote sensing image are more fully extracted; taking U gene positions of individuals with highest fitness in the father population as structure searching parameters, and then constructing a remote sensing image change detection model with the same structure as the sub-network model with high detection precision according to the structure searching parameters;
step 5) constructing a remote sensing image change detection model Y based on the structure search parameter Y:
constructing a feature extraction network comprising sequentially connected features,A remote sensing image change detection model Y of a feature fusion network and a detection network; the feature extraction network comprises a convolution layer and U ceil blocks connected in sequence, each ceil block consists of 3 convolution layers connected in sequence, and the convolution kernel size of the 3 convolution layers contained in the U ceil block
Figure BDA0003406707370000121
Is determined by the structure search parameter y:
Figure BDA0003406707370000122
Figure BDA0003406707370000123
Figure BDA0003406707370000124
the convolution kernel size, the number of convolution kernels and the convolution step size of convolution layers contained in the feature extraction network are respectively 3 x 3, 32 and 2;
in the feature extraction network, except that the convolution step size of the second convolution layer in the 3 convolution layers contained in the first ceil block, the fifth ceil block, the ninth ceil block and the fifteen ceil blocks is 2, the convolution step size of the 3 convolution layers contained in each ceil block is 1;
in the feature extraction network, the number of convolution kernels of 3 convolution layers contained in the first to fourth ceil blocks is 64, the number of convolution kernels of 3 convolution layers contained in the fifth to eight ceil blocks is 128, the number of convolution kernels of 3 convolution layers contained in the ninth to sixteen ceil blocks is 256, and the number of convolution kernels of 3 convolution layers contained in the sixteenth to twenty ceil blocks is 512;
the feature fusion network and the detection network have the same structures as those of the feature fusion network and the detection network in the step (2);
the structure of the remote sensing image change detection model constructed according to the structure search parameters is the same as the structure of the sub-network model corresponding to the individual with the highest fitness in the step (4), so that the remote sensing image change detection model can highly match the characteristics of the remote sensing image and fully extract the characteristics of the remote sensing image, and the change detection precision is improved;
step 6) carrying out iterative training on the remote sensing image change detection model Y:
(6a) the number of initialization iterations is O, the maximum number of iterations is O, O is more than or equal to 10000, and the current remote sensing image change detection model is YoAnd let o be 1, Yo＝Y；
In this embodiment, O30000 is designed to make model training more sufficient;
(6b) taking a training sample set as a remote sensing image change detection model YoThe input of (a) is propagated forward:
(6b1) the method comprises the steps that a feature extraction network carries out feature extraction on r A-moment remote sensing image blocks and r B-moment remote sensing image blocks of a training sample set at random to obtain feature maps of the r A-moment remote sensing image blocks and feature maps of the B-moment remote sensing image blocks, wherein r is more than or equal to 16, and the number d of sub-feature maps contained in the feature map of each A-moment remote sensing image block and the feature map of each B-moment remote sensing image block is 5;
in this embodiment, r is 32, and 5 sub-feature maps included in the feature map of the remote sensing image block at each time A, B are respectively from the convolutional layer, the fourth ceil block, the eighth ceil block, the sixteenth ceil block, and the twentieth block of the feature extraction network.
(6b2) The feature fusion network fuses the feature maps of the remote sensing image blocks at each A moment to obtain r fused feature maps at the A moment, and simultaneously fuses the feature maps of the remote sensing image blocks at each B moment to obtain r fused feature maps at the B moment;
