CN112733725B - Hyperspectral image change detection method based on multistage cyclic convolution self-coding network - Google Patents

Hyperspectral image change detection method based on multistage cyclic convolution self-coding network Download PDF

Info

Publication number
CN112733725B
CN112733725B CN202110036358.XA CN202110036358A CN112733725B CN 112733725 B CN112733725 B CN 112733725B CN 202110036358 A CN202110036358 A CN 202110036358A CN 112733725 B CN112733725 B CN 112733725B
Authority
CN
China
Prior art keywords
network
input
convolution
samples
gate
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.)
Active
Application number
CN202110036358.XA
Other languages
Chinese (zh)
Other versions
CN112733725A (en
Inventor
董文倩
杨宇菲
曲家慧
肖嵩
李云松
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 CN202110036358.XA priority Critical patent/CN112733725B/en
Publication of CN112733725A publication Critical patent/CN112733725A/en
Application granted granted Critical
Publication of CN112733725B publication Critical patent/CN112733725B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/13—Satellite images
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • 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
    • G06N3/047—Probabilistic or stochastic 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • 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
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • 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

Landscapes

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

Abstract

The application belongs to the technical field of image processing, and discloses a hyperspectral image change detection method based on a multistage cyclic convolution self-coding network, which comprises the steps of inputting two double-phase hyperspectral images acquired in the same area at different time, and preprocessing the images; generating a sample set, and selecting a training sample set and a test sample set; constructing a convolution self-encoder, introducing an attention mechanism, and forming a multi-stage convolution self-encoding sub-network by two parallel convolution self-encoders with attention modules; constructing a long-period memory neural network, wherein three parallel long-period memory neural networks form a circulating sub-network; training the built network model to obtain network parameters suitable for the model; and inputting all samples into a trained network for discrimination to obtain a final change detection result graph. The multi-level features extracted by the application give consideration to both high-level semantic information and low-level high-resolution detail information, and are beneficial to improving the precision of a change detection method.

Description

Hyperspectral image change detection method based on multistage cyclic convolution self-coding network
Technical Field
The application belongs to the technical field of image processing, and particularly relates to a hyperspectral image change detection method based on a multistage cyclic convolution self-coding network.
Background
At present: with the development of imaging technology, hyperspectral images are widely used in the remote sensing field. Because the hyperspectral image contains abundant spectral information and has high spectral resolution, the hyperspectral image can better describe similar objects compared with the traditional multispectral and panchromatic images, and becomes an important data source for human beings to accurately know the change of the earth ecological environment in real time. The change detection is used as an effective means for observing the surface change, is widely applied to the fields of land coverage rate change, environment monitoring, disaster monitoring evaluation, military reconnaissance and the like, and provides an important basis for researching the interaction between human activities and an earth ecological system.
Change detection has become a hotspot research direction, and many scholars propose different change detection methods, which demonstrate excellent performance. These methods are generally classified into three types, namely, an image algebra method, an image transformation method and a classification detection method. In order to quickly and simply detect the change of the ground object, the image algebra method measures the change intensity of the ground object by direct mathematical operation on corresponding wave bands, including a ratio method, a difference method, an image regression method and the like. Although the computational complexity is low, these methods have poor anti-interference capability and are susceptible to radiation and noise. The image transformation method transforms the remote sensing image into a specific feature space in which the change region can be highlighted, suppressing unchanged regions such as Principal Component Analysis (PCA), slow Feature Analysis (SFA), etc. While this approach eliminates data redundancy, data imbalance can affect its performance. Classification detection requires a priori knowledge to train the classifier, which can provide complete "from-to" information, but the performance of such methods depends on the classification accuracy of the classifier.
While the above methods all accomplish the task of detecting changes, many of them are not designed for hyperspectral images and there are limitations to applying them directly to hyperspectral images. Traditional methods simply extract shallow features manually, which limits their ability to express advanced features, thereby losing some detail. In addition, the hyperspectral image has very high spectral resolution and rich spectral detail information, and the method designed for the multispectral or panchromatic image cannot be used for aiming at the characteristics of the hyperspectral or panchromatic image, so that the information is fully utilized to obtain higher detection precision.
In recent years, deep learning has achieved remarkable results in change detection, and by using various deep learning models, the empty spectrum features of images are automatically extracted to mine deep features of hyperspectral images, so that the performance of change detection is improved. However, the existing deep learning model only extracts rough deep features, and the deep features usually ignore space details, which are unfavorable for obtaining details such as edges of images. Meanwhile, the Convolutional Neural Network (CNN) regards the multi-phase images as mutually independent inputs, ignoring the time dependence between them.
Through the above analysis, the problems and defects existing in the prior art are as follows: the existing deep learning model only extracts deep features containing abstract semantic information, but the deep features are rough and low in resolution, and the deep features usually ignore space details and are unfavorable for obtaining details such as edges in images. Meanwhile, the Convolutional Neural Network (CNN) regards multi-phase images as mutually independent inputs, and ignores the time dependency relationship between the images, so that the characteristics of the images are not fully utilized, and the performance of a change detection method is affected.
The difficulty of solving the problems and the defects is as follows: for a classical CNN model, only the deepest abstract semantic information is generally utilized, no proper model is used for extracting multi-level features, and how to further utilize the extracted multi-level features is also an openness problem; the problem of detecting the change is handled by using the traditional CNN network, special processing is needed for input to obtain the time sequence information, and in addition, how to efficiently extract the useful time sequence characteristics for time sequence input needs to be further studied.
The meaning of solving the problems and the defects is as follows: the neural network has higher resolution of the shallow features, contains more details, and is favorable for obtaining richer space-spectrum features by reasonably utilizing the shallow features, so that a change detection method can obtain a better change detection result diagram; the change detection task is regarded as being different from the simple two-classification problem, the time dimension information is fully utilized, the time sequence characteristics are effectively mined, and the accuracy of the change detection algorithm is improved.
Disclosure of Invention
Aiming at the problems existing in the prior art, a hyperspectral image change detection method based on a multistage cyclic convolution self-coding network is provided
The application is realized in such a way that the hyperspectral image change detection method based on the multistage cyclic convolution self-coding network comprises the following steps:
two double-phase hyperspectral images acquired at different time in the same region are input, and the images are preprocessed, so that the input data obey the same distribution, and the model can be effectively converged better and faster in the training process;
generating a sample set, reasonably selecting a training sample set and a test sample set, solving the problem of unbalanced samples, and enabling more unchanged samples to participate in network training, thereby being beneficial to obtaining better network parameters;
constructing a convolution self-encoder, introducing an attention mechanism, forming a multi-stage convolution self-encoding sub-network by two parallel convolution self-encoders with attention modules, acquiring multi-stage space-spectrum characteristics corresponding to each sample, acquiring more abundant information, and being beneficial to improving the performance of an algorithm;
the method comprises the steps of constructing a long-period memory neural network, forming a circulating sub-network by three parallel long-period memory neural networks, taking space-spectrum characteristics of the same scale of two time phases obtained by the multi-stage convolution self-coding sub-network as input to obtain space-spectrum-time characteristics, classifying by using a full-connection layer, and obtaining time sequence characteristics more efficiently by using a variant of the traditional circulating neural network, thereby being beneficial to obtaining a precisely-changed detection result graph;