(6c) the detection network encodes the fusion characteristic graph at each A moment and the corresponding fusion characteristic graph at each B moment, and performs up-sampling on the detection characteristic graph obtained by encoding to obtain r detection graphs;
(6d) calculating Y by using a contrast loss function and through r detection graphs and r label image blocks corresponding to the detection graphsoIs detected by the loss value qoThen is sampledBy gradient descent and by detecting the loss value qoFor YoWeight parameter theta ofoUpdating is carried out;
said detecting loss value q in step (6d)oAnd a weight parameter θoThe update formulas of (a) and (b) are respectively:
Figure BDA0003406707370000141
Figure BDA0003406707370000142
wherein r represents the number of the detection maps and the label image blocks, E represents the number of pixel points contained in each detection map and each label image block,
Figure BDA0003406707370000143
the pixel value of the e-th pixel point in the lambda detection graph is represented,
Figure BDA0003406707370000144
the pixel value r of the e-th pixel point in the lambda-th label image block is representeduRepresenting the number r of pixel points with pixel value 0 in r label image blockscExpressing the number of pixel points with the pixel value of 1 in the r label image blocks, wherein Max expresses a maximum function; gamma is learning rate, 1e-6 is not less than gamma not more than 0.1, qoRepresents the detection loss value, theta' represents the result after theta o is updated,
Figure BDA0003406707370000145
representing the partial derivative calculation.
In this embodiment, the initial learning rate γ is 0.001, when the network iterates to the 1 st ten thousand, the learning rate γ is 0.0001, when the network iterates to the 2 nd ten thousand, the learning rate γ is 0.00001, the optimizer function uses a random gradient to decrease the SGD, and the learning rate is attenuated when the network iterates to a certain number of times so as to prevent the loss function from falling into a local minimum;
(6e) judging whether O is true or not, if so, obtaining a trained remote sensing image change detection model Y', otherwise, making O be O +1, and executing the step (6 b);
step 7) obtaining a change detection result of the remote sensing image:
and carrying out forward propagation by taking the test sample set as the input of the remote sensing image change detection model Y' to obtain detection graphs of all test samples, assigning the pixel points with the pixel values smaller than 2 in the detection graphs to be 0, and assigning the rest pixel points to be 255 to obtain the detection results of all the test samples.
The technical effects of the present invention are further explained by simulation experiments as follows:
1. simulation conditions and contents:
637A, B time remote sensing images and corresponding label images are obtained from the LEVIR-CD data set and used for simulation experiments;
the simulation experiment is carried out on a server with a CPU model of Intel Xeon E5-2678 and a GPU model of GeForce GTX 1080. The operating system is a UBUNTU 18.04 system, the deep learning framework is PyTorch, and the programming language is Python 3.6;
the method is compared and simulated with the existing method for detecting the change of the remote sensing image based on the twin neural network. In order to quantitatively compare the change detection results, two remote sensing image change detection result evaluation indexes, namely F1 score and mIoU, are adopted in the experiment, the higher the two evaluation indexes are, the better the change detection result is, and the simulation result is shown in Table 1.
2. And (3) simulation result analysis:
TABLE 1
Figure BDA0003406707370000151
As can be seen from the table 1, compared with the existing twin neural network-based remote sensing image change detection method, F1 and mIoU are obviously improved, and the remote sensing image change detection model constructed by the method can be highly matched with the characteristics of a remote sensing image and fully extracts the characteristics of the remote sensing image, so that the change detection precision is improved, and the method has important practical significance.