performing supervised training on a built network model formed by a multi-stage convolution self-coding sub-network and a circulation sub-network together to obtain network parameters suitable for the model;
and sequentially placing all samples of the hyperspectral image into a trained network for discrimination to obtain a final change detection result graph.
Furthermore, the hyperspectral image change detection method based on the multistage cyclic convolution self-coding network inputs two double-phase hyperspectral images acquired in the same region at different timesAnd carrying out maximum and minimum normalization on the image, wherein the normalization formula is as follows:
wherein x is i Representing a picture element, x, in a hyperspectral image max And x min Respectively represent the maximum value and the minimum value of the hyperspectral image,is one pixel after normalization.
Further, a sample set is generated, and a training sample set I is selected t And test sample set I e ；
(1) The three-dimensional hyperspectral data are recorded as H multiplied by W multiplied by C, wherein H and W are the height and the width of a hyperspectral image respectively, and C is the band number of the hyperspectral image; for the hyperspectral image of the double phases, taking each pixel point of the image as the center, selecting a data block with the size of 5 multiplied by C as a sample, wherein the set of all the samples is the generated sample set;
(2) Calculating the proportion a and b of the changed sample and the unchanged sample in the reference image to the total sample respectively, and training a network by using more unchanged samples;
(3) Randomly selecting 20% of total samples as training sample set I t The remaining 80% is used as test sample set I e The ratio of the changed sample to the unchanged sample in the training sample set and the test sample set is a to b.
Further, a multistage convolution self-coding sub-network is constructed, two attention mechanism modules are introduced into the convolution self-coder, and multistage space-spectrum characteristics corresponding to each sample are obtained;
(1) The constructed convolutional self-encoder consists of two parts, encoder and decoder, for a pair of input samplesAndthe encoder comprises three convolution layers with a convolution kernel size of 3×3 and a step size of 2, so that the encoder obtains feature map sizes of 5×5× 0256,3 ×13× 2256,1 ×31×512, respectively; the decoder comprises three deconvolution layers with a convolution kernel size of 3 x 3 and a step size of 2, resulting in a feature map size of 3 x 256,5 x 5 x 256 and reconstructed samples +.>And->Batch normalization and ReLU activation function processing is used after each convolutional layer;
(2) Attention modules are introduced into the convolution self-encoder to improve the feature extraction effect, the attention modules are formed by connecting spatial attention modules and channel attention modules in series, and are embedded into a residual error module, and the specific structure of the adopted convolution self-encoder is as follows: first convolution layer- & gtattention module- & gtsecond convolution layer- & gtthird convolution layer- & gtfirst deconvolution layer- & gtsecond deconvolution layer- & gtattention module- & gtthird deconvolution layer;
(3) Two parallel convolutional self-encoders form a multi-stage convolutional self-encoding sub-network for a pair of input samplesAnd->After the multi-level features are extracted from the encoder through convolution respectively, the features with the same size in the same convolution self-encoder are cascaded to output three pairs of multi-level space-spectrum features with different sizes> And->The feature sizes obtained were 5×5× 512,3 ×3× 512,1 ×1×512, respectively.
Further, a cyclic subnetwork is constructed, three identical long-short-term memory neural networks are adopted, and the multi-stage convolution self-coding subnetwork is respectively adopted to obtainAnd->As the time sequence input of three cyclic neural networks, extracting the time dependency relationship of the three cyclic neural networks, and classifying the three cyclic neural networks by using a full-connection layer;
(1) The method comprises the steps of constructing a long-term and short-term memory neural network, wherein the long-term and short-term memory neural network consists of two cell units, and each cell unit comprises a forgetting gate, an input gate and an output gate;
first calculate the forgetting door f t The forgetting gate controls the forgetting degree of the content of the existing memory cell, namely decides which information is reserved and discarded, and the calculation formula of the forgetting gate is as follows:
wherein h is t-1 Is the hidden state of the last moment, W fi And W is fh Weight matrix of input-forget gate and hidden state-forget gate, b f Representing the bias of the forgetting gate, sigma is the activation function, normalize the output value to [0,1]；
Then calculate the input gate i t The input gate is responsible for updating the cell state, which is continuously regulated, by partially discarding the current memory content and adding new content in the cell state, the input gate is calculated as:
wherein h is t-1 Is the hidden state of the last moment, W ii And W is ih The weight matrix of the input-input gate and the hidden state-input gate, b i Representing the bias of the input gate, σ is the activation function.
The following is a memory cell c t Updating, wherein the calculation formula is as follows:
wherein +.The calculation formula of (2) is as follows:
wherein W is ci And W is ch The weight matrix of the input-memory cell and the hidden state-memory cell, respectively, tanh is an activation function, normalize the output value to [ -1,1]；
Finally, determining the value of the next hidden state by using an output gate, wherein the hidden state contains the previous input information, and the calculation formula of the output gate is as follows:
wherein h is t-1 Is the hidden state of the last moment, W oi And W is oh Input-output gate and hidden state-output, respectivelyWeight matrix of gates, b o Representing the bias of the input gate, sigma is the activation function, normalizing the output value to [0,1 ]]；
Output hidden state h t Is calculated by the following formula:
h t ＝o t tanh(c t )；
(2) Obtained from a multi-stage convolutional self-coding sub-networkAnd->Respectively used as time sequence input of three long-short-term memory neural networks, and the output of the last unit is the extracted space-spectrum-time characteristic and is marked as f 1 ,f 2 And f 3 ；
(3) Will f 1 ,f 2 And f 3 The method comprises the steps of respectively passing through a full-connection layer, setting the number of input nodes and the number of output nodes as 1024 and 256, 512 and 128, 128 and 32, cascading the output of the full-connection layer, and then passing through two layers of full-connection, and obtaining the probability of change and unchanged by utilizing a softmax function to obtain a final prediction label.
Further, performing supervised training on a built network model formed by the two sub-networks together to obtain network parameters suitable for the model;
(1) Training sample I to be labeled t Extracting 16 samples which are not repeated randomly each time as a batch, inputting the samples into a network model to be trained, and outputting label prediction of the training samples;
(2) Calculating a loss function between the predicted label and the real label of the reference image using the following loss function formula:
L＝L rec1 +L rec2 +L E ；
where L is the final loss function,and->Respectively T 1 Time convolution is from the j-th band of input samples and reconstructed samples of the encoder,/th band of samples>And->Respectively T 2 The j-th band of the input samples and reconstructed samples of the time convolution self-encoder, C representing the band number, y and +.>Respectively representing a real label and a predicted label of a training sample;
(3) Training the network parameters by using a random gradient descent method until the network converges, and storing the optimal network parameters to finish the discrimination of two types, namely, change and unchanged.
By combining all the technical schemes, the application has the advantages and positive effects that: the application extracts the multi-level space-spectrum characteristics, gives consideration to the high-level semantic information and the high-resolution detailed information of the images, learns the time dependency relationship between multi-phase hyperspectral images, and improves the performance of the change detection method.
The application utilizes convolution self-coding sub-network to extract multi-stage space-spectrum characteristics, wherein each stage of characteristics are obtained by cascading characteristics with the same size in an encoder and a decoder, complementary and redundant information is provided, the multi-stage characteristics not only comprise deep characteristics with abstract high-level semantic information, but also cover high resolution details of capturing images by shallow characteristics, so that the obtained space-spectrum characteristics are richer, and compared with the prior art, the method utilizes single-stage characteristics and only comprises coarse deep information, the precision of a change detection method is improved.
The application uses the cyclic sub-network to treat two inputs of the change detection task as continuous time sequences, thereby extracting the time sequence characteristics of the double-phase hyperspectral image. The convolution self-coding and the cyclic neural network are combined, so that the space-spectrum-time sequence characteristics of the hyperspectral image can be fully mined according to the characteristics of the change detection task.
The application defines a joint loss function, wherein the first part of the joint loss function minimizes the reconstruction error of the convolution self-encoder, the second part considers the cross entropy of the final prediction label and the real label, and the two parts are combined to train the network in an end-to-end mode, thereby ensuring that the convolution self-encoder effectively extracts space-spectrum characteristics and finally obtaining the effect of a change detection result graph.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings that are needed in the embodiments of the present application will be briefly described below, and it is obvious that the drawings described below are only some embodiments of the present application, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
Fig. 1 is a flowchart of a hyperspectral image change detection method according to an embodiment of the present application.
Fig. 2 is a schematic diagram of an internal structure of a single cell unit in the long-short-term memory neural network according to an embodiment of the present application.
Fig. 3 is a schematic diagram of a hyperspectral image of two phases provided by an embodiment of the present application.
FIG. 4 is a graph of detection results of different change detection methods according to an embodiment of the present application;
in fig. 4: (a) is a group-trunk standard chart; (b) is a CVA method result graph; (c) is a PCA method result chart; (d) is an ELM method result graph; (e) is a graph of the results of the present application.
Detailed Description
The present application will be described in further detail with reference to the following examples in order to make the objects, technical solutions and advantages of the present application more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the application.
Aiming at the problems existing in the prior art, the application provides a hyperspectral image change detection method based on a multistage cyclic convolution self-coding network, and the application is described in detail below with reference to the accompanying drawings.
As shown in fig. 1, the hyperspectral image change detection method based on the multistage cyclic convolution self-coding network provided by the application comprises the following steps:
s101: inputting two double-phase hyperspectral images acquired at different times in the same region, and preprocessing the images;
s102: generating a sample set, and selecting a training sample set and a test sample set;
s103: constructing a convolution self-encoder, introducing an attention mechanism, and forming a multi-stage convolution self-encoding sub-network by two parallel convolution self-encoders with attention modules, thereby acquiring multi-stage space-spectrum characteristics corresponding to each sample;
s104: constructing a long-term memory neural network, forming a circulating sub-network by three parallel long-term memory neural networks, taking space-spectrum characteristics of the same scale of two time phases obtained by the multi-stage convolution self-coding sub-network as input, obtaining space-spectrum-time characteristics, and classifying by utilizing a full-connection layer;
s105: performing supervised training on a built network model formed by a multi-stage convolution self-coding sub-network and a circulation sub-network together to obtain network parameters suitable for the model;
s106: and sequentially placing all samples of the hyperspectral image into a trained network for discrimination to obtain a final change detection result graph.
The method for detecting the hyperspectral image change based on the multistage cyclic convolution self-coding network provided by the application can be implemented by adopting other steps by a person skilled in the art, and the method for detecting the hyperspectral image change based on the multistage cyclic convolution self-coding network provided by the application of fig. 1 is only one specific embodiment.
The technical scheme of the application is further described below with reference to the accompanying drawings.
As shown in fig. 1, the implementation process of the hyperspectral image change detection method based on the multistage cyclic convolution self-coding network provided by the embodiment of the application is as follows:
(1) Two double-phase hyperspectral images acquired in different time in the same region are inputAnd carrying out maximum and minimum normalization on the image, wherein the normalization formula is as follows:
wherein x is i Representing a picture element, x, in a hyperspectral image max And x min Respectively represent the maximum value and the minimum value of the hyperspectral image,is one pixel after normalization.
(2) Generating a sample set, and selecting a training sample set I t And test sample set I e ；