Claims (5)

1. A remote sensing image change detection method based on neural network structure search is characterized by comprising the following steps:
(1) acquiring a training sample set, a verification sample set and a test sample set:
(1a) obtaining A, B T remote sensing images HA (changed appearance) with the time of R multiplied by R from a remote sensing image change detection data set1,…,HAt,…,HAT}、HB＝{HB1,…,HBt,…,HBTAnd a label image HL ═ HL of a region where HB HAs changed with respect to HA { HL }1,…,HLt,…,HLTWherein R is more than or equal to 256, T is more than or equal to 200, HAt、HBtThe t-th remote sensing image HL representing the A moment and the B momenttRepresents HBtWith respect to HAtA label image of the changed region;
(1b) HA each remote sensing image at A and B momentst、HBtAnd label image HLtCutting the image blocks into Z image blocks with the size of p × p respectively to obtain a remote sensing image block set HA '═ HA'1,…,HA′t,…,HA′TRemote sensing image block set HB ' ═ HB ' corresponding to the remote sensing image blocks '1,…,HB′t,…,HB′T} and a set of label image blocks HL ' l ═ HL ' corresponding to HL '1,…,HL′t,…,HL′TAnd (c) the step of (c) in which,
Figure FDA0003406707360000011
HA′trepresents HAtA set of corresponding remote sensing image blocks,
Figure FDA0003406707360000012
HB′trepresents HBtA set of corresponding remote sensing image blocks,
Figure FDA0003406707360000013
HL′tto denote HLtCorresponding label graphA set of the image blocks is obtained,
Figure FDA0003406707360000014
Figure FDA0003406707360000015
are respectively HA't、HB′t、HL′tThe z-th image block of (1);
(1c) more than half of the image blocks are randomly extracted from the image block sets HA ', HB ' and HL ' to form a training sample set, and half of the image blocks in the rest image blocks form a verification sample set, and the other half of the image blocks form a test sample set;
(2) constructing a remote sensing image change detection super neural network model M:
constructing a remote sensing image change detection super neural network model M comprising a feature extraction super network, a feature fusion network and a detection network which are connected in sequence, wherein the feature extraction super network comprises a convolution layer and U block blocks connected with the convolution layer in sequence, and each block comprises v unit blocks which are arranged in parallel and are composed of x convolution layers connected in sequence; the feature fusion network comprises a plurality of fusion blocks which are connected in sequence, wherein each fusion block comprises an upsampling layer, a concatemate layer and a convolutional layer which are connected in sequence; the detection network comprises a cascade coding layer and an upper sampling layer;
(3) carrying out iterative training on the remote sensing image change detection super neural network model M:
(3a) the number of initialization iterations is N, the maximum number of iterations is N, N is more than or equal to 10000, and the current remote sensing image change detection super neural network model is MnAnd let n equal to 1, Mn＝M；
(3b) Super neural network model M for detecting changes of training sample set as remote sensing imagenThe input of (a) is propagated forward:
(3b1) the method comprises the steps that a feature extraction super network carries out feature extraction on B random A-moment remote sensing image blocks and B random B-moment remote sensing image blocks of a training sample set to obtain feature maps of the B A-moment remote sensing image blocks and feature maps of the B-moment remote sensing image blocks, wherein B is more than or equal to 16, the feature map of each A-moment remote sensing image block and the feature map of each B-moment remote sensing image block comprise d sub-feature maps with equal quantity, and U is more than d and more than 2;
(3b2) the feature fusion network fuses the feature maps of the remote sensing image blocks at each A moment to obtain B fused feature maps at the A moments, and simultaneously fuses the feature maps of the remote sensing image blocks at each B moment to obtain B fused feature maps at the B moments;
(3c) the detection network encodes the fusion characteristic graph at each A moment and the corresponding fusion characteristic graph at each B moment, and performs up-sampling on the detection characteristic graph obtained by encoding to obtain B detection graphs;
(3d) calculating M by using a contrast loss function and through the b detection graphs and the corresponding b label image blocksnDetected loss value l ofnThen using a gradient descent method and detecting the loss value lnTo MnWeight parameter w ofnUpdating is carried out;
(3e) judging whether N is true or not, if so, obtaining a trained remote sensing image change detection super neural network model M', otherwise, making N be N +1, and executing the step (3 b);
(4) searching the trained remote sensing image change detection hyper-network neural model M' by adopting a genetic algorithm to obtain a structure search parameter y:
searching the trained remote sensing image change detection hyper-neural network model M' by adopting a genetic algorithm to obtain a structure search parameter
Figure FDA0003406707360000031
Wherein the content of the first and second substances,
Figure FDA0003406707360000032
indicates a value of
Figure FDA0003406707360000033
The gene position corresponding to the u block of (1),
Figure FDA0003406707360000034
Figure FDA0003406707360000035
is an integer;
(5) constructing a remote sensing image change detection model Y based on a structure search parameter Y:
constructing a remote sensing image change detection model Y comprising a feature extraction network, a feature fusion network and a detection network which are connected in sequence; the feature extraction network comprises a convolution layer and U ceil blocks connected in sequence, each ceil block consists of 3 convolution layers connected in sequence, and the convolution kernel size of the 3 convolution layers contained in the U ceil block
Figure FDA0003406707360000036
Is determined by the structure search parameter y:
Figure FDA0003406707360000037
Figure FDA0003406707360000038
Figure FDA0003406707360000039
the feature fusion network and the detection network have the same structures as those of the feature fusion network and the detection network in the step (2);
(6) carrying out iterative training on the remote sensing image change detection model Y:
(6a) the number of initialization iterations is O, the maximum number of iterations is O, O is more than or equal to 10000, and the current remote sensing image change detection model is YoAnd let o be 1, Yo＝Y；
(6b) Taking a training sample set as a remote sensing image change detection model YoThe input of (a) is propagated forward:
(6b1) the method comprises the steps that a feature extraction network carries out feature extraction on r A-moment remote sensing image blocks and r B-moment remote sensing image blocks of a training sample set at random to obtain feature maps of the r A-moment remote sensing image blocks and feature maps of the B-moment remote sensing image blocks, wherein r is more than or equal to 16, the feature map of each A-moment remote sensing image block and the feature map of each B-moment remote sensing image block comprise d sub-feature maps with equal quantity, and U is more than d and more than 2;
(6b2) the feature fusion network fuses the feature maps of the remote sensing image blocks at each A moment to obtain r fused feature maps at the A moment, and simultaneously fuses the feature maps of the remote sensing image blocks at each B moment to obtain r fused feature maps at the B moment;
(6c) the detection network encodes the fusion characteristic graph at each A moment and the corresponding fusion characteristic graph at each B moment, and performs up-sampling on the detection characteristic graph obtained by encoding to obtain r detection graphs;
(6d) calculating Y by using a contrast loss function and through r detection graphs and r label image blocks corresponding to the detection graphsoIs detected by the loss value qoThen using a gradient descent method and detecting the loss value qoFor YoWeight parameter theta ofoUpdating is carried out;
(6e) judging whether O is true or not, if so, obtaining a trained remote sensing image change detection super neural network model Y', otherwise, making O be O +1, and executing the step (6 b);
(7) obtaining a change detection result of the remote sensing image:
and carrying out forward propagation by taking the test sample set as the input of the remote sensing image change detection model Y' to obtain detection graphs of all test samples, assigning the pixel points with the pixel values smaller than 2 in the detection graphs to be 0, and assigning the rest pixel points to be 255 to obtain the detection results of all the test samples.
2. The method for detecting changes in remote sensing images based on neural network structure search according to claim 1, wherein the remote sensing image change detection super neural network model M in step (2) is a model in which:
a feature extraction super network, the convolution kernel size of convolution layers included in the feature extraction super network is 3 × 3, the number of block blocks included in the feature extraction super network is 20, the number of unit blocks included in each block is 4, the number of convolution layers included in each unit block is 3, the convolution kernel sizes of 3 convolution layers included in a first unit block are 1 × 1, 3 × 3 and 1 × 1 respectively, the convolution kernel sizes of 3 convolution layers included in a second unit block are 1 × 1, 5 × 5 and 1 × 1 respectively, the convolution kernel sizes of 3 convolution layers included in a third unit block are 1 × 1, 7 × 7 and 1 × 1 respectively, and the convolution kernel sizes of 3 convolution layers included in a fourth unit block are all 3 × 3;
the feature fusion network includes 4 fusion blocks, and each fusion block includes convolution layers each having a convolution kernel size of 3 × 3.
3. The method for detecting changes in remotely sensed images based on neural network structure search as claimed in claim 1, wherein said detection loss value l in step (3d)nAnd a weight parameter wnThe update formulas of (a) and (b) are respectively:
Figure FDA0003406707360000051
Figure FDA0003406707360000052
wherein b represents the number of the detection maps and the label image blocks, E represents the number of pixel points contained in each detection map and each label image block,
Figure FDA0003406707360000053
the pixel value of the e-th pixel point in the lambda detection graph is represented,
Figure FDA0003406707360000054
b represents the pixel value of the e-th pixel point in the lambda-th label image blockuRepresenting the number of pixel points with the pixel value of 0 in the b label image blocks, bcRepresenting a pixel having a pixel value of 1 in b labeled image blocksThe number of prime points, Max, represents a maximum function; eta is learning rate, 1e-6 is not less than 0.1, lnRepresents the detection loss value, w' represents wnAs a result of the update, the result of the update,