(2a) The three-dimensional hyperspectral data are recorded as H multiplied by W multiplied by C, wherein H and W are the height and the width of a hyperspectral image respectively, and C is the band number of the hyperspectral image; the change detection task needs to perform classification judgment on change and unchanged for each pixel one by one, and needs to combine neighborhood information around the pixel, so that for a double-phase hyperspectral image, a data block with the size of 5 multiplied by C pixels is selected as a sample by taking each pixel point of the image as the center, and a set of all samples is a generated sample set;
(2b) In practical cases, the proportion of the changed samples is usually small, the detection performance is greatly influenced by the 1:1 selected samples, in order to solve the problem of sample imbalance, firstly, the proportion a and b of the changed samples and the unchanged samples in the reference image respectively accounting for the total samples are calculated, and more unchanged samples are used for training a network;
(2c) Randomly selecting 20% of total samples as training sample set I t The remaining 80% is used as test sample set I e The ratio of the changed sample to the unchanged sample in the training sample set and the test sample set is a to b.
(3) Constructing a multistage convolution self-coding sub-network, and introducing two attention mechanism modules into a convolution self-coder to obtain multistage space-spectrum characteristics corresponding to each sample;
(3a) The constructed convolutional self-encoder consists of two parts, encoder and decoder, for a pair of input samplesAnd->The encoder comprises three convolution layers with a convolution kernel size of 3×3 and a step size of 2, so that the encoder obtains feature map sizes of 5×5× 0256,3 ×13× 2256,1 ×31×512, respectively; the decoder comprises three deconvolution layers with a convolution kernel size of 3×3 and a step size of 2, resulting in a feature map size of 3×3× 256,5 ×5×256, and reconstructed 5×5×c samples, respectivelyAnd->Batch normalization and ReLU activation function processing is used after each convolutional layer;
(3b) Attention modules are introduced into the convolution self-encoder to improve the feature extraction effect, the attention modules are formed by connecting spatial attention modules and channel attention modules in series and are embedded into a residual error module, so that the specific structure of the convolution self-encoder adopted by the application is as follows: first convolution layer- & gtattention module- & gtsecond convolution layer- & gtthird convolution layer- & gtfirst deconvolution layer- & gtsecond deconvolution layer- & gtattention module- & gtthird deconvolution layer;
(3c) Two parallel convolutional self-encoders form a multi-stage convolutional self-encoding sub-network for a pair of input samplesAnd->After the multi-level features are extracted from the encoder through convolution respectively, the features with the same size in the same convolution self-encoder are cascaded to output three pairs of multi-level space-spectrum features with different sizes> And->The feature sizes obtained were 5×5× 512,3 ×3× 512,1 ×1×512, respectively.
(4) Constructing a cyclic sub-network, and respectively obtaining the multi-level convolution self-coding sub-network by adopting three identical long-short-period memory neural networksAnd->As the time sequence input of three cyclic neural networks, extracting the time dependency relationship of the three cyclic neural networks, and classifying the three cyclic neural networks by using a full-connection layer;
(4a) The method comprises the steps of constructing a long-term and short-term memory neural network, wherein the long-term and short-term memory neural network consists of two cell units, and each cell unit comprises a forgetting gate, an input gate and an output gate;
first calculate the forgetting door f t The forgetting gate controls the degree to which the contents of the existing memory cells are forgotten, i.e., decides which information to retain and discard, forgetsThe calculation formula of the door is:
wherein h is t-1 Is the hidden state of the last moment, W fi And W is fh Weight matrix of input-forget gate and hidden state-forget gate, b f Representing the bias of the forgetting gate, sigma is the activation function, normalize the output value to [0,1]；
Then calculate the input gate i t The input gate is responsible for updating the cell state, which is continuously regulated, by partially discarding the current memory content and adding new content in the cell state, the input gate is calculated as:
wherein h is t-1 Is the hidden state of the last moment, W ii And W is ih The weight matrix of the input-input gate and the hidden state-input gate, b i Representing the bias of the input gate, σ is the activation function.
The following is a memory cell c t Updating, wherein the calculation formula is as follows:
wherein +.The calculation formula of (2) is as follows:
wherein W is ci And W is ch The weight matrix of the input-memory cell and the hidden state-memory cell, respectively, tanh is an activation function, normalize the output value to [ -1,1]；
Finally, determining the value of the next hidden state by using an output gate, wherein the hidden state contains the previous input information, and the calculation formula of the output gate is as follows:
wherein h is t-1 Is the hidden state of the last moment, W oi And W is oh Weight matrix of input-output gate and hidden state-output gate, b o Representing the bias of the input gate, sigma is the activation function, normalizing the output value to [0,1 ]]；
Output hidden state h t Is calculated by the following formula:
h t ＝o t tanh(c t )；
(4b) Obtained from a multi-stage convolutional self-coding sub-networkAnd->Respectively used as time sequence input of three long-short-term memory neural networks, and the output of the last unit is the extracted space-spectrum-time characteristic and is marked as f 1 ,f 2 And f 3 ；
(4c) Will f 1 ,f 2 And f 3 The method comprises the steps of respectively passing through a full-connection layer, setting the number of input nodes and the number of output nodes as 1024 and 256, 512 and 128, 128 and 32, cascading the output of the full-connection layer, and then passing through two layers of full-connection, and obtaining the probability of change and unchanged by utilizing a softmax function to obtain a final prediction label.
(5) Performing supervised training on a built network model formed by the two sub-networks together, so as to obtain network parameters suitable for the model;
(5a) Training sample I to be labeled t Extracting 16 samples which are not repeated randomly each time as a batch, inputting the samples into a network model to be trained, and outputting label prediction of the training samples;
(5b) Calculating a loss function between the predicted label and the real label of the reference image using the following loss function formula:
L＝L rec1 +L rec2 +L E ；
where L is the final loss function,and->Respectively T 1 Time convolution is from the j-th band of input samples and reconstructed samples of the encoder,/th band of samples>And->Respectively T 2 The j-th band of the input samples and reconstructed samples of the time convolution self-encoder, C representing the band number, y and +.>Respectively representing a real label and a predicted label of a training sample;
(5c) Training network parameters by using a random gradient descent method, and traversing the training process through the whole training sample set I t Called one training round, the total training process is 100 training rounds, and the learning rate is set to be 0.0001. After training, the network model detects the hyperspectral image, thereby completing the two kinds of change and unchangedIs determined.
(6) And sequentially placing all samples of the hyperspectral image into a trained network for discrimination to obtain a final change detection result graph.
The technical effects of the present application are described in detail below in conjunction with simulation experiments:
1. simulation experiment conditions:
the hardware platform of the simulation experiment of the application is: the processor is Intel i9-10900 CPU and the memory is 16GB.
The software platform of the simulation experiment of the application is: linux18.04 operating system, python 3.7 and pytorch1.5.
The hyperspectral image used in the simulation experiment adopts farmland images of salt city of Jiangsu province acquired by an Earth Observing-1 (EO-1) hyperserving sensor, the two images are respectively shot on 3 months in 2006 and 23 months in 2007, the final image size is 420 multiplied by 140 multiplied by 154, the spatial resolution is 30m, the spectral range is 0.4-2.5 mu m, and the spectral resolution is 10nm.
2. Experimental content and outcome analysis:
the simulation experiment of the application adopts the application and three prior technologies (CVA, PCA and ELM) to respectively carry out variation detection on the input hyperspectral images of Jiangsu salt city farmland, so as to obtain a final variation detection result graph.
The contrast change detection method in the prior art used in the application is as follows:
the prior art CVA Change detection method is a Change detection method proposed by Malila et al in the literature "Change-vector analysis in multitemporal space: a tool to detect and categorize land-cover Change processes using high temporal-resolution satellite data [ J ]. Remote Sensing of Environment,1994,48 (2): 231-244 ].
The prior art principal component analysis PCA change detection method is a change detection method proposed by Deng et al in the literature "PCA-based land-use change detection and analysis using multitemporal and multisensory satellite data [ J ]. Int.J. remote Sens., vol.29, no.16, pp.4823-4838,2008 ].
The ELM change detection method of the prior art extreme learning machine is a change detection method proposed by Huang et al in the literature "Extreme learning machine: theory and applications [ J ]. Neurochemusting, vol.70, nos.1-3, pp.489-501, dec.2006 ].
The effects of the present application will be further described with reference to a change detection result chart in fig. 4.
As can be seen from fig. 4 (b), the result graph obtained by the existing CVA method retains a large amount of noise, mainly because the method directly performs numerical calculation, does not process noise in data, has poor anti-interference capability, and has a large amount of omission in the result, so that the accuracy of the method is poor.
As can be seen from fig. 4 (c), the result graph obtained by the existing PCA method has fewer noise points compared with the CVA result, mainly because the PCA method converts information into a specific space and extracts only key information, effectively eliminates noise and redundant information, but also causes loss of some information, which causes missing detection conditions to occur and affects the accuracy of the change detection method.
As can be seen from fig. 4 (d), the result diagram obtained by the conventional ELM method not only contains considerable noise, but also has unclear edges, and high-frequency information is not sufficiently extracted, resulting in unclear picture details. The ELM method acts as a simple classifier whose accuracy is not adequate for the task of detecting changes containing detailed and complex information.
As can be seen from fig. 4 (e), the result graph obtained by the present application has the best effect compared with the results obtained by the existing three methods, and not only contains few noise points, but also has clear edges of the changed and unchanged regions, and voids in some regions are detected, including more details.
And respectively objectively evaluating the change detection result graphs obtained by the four methods by using two evaluation indexes (total accuracy OA and chi-square coefficient KAPPA). The total accuracy OA represents the proportion of correctly classified samples to the total samples, and the closer the OA value is to 1, the higher the detection accuracy is; the consistency of the result obtained by the KAPPA coefficient KAPPA characterization and the reference graph shows that the closer the KAPPA value is to 1, the better the method performance is. The values of the various statistical evaluation indexes are plotted in table 1.
TABLE 1 quantitative analysis Table of the results of the present application on the variation detection of hyperspectral images of Jiangsu salt city farmland
As can be seen from the combination of Table 1, the total accuracy OA of the application reaches 98.09%, the KAPPA value reaches 0.9555, and compared with the best Effect (ELM) in the current comparison method, the method respectively improves by 4.28% and 0.0942, which are both significantly higher than the prior art method, and the application can better detect the change region, and the performance is significantly better than the prior art method.
In summary, the application fully excavates the semantic information and the high resolution details of the hyperspectral image by constructing the end-to-end network model based on the multistage cyclic convolution self-coding network and utilizing the mode of extracting multistage characteristics, thereby overcoming the problem that the traditional method only extracts coarse deep semantic information, and simultaneously, the hyperspectral images of two time phases are regarded as continuous sequences with time dependency relationships rather than mutually independent relationships through the long-short-term memory neural network, thereby ensuring that the time sequence characteristics are extracted and effectively improving the precision of the hyperspectral image change detection method based on the multistage cyclic convolution self-coding network.
It should be noted that the embodiments of the present application can be realized in hardware, software, or a combination of software and hardware. The hardware portion may be implemented using dedicated logic; the software portions may be stored in a memory and executed by a suitable instruction execution system, such as a microprocessor or special purpose design hardware. Those of ordinary skill in the art will appreciate that the apparatus and methods described above may be implemented using computer executable instructions and/or embodied in processor control code, such as provided on a carrier medium such as a magnetic disk, CD or DVD-ROM, a programmable memory such as read only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The device of the present application and its modules may be implemented by hardware circuitry, such as very large scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., as well as software executed by various types of processors, or by a combination of the above hardware circuitry and software, such as firmware.
The foregoing is merely illustrative of specific embodiments of the present application, and the scope of the application is not limited thereto, but any modifications, equivalents, improvements and alternatives falling within the spirit and principles of the present application will be apparent to those skilled in the art within the scope of the present application.

Claims (5)

1. The hyperspectral image change detection method based on the multistage cyclic convolution self-coding network is characterized by comprising the following steps of:
inputting two double-phase hyperspectral images acquired at different times in the same region, and preprocessing the images;
generating a sample set, and selecting a training sample set and a test sample set;
constructing a convolution self-encoder, introducing an attention mechanism, forming a multi-stage convolution self-encoding sub-network by two parallel convolution self-encoders with attention modules, and acquiring multi-stage space-spectrum characteristics corresponding to each sample;
constructing a long-term memory neural network, forming a circulating sub-network by three parallel long-term memory neural networks, taking space-spectrum characteristics of the same scale of two time phases obtained by the multi-stage convolution self-coding sub-network as input, obtaining space-spectrum-time characteristics, and classifying by utilizing a full-connection layer;
performing supervised training on a built network model formed by a multi-stage convolution self-coding sub-network and a circulation sub-network together to obtain network parameters suitable for the model;
sequentially placing all samples of the hyperspectral image into a trained network for discrimination to obtain a final change detection result graph;
constructing a multistage convolution self-coding sub-network, introducing two attention mechanism modules into the convolution self-coder, and respectively extracting multistage space-spectrum characteristics from hyperspectral images of two time phases;
(1) The constructed convolutional self-encoder consists of two parts, encoder and decoder, for a pair of input samplesAnd->The encoder comprises three convolution layers with a convolution kernel size of 3×3 and a step size of 2, so that the encoder obtains feature map sizes of 5×5× 0256,3 ×13× 2256,1 ×31×512, respectively; the decoder comprises three deconvolution layers with a convolution kernel size of 3 x 3 and a step size of 2, resulting in a feature map size of 3 x 256,5 x 5 x 256 and reconstructed samples +.>And->Batch normalization and ReLU activation function processing is used after each convolutional layer;
(2) Attention modules are introduced into the convolution self-encoder to improve the feature extraction effect, the attention modules are formed by connecting spatial attention modules and channel attention modules in series, and are embedded into a residual error module, and the specific structure of the adopted convolution self-encoder is as follows: first convolution layer- & gtattention module- & gtsecond convolution layer- & gtthird convolution layer- & gtfirst deconvolution layer- & gtsecond deconvolution layer- & gtattention module- & gtthird deconvolution layer;
(3) Two parallel convolutional self-encoders form a multi-stage convolutional self-encoding sub-network for a pair of input samplesAndafter the multi-level features are extracted from the encoder through convolution respectively, the features with the same size in the same convolution self-encoder are cascaded to output three pairs of multi-level space-spectrum features with different sizes> And->The feature sizes obtained were 5×5× 512,3 ×3× 512,1 ×1×512, respectively.
2. The hyperspectral image variation detecting method based on multi-stage cyclic convolution self-encoding network as claimed in claim 1, wherein the hyperspectral image variation detecting method based on multi-stage cyclic convolution self-encoding network inputs two double-phase hyperspectral images acquired at different times in the same regionAnd carrying out maximum and minimum normalization on the image, wherein the normalization formula is as follows:
wherein x is i Representing a picture element, x, in a hyperspectral image max And x min Respectively represent the maximum value and the minimum value of the hyperspectral image,is one pixel after normalization.
3. The hyperspectral image variation detection method based on the multi-stage cyclic convolution self-coding network according to claim 1, wherein a sample set is generated and a training sample set I is selected t And test sample set I e ；
(1) The three-dimensional hyperspectral data are recorded as H multiplied by W multiplied by C, wherein H and W are the height and the width of a hyperspectral image respectively, and C is the band number of the hyperspectral image; for the hyperspectral image of the double phases, taking each pixel point of the image as the center, selecting a data block with the size of 5 multiplied by C as a sample, wherein the set of all the samples is the generated sample set;
(2) Calculating the proportion a and b of the changed sample and the unchanged sample in the reference image to the total sample respectively, solving the problem of unbalanced sample, and training a network by using more unchanged samples;
(3) Randomly selecting 20% of total samples as training sample set I t The remaining 80% is used as test sample set I e The ratio of the changed sample to the unchanged sample in the training sample set and the test sample set is a to b.
4. The hyperspectral image change detection method based on multi-stage cyclic convolution self-coding network as claimed in claim 1, wherein a cyclic subnetwork is constructed, three identical long-term and short-term memory neural networks are adopted, and the multi-stage cyclic convolution self-coding subnetwork is respectively adopted to obtainAnd->As the time sequence input of three cyclic neural networks, extracting the time dependency relationship of the three cyclic neural networks, and classifying the three cyclic neural networks by using a full-connection layer;
(1) The method comprises the steps of constructing a long-term and short-term memory neural network, wherein the long-term and short-term memory neural network consists of two cell units, and each cell unit comprises a forgetting gate, an input gate and an output gate;
first calculate the forgetting door f t The forgetting gate controls the forgetting degree of the content of the existing memory cell, namely decides which information is reserved and discarded, and the calculation formula of the forgetting gate is as follows:
wherein h is t-1 Is the hidden state of the last moment, W fi And W is fh Weight matrix of input-forget gate and hidden state-forget gate, b f Representing the bias of the forgetting gate, sigma is the activation function, normalize the output value to [0,1]；
Then calculate the input gate i t The input gate is responsible for updating the cell state, which is continuously regulated, by partially discarding the current memory content and adding new content in the cell state, the input gate is calculated as:
wherein h is t-1 Is the hidden state of the last moment, W ii And W is ih The weight matrix of the input-input gate and the hidden state-input gate, b i Representing the bias of the input gate, σ being the activation function;
the following is a memory cell c t Updating, wherein the calculation formula is as follows:
wherein +.The calculation formula of (2) is as follows:
wherein W is ci And W is ch The weight matrix of the input-memory cell and the hidden state-memory cell, respectively, tanh is an activation function, normalize the output value to [ -1,1]；
Finally, determining the value of the next hidden state by using an output gate, wherein the hidden state contains the previous input information, and the calculation formula of the output gate is as follows:
wherein h is t-1 Is the hidden state of the last moment, W oi And W is oh Weight matrix of input-output gate and hidden state-output gate, b o Representing the bias of the input gate, sigma is the activation function, normalizing the output value to [0,1 ]]；
Output hidden state h t Is calculated by the following formula:
h t ＝o t tanh(c t )；
(2) Obtained from a multi-stage convolutional self-coding sub-networkAnd->Respectively used as time sequence input of three long-short-term memory neural networks, and the output of the last unit is the extracted space-spectrum-time characteristic and is marked as f 1 ,f 2 And f 3 ；
(3) Will f 1 ,f 2 And f 3 Respectively passing through a full-connection layer, setting the number of input nodes and the number of output nodes as 1024 and 256, 512 and 128, 128 and 32, cascading the output of the full-connection layer, and passing through two full-connection layers and utilizing softmax functionThe number is changed and the probability of unchanged, and the final predictive label is obtained.
5. The hyperspectral image change detection method based on the multistage cyclic convolution self-coding network according to claim 1, wherein the constructed network model formed by two sub-networks together is subjected to supervised training to obtain network parameters suitable for the model;
(1) Training sample I to be labeled t Extracting 16 samples which are not repeated randomly each time as a batch, inputting the samples into a network model to be trained, and outputting label prediction of the training samples;
(2) Calculating a loss function between the predicted label and the real label of the reference image using the following loss function formula:
L＝L rec1 +L rec2 +L E ；
where L is the final loss function,and->Respectively T 1 Time convolution is from the j-th band of input samples and reconstructed samples of the encoder,/th band of samples>And->Respectively T 2 The j-th band of the input samples and reconstructed samples of the time convolution self-encoder, C representing the band number, y and +.>Respectively representing a real label and a predicted label of a training sample;
(3) Training the network parameters by using a random gradient descent method until the network converges, and storing the optimal network parameters to finish the discrimination of two types, namely, change and unchanged.
CN202110036358.XA 2021-01-12 2021-01-12 Hyperspectral image change detection method based on multistage cyclic convolution self-coding network Active CN112733725B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110036358.XA CN112733725B (en) 2021-01-12 2021-01-12 Hyperspectral image change detection method based on multistage cyclic convolution self-coding network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110036358.XA CN112733725B (en) 2021-01-12 2021-01-12 Hyperspectral image change detection method based on multistage cyclic convolution self-coding network

Publications (2)

Publication Number Publication Date
CN112733725A CN112733725A (en) 2021-04-30
CN112733725B true CN112733725B (en) 2023-09-22

Family

ID=75590538

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110036358.XA Active CN112733725B (en) 2021-01-12 2021-01-12 Hyperspectral image change detection method based on multistage cyclic convolution self-coding network

Country Status (1)

Country Link
CN (1) CN112733725B (en)

Families Citing this family (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113781284A (en) * 2021-06-30 2021-12-10 华南农业大学 Zero watermark construction method based on depth attention self-encoder
CN113421276B (en) * 2021-07-02 2023-07-21 深圳大学 Image processing method, device and storage medium
CN113433081B (en) * 2021-07-06 2022-05-24 江南大学 Dynamic neural network-based quality detection method and system in fruit and vegetable drying process
CN113591816B (en) * 2021-09-29 2022-02-22 广东省科学院智能制造研究所 Hyperspectral anomaly detection method and system based on self-supervision guide coding network
CN113960090A (en) * 2021-10-15 2022-01-21 电子科技大学长三角研究院(湖州) LSTM neural network algorithm-based soil Cd element spectrum qualitative analysis method
CN114124823B (en) * 2021-10-18 2023-08-11 西安电子科技大学 Self-adaptive routing method, system and equipment oriented to high dynamic network topology
CN114862906B (en) * 2022-04-11 2024-05-07 中山大学 Visible light positioning and tracking method based on bidirectional cyclic convolutional neural network
CN115019166B (en) * 2022-05-24 2024-02-09 深圳大学 Method, device, medium and terminal for extracting marsh wetland information
CN115240072B (en) * 2022-08-04 2024-03-12 南京审计大学 Hyperspectral multi-class change detection method based on multidirectional multi-scale spectrum-space residual convolution neural network
CN115471677B (en) * 2022-09-15 2023-09-29 贵州大学 Hyperspectral image classification method based on double-channel sparse network
CN116309497B (en) * 2023-03-26 2023-10-03 湖南医药学院 Image recognition-based auxiliary analysis method for cancer cell counting and prognosis prediction
CN116384237A (en) * 2023-03-29 2023-07-04 大连海事大学 Thermal infrared atmospheric parameter inversion method and device and electronic equipment
CN116563711B (en) * 2023-05-17 2024-02-09 大连民族大学 Hyperspectral target detection method of binary encoder network based on momentum update

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110084159A (en) * 2019-04-15 2019-08-02 西安电子科技大学 Hyperspectral image classification method based on the multistage empty spectrum information CNN of joint
CN111126282A (en) * 2019-12-25 2020-05-08 中国矿业大学 Remote sensing image content description method based on variation self-attention reinforcement learning
CN112052758A (en) * 2020-08-25 2020-12-08 西安电子科技大学 Hyperspectral image classification method based on attention mechanism and recurrent neural network
CN112052755A (en) * 2020-08-24 2020-12-08 西安电子科技大学 Semantic convolution hyperspectral image classification method based on multi-path attention mechanism
WO2020258611A1 (en) * 2019-06-28 2020-12-30 山东科技大学 Lymph node ct detection system employing recurrent spatio-temporal attention mechanism

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110084159A (en) * 2019-04-15 2019-08-02 西安电子科技大学 Hyperspectral image classification method based on the multistage empty spectrum information CNN of joint
WO2020258611A1 (en) * 2019-06-28 2020-12-30 山东科技大学 Lymph node ct detection system employing recurrent spatio-temporal attention mechanism
CN111126282A (en) * 2019-12-25 2020-05-08 中国矿业大学 Remote sensing image content description method based on variation self-attention reinforcement learning
CN112052755A (en) * 2020-08-24 2020-12-08 西安电子科技大学 Semantic convolution hyperspectral image classification method based on multi-path attention mechanism
CN112052758A (en) * 2020-08-25 2020-12-08 西安电子科技大学 Hyperspectral image classification method based on attention mechanism and recurrent neural network

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
基于深度卷积长短时神经网络的视频帧预测;张德正;翁理国;夏旻;曹辉;;计算机应用(第06期);全文 *
结合深度学习和超像元的高分遥感影像变化检测;王艳恒;高连如;陈正超;张兵;;中国图象图形学报(第06期);全文 *

Also Published As

Publication number Publication date
CN112733725A (en) 2021-04-30

Similar Documents

Publication Publication Date Title
CN112733725B (en) Hyperspectral image change detection method based on multistage cyclic convolution self-coding network
Cai et al. BS-Nets: An end-to-end framework for band selection of hyperspectral image
Zhao et al. Spectral-spatial classification of hyperspectral imagery based on stacked sparse autoencoder and random forest
CN109949278B (en) Hyperspectral anomaly detection method based on antagonistic self-coding network
Sathishkumar et al. Forest fire and smoke detection using deep learning-based learning without forgetting
CN110657984B (en) Planetary gearbox fault diagnosis method based on reinforced capsule network
CN113392931B (en) Hyperspectral open set classification method based on self-supervision learning and multitask learning
Qi et al. Deep spectral convolution network for hyperspectral image unmixing with spectral library
CN114037891A (en) High-resolution remote sensing image building extraction method and device based on U-shaped attention control network
CN111627007A (en) Spacecraft defect detection method based on self-optimization matching network image stitching
CN115439708A (en) Image data processing method and device
CN101826160A (en) Hyperspectral image classification method based on immune evolutionary strategy
Kang et al. YOLO-FA: Type-1 fuzzy attention based YOLO detector for vehicle detection
CN113705538A (en) High-resolution remote sensing image road change detection device and method based on deep learning
CN113887559A (en) Brain-computer information fusion classification method and system for brain off-loop application
CN117636298A (en) Vehicle re-identification method, system and storage medium based on multi-scale feature learning
CN116935240A (en) Surface coverage classification system and method for multi-scale perception pyramid
CN113343861B (en) Remote sensing image water body region extraction method based on neural network model
Wu et al. Unsupervised encoder-decoder network under spatial and spectral guidance for hyperspectral and multispectral image fusion
Huan et al. SAR multi‐target interactive motion recognition based on convolutional neural networks
Xu et al. Based on improved CNN bearing fault detection
CN114140524A (en) Closed loop detection system and method for multi-scale feature fusion
US11657282B2 (en) Efficient inferencing with fast pointwise convolution
CN113657351A (en) High-resolution remote sensing image forest and grass change detection device and method based on deep learning
CN113947712A (en) Hyperspectral anomaly detection method and system based on capsule differential countermeasure network

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
GR01 Patent grant
GR01 Patent grant