Figure FDA0003406707360000055
representing the partial derivative calculation.
4. The remote sensing image change detection method based on neural network structure search of claim 1, characterized in that, the step (4) of searching the trained remote sensing image change detection hyper-neural network model M' by using genetic algorithm comprises the following steps:
(4a) initializing genetic algorithm parameters: the iteration number is G, the maximum iteration number is G, the number of elite reservations is K, and the father population is P ═ P1,…,pi,…,pNPAnd the sub-population is S { }, wherein NP represents the number of the parent population individuals, and piThe represented ith parent population individual,
Figure FDA0003406707360000056
Figure FDA0003406707360000057
indicates that the u block corresponds to a value of
Figure FDA0003406707360000058
The gene position of (a) is determined,
Figure FDA0003406707360000059
Figure FDA00034067073600000510
is an integer, and let g be 1;
(4b) taking PC as cross probability and PM as variation probability to each father population individual piPerforming crossover operation and mutation operation, and adding new individuals generated by crossover and mutation into the sub-population S to obtain the sub-population S ═ { S }1,…,sα,…,sNSWhere NS denotes the number of sub-population individuals, sαRepresents the alpha sub-population individuals;
(4c) combining NP father population individuals in the father population P and NS sub-population individuals in the sub-population S into a temporary population T ═ { T }1,…,tβ,…,tNTWhere NT denotes the number of temporary population individuals, NT ═ NP + NS, tβIndicates the beta-th temporary population of individuals,
Figure FDA0003406707360000061
Figure FDA0003406707360000062
indicates that the u block corresponds to a value of
Figure FDA0003406707360000063
The gene position of (a) is determined,
Figure FDA0003406707360000064
Figure FDA0003406707360000065
is an integer;
(4d) obtaining individual t of each temporary populationβCorresponding remote sensing image change detection sub-network model Mβ：
Removing the first block of the characteristic extraction hyper-network in the trained remote sensing image change detection hyper-neural network model M
Figure FDA0003406707360000066
The unit blocks except the unit block are obtained to obtain tβCorresponding remote sensing image change detection sub-network model Mβ；
(4e) Using the verification sample set as a sub-network model MβThe input of the verification sample is transmitted forward to obtain a detection image of each verification sample, the pixel points with the pixel values smaller than 2 in each detection image are assigned to be 0, and the rest pixel points are assigned to be 1 to obtain all verificationsThe detection result graphs of the samples, and M is calculated through all the detection result graphs and the corresponding label image blocksβF1 score, and then the F1 score as tβIs a fitness value fβWherein, the F1 score is a change detection result evaluation index;
(4f) selecting K temporary population individuals with the maximum fitness value in the temporary population T, and selecting
Figure FDA0003406707360000067
Selecting NP-K in the rest NT-K temporary population individuals in the T for probability to form a new parent population P '═ { P'1,…,p′i,…,p′NP}；
(4g) Judging whether G is satisfied or not, if so, determining the U gene positions of the individual with the highest fitness value in P
Figure FDA00034067073600000612
As the structure search parameter y, among others,
Figure FDA0003406707360000068
indicates that the u block corresponds to a value of
Figure FDA0003406707360000069
The gene position of (a) is determined,
Figure FDA00034067073600000610
Figure FDA00034067073600000611
otherwise, let g be g +1, and perform step (4 b).
5. The method for detecting changes in remotely sensed images based on neural network structure search as claimed in claim 1, wherein said detection loss value q in step (6d) is determined by said methodoAnd a weight parameter θoThe update formulas of (a) and (b) are respectively:
Figure FDA0003406707360000071
Figure FDA0003406707360000072
wherein r represents the number of the detection maps and the label image blocks, E represents the number of pixel points contained in each detection map and each label image block,
Figure FDA0003406707360000073
the pixel value of the e-th pixel point in the lambda detection graph is represented,
Figure FDA0003406707360000074
the pixel value r of the e-th pixel point in the lambda-th label image block is representeduRepresenting the number r of pixel points with pixel value 0 in r label image blockscExpressing the number of pixel points with the pixel value of 1 in the r label image blocks, wherein Max expresses a maximum function; gamma is learning rate, 1e-6 is not less than gamma not more than 0.1, qoRepresents the detection loss value, and theta' represents thetaoAs a result of the update, the result of the update,
Figure FDA0003406707360000075
representing the partial derivative calculation.
CN202111515285.9A 2021-12-13 2021-12-13 Remote sensing image change detection method based on neural network structure search Pending CN114187530A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111515285.9A CN114187530A (en) 2021-12-13 2021-12-13 Remote sensing image change detection method based on neural network structure search

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111515285.9A CN114187530A (en) 2021-12-13 2021-12-13 Remote sensing image change detection method based on neural network structure search

Publications (1)

Publication Number Publication Date
CN114187530A true CN114187530A (en) 2022-03-15

Family

ID=80604638

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111515285.9A Pending CN114187530A (en) 2021-12-13 2021-12-13 Remote sensing image change detection method based on neural network structure search

Country Status (1)

Country Link
CN (1) CN114187530A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115019174A (en) * 2022-06-10 2022-09-06 西安电子科技大学 Up-sampling remote sensing image target identification method based on pixel recombination and attention
CN116310851A (en) * 2023-05-26 2023-06-23 中国科学院空天信息创新研究院 Remote sensing image change detection method

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115019174A (en) * 2022-06-10 2022-09-06 西安电子科技大学 Up-sampling remote sensing image target identification method based on pixel recombination and attention
CN115019174B (en) * 2022-06-10 2023-06-16 西安电子科技大学 Up-sampling remote sensing image target recognition method based on pixel recombination and attention
CN116310851A (en) * 2023-05-26 2023-06-23 中国科学院空天信息创新研究院 Remote sensing image change detection method
CN116310851B (en) * 2023-05-26 2023-08-15 中国科学院空天信息创新研究院 Remote sensing image change detection method

Similar Documents

Publication Publication Date Title
CN113919448B (en) Method for analyzing influence factors of carbon dioxide concentration prediction at any time-space position
CN111612066B (en) Remote sensing image classification method based on depth fusion convolutional neural network
CN111259930B (en) General target detection method of self-adaptive attention guidance mechanism
CN110188765B (en) Image semantic segmentation model generation method, device, equipment and storage medium
CN110287927B (en) Remote sensing image target detection method based on depth multi-scale and context learning
CN112634292B (en) Asphalt pavement crack image segmentation method based on deep convolutional neural network
CN109784283B (en) Remote sensing image target extraction method based on scene recognition task
CN108446470B (en) Medical facility accessibility analysis method based on vehicle trajectory data and population distribution
CN112418212B (en) YOLOv3 algorithm based on EIoU improvement
CN114187530A (en) Remote sensing image change detection method based on neural network structure search
CN110246148B (en) Multi-modal significance detection method for depth information fusion and attention learning
CN113780296A (en) Remote sensing image semantic segmentation method and system based on multi-scale information fusion
JP2020126613A (en) Method for automatically evaluating labeling reliability of training image for use in deep learning network to analyze image, and reliability-evaluating device using the same
CN113486981A (en) RGB image classification method based on multi-scale feature attention fusion network
CN111210869A (en) Protein cryoelectron microscope structure analysis model training method and analysis method
CN110659601B (en) Depth full convolution network remote sensing image dense vehicle detection method based on central point
CN112668809B (en) Method for establishing autism children rehabilitation effect prediction model
CN110852369B (en) Hyperspectral image classification method combining 3D/2D convolutional network and adaptive spectrum unmixing
CN109614883A (en) A kind of tight sand crack intelligent identification Method based on convolutional neural networks
CN111985325A (en) Aerial small target rapid identification method in extra-high voltage environment evaluation
CN110210550A (en) Image fine granularity recognition methods based on integrated study strategy
CN113177456A (en) Remote sensing target detection method based on single-stage full convolution network and multi-feature fusion
CN116934780A (en) Deep learning-based electric imaging logging image crack segmentation method and system
CN114529552A (en) Remote sensing image building segmentation method based on geometric contour vertex prediction
CN113313000A (en) Gas-liquid two-phase flow intelligent identification method based on optical image

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination