A network construction method for local and global deep contrast feature learning for change detection between images

Abstract

The present invention discloses a method for constructing a local and global deep contrast feature learning network for detecting changes between images, comprising: encoding an input image to obtain a global high-dimensional representation feature; constructing positive sample data and negative sample data whose feature domains are similar to the global features; measuring the degree of correlation between the global features and the constructed positive and negative sample data; establishing a constrained prior model for hierarchical input of the feature domain; obtaining the optimal encoding representation feature; iteratively optimizing the probability model to solve the difference probability map matrix; and outputting a binary change result map. The present invention makes full use of the structural information of the hierarchical input of the feature domain to construct a network loss function based on the global features, local features and prior losses of positive and negative samples. By optimizing the loss function, the parameters of the encoder are updated, so that the encoder can effectively learn the local and global deep contrast features between images, reflecting the correlation between abstract semantic information and spatial information.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing image processing, and in particular relates to a method for constructing a local and global depth contrast feature learning network for inter-image change detection. Background Art

[0002] Change detection in remote sensing images is the process of quantitatively analyzing and identifying the characteristics of surface changes in multi-temporal remote sensing images acquired over the same area at different times. The goal of change detection is to identify areas of change and represent them using binary images. Change detection can accurately and timely detect Earth's surface features, providing a foundation for a better understanding of the relationships and interactions between humans and natural phenomena, enabling better resource management and utilization. This technology has both urgent scientific applications and broad potential. This technology has been widely applied in areas such as forest fire prevention, environmental change, glacial changes, urban expansion, farmland monitoring, flood warning, road analysis, and post-earthquake disaster detection.

[0003] Most current change detection methods target homologous images, that is, images acquired by the same sensor. For example, in SAR imagery, unchanged regions within an image exhibit identical features, allowing for direct comparison and relatively simple change detection. However, this approach faces significant limitations in practical applications. To detect changes in a region after an event, not only pre-event images of the region are required, but also post-event images from the same sensor. While SAR images are unaffected by weather and light and can operate 24 / 7, they are often affected by multiplicative speckle noise due to the characteristics of radar echoes. Optical imagery offers nearly global coverage, but using only optical sensors can significantly impact image quality due to lighting and weather conditions, making rapid detection difficult in critical situations. For example, in earthquake rescue operations, weather-related issues make it difficult for optical sensors to obtain high-quality images of the affected area immediately. SAR sensors, however, are immune to weather interference, so SAR images are often the primary means of acquisition. Due to the limited coverage of SAR images, pre-earthquake SAR images of the affected area are difficult to compare with. Optical images, however, offer wide coverage, making pre-earthquake images of the affected area readily available. Therefore, combining the strengths of both optical and SAR imagery for change detection analysis is a promising approach. Therefore, change detection in heterogeneous images has become a research focus, but it also presents a huge challenge: how to measure the difference information between heterogeneous images. An effective network-based approach is an end-to-end method [Gong MG, Zhao JJ, Liu J, et al. Change Detection in Synthetic Aperture Radar Images Based on Deep Neural Networks[J]. IEEE Transactions on Neural Networks and Learning Systems, 2017, 27(1): 125-138.]. A trained deep neural network is directly used to generate a change detection map for the two input images. This method omits the process of generating the difference image, and the network output is the change detection binary image.Another method is to generate a difference map based on a network, such as the SCNN method [Liu J, Gong MG, Qin K, et al. A Deep Convolutional Coupling Network for Change Detection Based on Heterogeneous Optical and Radar Images [J]. IEEE Transactions on Neural Networks and Learning Systems, 2018, 29 (3): 545-559.]. Unsupervised learning is used to extract local features of heterogeneous images and convert them into the same feature space for comparison. The neural network is optimized and trained to obtain a difference map. The difference map is then analyzed using a clustering method to obtain a binary change detection image. Neither of the above two methods maximizes the correlation between global features and local features in the feature domain and between similar features (positive samples), nor minimizes the correlation between features that are dissimilar to global features (negative samples). It is difficult to effectively establish spatial correlation. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for constructing a local and global deep contrast feature learning network for inter-image change detection. This method improves the information utilization of samples by hierarchically inputting structural information in the feature domain; and uses the contrast loss function of positive and negative samples of constructed images to learn model parameters without the need for manual feature analysis. It is an unsupervised method.

[0005] The technical solution to achieve the purpose of the present invention is: a method for constructing a local and global deep contrast feature learning network for inter-image change detection, comprising the following steps:

[0006] (1) Use the network to encode the input image to obtain global high-dimensional representation features and intermediate layer features;

[0007] (2) Construct global positive and negative sample data and local positive and negative sample data;

[0008] (3) Construct contrast loss function;

[0009] (4) Solve the difference probability map matrix;

[0010] (5) Cluster the difference probability map matrix and output the binary change result map.

[0011] Furthermore, input an image X1∈R of the same scene at different times W×H and an image X2∈R W×H , W represents the width of the image, and H represents the height of the image.

[0012] Furthermore, the original image is randomly taken into blocks x by pixel i ,(i=1,2,...,N),block x i Pull into a one-dimensional vector as the encoder E ψ (x i ) input image, obtain the global high-dimensional representation features of the feature domain (high-dimensional features represent the last layer features) y ψ and the intermediate layer features M ψ .

[0013] Furthermore, global positive and negative sample data and local positive and negative sample data are constructed. The specific process is as follows:

[0014] (1) Global positive and negative sample data: intermediate layer feature M ψ As the global positive sample data y ψ Merge the intermediate layer features of the first image of each batch with the end features of the batch to construct the global negative sample data M_prime ψ ;

[0015] (2) Local positive and negative sample data: The middle layer feature M ψ and the global high-dimensional feature y ψ Splicing to obtain local positive sample data y_M ψ . The intermediate layer feature M_prime of the global negative sample ψ and high-dimensional features y ψ Splicing to obtain local negative sample data y_M_prime ψ .

[0016] Furthermore, the contrast loss function is constructed as follows:

[0017] (1) Global feature loss function Defined as:

[0018]

[0019] Among them, G ω represents the global feature discriminator, S p represents the activation function, and E represents the expectation.

[0020] (2) Local feature loss function Defined as:

[0021]

[0022] Among them, L ω represents the local feature discriminator, S p represents the activation function, and E represents the expectation.

[0023] (3) Prior loss function PRIOR φ,ψ for:

[0024] PRIOR φ,ψ =-(log(P φ (y rand )+log(1-P φ (y ψ )))

[0025] Among them, P φ represents the prior discriminator, y rand is a random vector that obeys the uniform distribution U(0,1), and y ψ Keep the size consistent.

[0026] The present invention constructs positive sample data related to global features and negative sample data unrelated to them in the feature domain, proposes a network framework for local and global deep contrast feature learning, and constructs a network loss function containing three items: global features, local features, and prior loss to learn model parameters, measure the similarity between global features and positive sample data and negative sample data, and effectively reflect the correlation between abstract semantic information and spatial information; in the prior loss, the high-dimensional representation has the characteristics of the expected prior distribution. Compared with the existing technology, the present invention organically integrates high-level semantic information and low-level spatial information to form a more refined feature representation. At the same time, through the contrast feature learning framework, a more discriminative abstract semantic feature is formed. The model has fewer parameters, simpler optimization, and stronger generalization ability. It can be widely used in heterogeneous remote sensing image change detection, lighting-robust video surveillance, etc.

[0027] The present invention is further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is the overall flow chart of the present invention.

[0029] Figure 2 This is the overall framework diagram of the model.

[0030] Figure 3 It is a global feature diagram.

[0031] Figure 4 It is a local feature diagram.

[0032] Figure 5 It is a schematic diagram of prior constraints.

[0033] Figure 6(a) is the Tianhe Airport dataset: optical image in July 2002.

[0034] Figure 6(b) is the Tianhe Airport dataset: optical image in June 2013.

[0035] Figure 7(a) is the NJUST dataset: optical images from September 2013.

[0036] Figure 7(b) is the NJUST dataset: optical images from March 2019

[0037] Figure 8(a) is the Yellow River dataset: SAR image in June 2008.

[0038] Figure 8(b) is the Yellow River dataset: SAR image in June 2009

[0039] Figure 9(a) shows the change detection results of the Tianhe Airport dataset using the NR-FLICM method. Figure 9(b) shows the change detection results of the Tianhe Airport dataset using the CWNN method.

[0040] Figure 9(c) shows the change detection results of the Tianhe Airport dataset using the SCNN method. Figure 9(d) shows the change detection results of the Tianhe Airport dataset using the cGAN method. Figure 9(e) shows the change detection results of the Tianhe Airport dataset using the present invention.

[0041] Figure 9(f) is the reference image of the Tianhe Airport dataset.

[0042] Figure 10(a) shows the change detection results of the NJUST dataset using the NR-FLICM method. Figure 10(b) shows the change detection results of the NJUST dataset using the CWNN method.

[0043] Figure 10(c) shows the change detection results of the NJUST dataset using the SCNN method. Figure 10(d) shows the change detection results of the NJUST dataset using the cGAN method. Figure 10(e) shows the change detection results of the NJUST dataset using the present invention.

[0044] Figure 10(f) is a reference image of the NJUST dataset.

[0045] Figure 11(a) shows the change detection results of the Yellow River dataset using the NR-FLICM method.

[0046] Figure 11(b) shows the change detection results of the Yellow River dataset using the CWNN method.

[0047] Figure 11(c) shows the change detection results of the Yellow River dataset using the SCNN method.

[0048] Figure 11(d) shows the change detection results of the Yellow River dataset using the cGAN method.

[0049] FIG11( e ) is a diagram showing the change detection results of the Yellow River dataset using the present invention.

[0050] Figure 11(f) is the reference map of the Yellow River dataset. DETAILED DESCRIPTION

[0051] Analysis shows that similar samples have high mutual information in the same area, which is conducive to the comprehensive construction of image information representation and the sharing of information from all aspects of the data. At the same time, by measuring the similarity between hierarchical structural information, the model's feature representation of the image has a smaller loss value with positive samples and a larger loss value with negative samples, thereby effectively and accurately extracting the image's feature information. Furthermore, although low-level features are detailed, they lack semantic information. Therefore, fusing high-level semantic information with low-level spatial information can produce more refined feature representations. Moreover, because the model can learn more discriminative contrast semantic features in the feature space at the abstract semantic level, the model has stronger generalization capabilities.

[0052] This paper proposes a new unsupervised change detection method for multi-source heterogeneous images based on a contrastive learning model. The network parameters are trained by optimizing the objective function constructed based on the contrastive learning model. The contrastive learning model is then used to extract image features. Images of different properties are converted into the same feature space, and then compared to generate a difference map.

[0053] Combine Figures 1 to 5 , the implementation process of the present invention is described in detail below, the steps are as follows:

[0054] Step 1: Input two remote sensing images from different sources

[0055] Input a SAR image X1∈R of the same scene at different times W×H and an optical image X2∈R W×H , the image width W = 523 and the image height H = 447. Randomly initialize the encoder E ψ (X) network parameters ψ and discriminator and P φ Parameters.

[0056] Step 2: Encode feature representation

[0057] (1) Encoder E ψ (X) Network structure

[0058] An image X is used as the input of the encoding network, and the encoding feature representation is obtained through the convolutional layer and the fully connected layer. Among them, y ψ represents the high-dimensional features of the encoded feature representation, M ψ represents the intermediate layer features of the encoded feature representation, Indicates the nth fc fully connected layers, Indicates the nth covConvolutional layers. Number of convolutional layers n cov =4, number of fully connected layers n fc =1.

[0059] (2) Randomly select a block x from each pixel of an input SAR image X1 i ,(i=1,2,...,N), take the number of blocks N=233781, block x i Pull into a one-dimensional vector as encoder E ψ (x i ) is used as the original image input to obtain the global representation feature y of the SAR image X1 in the feature domain ψ and the global intermediate layer feature M ψ .

[0060] (3) Randomly select a block x from another input optical image X2 by pixel i ,(i=1,2,...,N), take the number of blocks N=233781, block x i Pull into a one-dimensional vector as encoder E ψ (x i ) is input as the original image, and the global representation feature y of the optical image X2 in the feature domain is obtained ψ and the global intermediate layer feature M ψ .

[0061] Step 3: Construct positive and negative samples in the feature domain

[0062] Construct global positive and negative sample data and local positive and negative sample data. The specific process is:

[0063] (1) Global positive and negative sample data: intermediate layer feature M ψ As the global positive sample data y ψ Merge the intermediate layer features of the first image of each batch with the end features of the batch to construct the global negative sample data M_prime ψ ;

[0064] (2) Local positive and negative sample data: The middle layer feature M ψ and the global high-dimensional feature y ψ Splicing to obtain local positive sample data y_M ψ . The intermediate layer feature M_prime of the global negative sample ψ and high-dimensional features y ψ Splicing to obtain local negative sample data y_M_prime ψ .

[0065] Step 4: Construct contrast loss function

[0066] (1) Global feature loss function Defined as:

[0067]

[0068] Among them, the global feature discriminator is the discriminator of the convolutional fully connected structure, S p Represents the softplus activation function: S(x)=log(1+e x ), E represents the expectation.

[0069] (2) Local feature loss function Defined as:

[0070]

[0071] Among them, the local feature discriminator is a 1×1 convolution discriminator, S p Represents the softplus activation function: S(x)=log(1+e x ), E represents the expectation.

[0072] (3) Prior loss function PRIOR φ,ψ for:

[0073] PRIOR φ,ψ =-(log(P φ (y rand )+log(1-P φ (y ψ )))

[0074] Among them, the prior discriminator P φ is the fully connected discriminator, y rand is a random vector that obeys the uniform distribution U(0,1), and y ψ Keep the size consistent.

[0075] (4) The overall loss function is:

[0076]

[0077] Among them, α, β and γ are regularization parameters.

[0078] (5) Use the back propagation algorithm to train the network, iterate repeatedly, update the network parameters: ψ, ω1, ω2, φ, and continuously optimize the network until the loss converges.

[0079] (6) Obtain the encoding feature representation E of the two input images respectively ψ (X) model parameters.

[0080] (7) SAR image X1 and optical image X2 are obtained by optimizing encoder E ψ(X), and obtain the trained image features F1(X1) and F2(X2) respectively.

[0081] Step 5: Generate a difference map

[0082] (1) Based on the trained image global features F1(X1) and F2(X2), the probability map matrix P is used u The model is as follows:

[0083]

[0084] Among them, P u It is a probability map matrix. The value of each pixel represents the probability that the pixel position has not changed. Its size is the input image size, and λ is the regularization factor.

[0085] (2) When F1(X1)-F2(X2) is fixed, the probability map matrix P u The solution is:

[0086]

[0087] P u This is the difference map for change detection.

[0088] Step 6: Analyze the difference map

[0089] The local neighborhood information fuzzy C-means clustering algorithm (FLICM) [Krinidis S, Chatzis V. Arobust fuzzy local information C-means clustering algorithm [J]. IEEE transactions on image processing, 2010, 19 (5): 1328-1337.] is used to cluster the difference map P u Perform analysis to obtain the final binary change map.

[0090] Step 7: Output binary change map

[0091] Output the binary change result map obtained by difference map analysis.

[0092] The effect of the present invention can be further illustrated by the following simulation experiments:

[0093] Simulation conditions

[0094] The simulation experiment uses three sets of real remote sensing image data. Each set of data contains two registered remote sensing images from different periods and one manually annotated binary image.

[0095] As shown in Figures 6(a) and 6(b), the first set is the Tianhe Airport dataset. This heterogeneous image dataset consists of one SAR image and one optical image. This dataset includes a panchromatic image acquired from Landsat-7 and a color optical image acquired from Google Earth. Specifically, it depicts the vicinity of Wuhan Tianhe International Airport, showing the actual changes in the highway-airport area before and after the airport expansion. The images are 666×615 pixels in size and were acquired in July 2002 and June 2013, respectively. Due to the different acquisition methods of the two remote sensing images, the differences in their properties pose a challenge to the change detection task.

[0096] The second dataset, the NJUST dataset, is a homologous dataset consisting of two optical images, as shown in Figures 7(a) and 7(b). This dataset uses Google Earth images acquired in the Nanjing University of Science and Technology area in September 2013 and March 2019, respectively. The images are 523×447 pixels in size. They show the Nanjing University of Science and Technology building complex before and after its transformation. Due to seasonal variations, subtle changes occur in the roads, dormitories, and trees near the gymnasium. These are interfering factors and should be excluded.

[0097] As shown in Figures 8(a) and 8(b), the third dataset is the Yellow River dataset, a homologous dataset consisting of two SAR images. This dataset was acquired by Radarsat-2 in the Yellow River estuary region of China in June 2008 and June 2009. The original image size is 7666×7692, which is too large to collect prior information for integrating the images over such a large area. Therefore, a subregion size of 257×289 was selected for the change detection experiment in this invention. The reference image, obtained based on field observations and manual calibration, shows the changes in farmland along the Yellow River and the expansion of the river channel. It is worth noting that the image acquired in 2009 has a higher speckle noise level. The significant difference in speckle noise levels between the two images used may complicate the change detection process.

[0098] The simulation experiments were completed using C++ and Python under Windows 10 operating system. i Set to 5; for the Tianhe Airport dataset, block x i Set to 7; for the NJUST dataset, block x i Set to 11. The weight parameters α, β and γ are set to 0.6, 0.8 and 0.12 respectively. Encoder E ψ (x i ) consists of 4 layers of 2D convolutional layers and 1 layer of linear layer, with 64 output channels, 4×4 convolution kernel size, and Sigmoid activation function. Global feature discriminator It consists of 2 layers of 2D convolutional layers and 3 layers of linear layers, with 1 output channel and a convolution kernel size of 3×3. It is composed of three layers of two-dimensional convolutional layers, the number of output channels is 1, and the convolution kernel size is 1×1. φ The network consists of three linear layers with one output channel and a convolution kernel size of 3 × 3. The Adam optimizer is used for optimization with a step size of (1, 1), a batch size of 64, and a learning rate of 0.0001.

[0099] The present invention adopts six evaluation indicators, namely: false positive (FP), false negative (FN), overall classification errors (OE), percentage correct classification (PCC), Kappa coefficient (KC) and area under the ROC (Receiveroperating characteristic) curve (AUC).

[0100] Simulation content

[0101] The present invention uses the real Tianhe Airport dataset, NJUST dataset and Yellow River dataset to test the change detection performance of the algorithm. In order to test the performance of the algorithm of the present invention, the proposed local and global deep comparative feature learning network for change detection between images (Deep Comparative Learning based on Local and Global, DCL) is compared with the neighborhood-based ratio difference map operator (NR-FLICM) [Gong M, Cao Y, Wu Q.Aneighborhood-based ratio approachfor change detection in SAR images[J].IEEE Geoscience and Remote Sensing Letters,2011,9(2):307-311.], the convolutional wavelet network (CWNN) [Gao F, Wang X, Gao YH, et al.SeaIce Change Detection in SAR Images Based on Convolutional-Wavelet NeuralNetworks[J].IEEE Geoscience and Remote Sensing Letters,2019,16(8):1240-1244.], SCCN and conditional adversarial generative network (cGAN) [Niu XD, Gong MG, Zhan T, et al.AConditional Adversarial Network for Change Detection in Heterogeneous Images[J].IEEE Geoscience and Remote Sensing Letters,2019,16(1):45-49.] four algorithms are compared.

[0102] Analysis of simulation experiment results

[0103] Table 1 shows the comparison results of the first set of Tianhe Airport data sets under different change detection algorithms. Figure 9(a) to Figure 9(f) The following are the change detection results of the Tianhe dataset under different change detection algorithms. The experimental results show that the CWNN algorithm cannot distinguish the unchanged areas of heterogeneous images well on heterogeneous datasets, with a large number of false detection areas, and the detection effect is greatly reduced. Although SCCN has fewer false detection areas, it only detects some airport change areas, and the difference map is noisy and difficult to analyze. cGAN can detect airport change areas, but it is affected by edges and SAR noise, resulting in a large number of false detections. Figure 9(a) to Figure 9(f)It can be seen that the algorithm proposed in this paper outperforms other deep learning-based algorithms and can more accurately detect the changed areas of Tianhe Airport. The PCC, KC, and AUC values ​​are all significantly improved. The KC coefficient is 0.1425 higher than that of the well-performing SCCN. This is because the local and global deep contrast feature learning network proposed in this paper for inter-image change detection combines information between hierarchical features.

[0104] Table 1 Comparison of change detection evaluation indicators for the first group of Tianhe Airport datasets

[0105]

[0106] Table 2 shows the comparison results of the second set of NJUST datasets under different change detection algorithms. Figure 10(a) to Figure 10(f) The following is a graph showing the change detection results of the NJUST dataset under different change detection algorithms. The NJUST dataset consists of two images with deviations in shooting angles and obvious seasonal changes. Our task is to detect changes in buildings, and changes in grass and trees on both sides of the road are factors that interfere with detection. The experimental results show that the CWNN algorithm is unable to detect accurate results due to its narrow applicability and interference from many scene factors. Although the SCCN algorithm and the cGAN algorithm can detect some changed buildings, they are affected by inconsistent shooting angles and seasonal changes. The algorithm of the present invention can better overcome these interference factors, and PCC, KC and AUC are significantly improved. Therefore, the local and global deep contrast feature learning network for inter-image change detection proposed in the present invention can also be used for detection on homologous images, and its change detection effect is significant.

[0107] Table 2 Comparison of change detection evaluation indicators of the second set of NJUST datasets

[0108]

[0109] Table 3 shows the comparison results of the third set of Yellow River datasets under different change detection algorithms. Figures 11(a) to 11(f) Figure 2 shows the change detection results for the Yellow River dataset using different change detection algorithms. Since the CWNN algorithm is specifically designed for detecting homologous SAR images, it outperforms the proposed algorithm on this dataset. However, CWNN performs poorly on other datasets. While not as good as CWNN, the proposed algorithm performs better than other deep learning heterogeneous source detection algorithms and can also detect changed areas relatively well, validating the theoretical feasibility of this algorithm.

[0110] Table 3 Comparison of change detection evaluation indicators for the third set of Yellow River datasets

[0111]

[0112] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for constructing a local and global deep contrast feature learning network for inter-image change detection, characterized in that: The steps include: (1) Use the network to encode the input image to obtain global high-dimensional representation features and intermediate layer features; Input two images of the same scene at different times: X1∈R W×H and X2∈R W×H , W represents the width of the image, H represents the height of the image; randomly take blocks x of the original image by pixel i , i=1,2,...,N, block x i Pull into a one-dimensional vector as the encoder E ψ (x i ) input to obtain the global high-dimensional representation feature y of the feature domain ψ and the intermediate layer features M ψ , the global high-dimensional feature represents the last layer of features; the specific steps of encoding feature representation are as follows: 1) Encoder E ψ (X) Network structure An image X is used as the input of the encoding network, and the encoding feature representation is obtained through the convolutional layer and the fully connected layer. Among them, y ψ represents the high-dimensional feature representation of the encoded feature representation, M ψ represents the intermediate layer feature representation of the encoded feature representation, Indicates the nth fc fully connected layers, Indicates the nth cov convolutional layers; 2) Randomly take a block x from an input image X1 by pixel i , i=1,2,...,N, N represents the number of blocks in the X1 image, block x i Pull into a one-dimensional vector as encoder E ψ (x i ) input, and obtain the global representation feature y of the SAR image X1 in the feature domain ψ and the global intermediate layer feature M ψ ; 3) Randomly take a block x of pixels from another input image X2 i , i=1,2,...,N, N represents the number of blocks in the X2 image, block x i Pull into a one-dimensional vector as encoder E ψ (x i ) input, and obtain the global representation feature y of the optical image X2 in the feature domain ψ and the global intermediate layer feature M ψ ; (2) Construct global positive and negative sample data and local positive and negative sample data. The specific process is as follows: Global positive and negative sample data: intermediate layer feature M ψ As global positive sample data; merge the intermediate layer features of the first image of each batch with the end features of the batch to construct the global negative sample data M_prime ψ ; Local positive and negative sample data: the middle layer feature M ψ and the global high-dimensional feature y ψ Splicing to obtain local positive sample data y_M ψ ; The intermediate layer feature M_prime of the global negative sample ψ and high-dimensional features y ψ Splicing to obtain local negative sample data y_M_prime ψ ; (3) Construct contrast loss function: Global feature loss function Defined as: in, represents the global feature discriminator, S p represents the activation function, and E represents the expectation; Local feature loss function Defined as: in, represents the local feature discriminator; Prior loss function PRIOR φ,ψ for: PRIOR φ,ψ =-(log(P φ (and rand )+log(1-P φ (and ψ ))) Among them, P φ represents the prior discriminator, y rand is a random vector that obeys the uniform distribution U(0,1), and y ψ Keep the size consistent; The three parts of loss constitute the total loss function, which is defined as follows: Among them, α, β and γ are regularization parameters; (4) Solve the difference probability map matrix; (5) Cluster the difference probability map matrix and output the binary change result map.

2. The method for constructing a local and global deep contrast feature learning network for inter-image change detection according to claim 1, characterized in that: The global feature discriminator It is a discriminator of convolutional fully connected structure and a local feature discriminator is a 1×1 convolution discriminator, a priori discriminator P φ is a fully connected discriminator; activation function S p is the softplus function, and its expression is as follows: S p (x)=log(1+e x )。 3. The method for constructing a local and global deep contrast feature learning network for inter-image change detection according to claim 1, characterized in that: Solving the difference probability map matrix includes the following steps: 1) Based on the trained global image features F1(X1) and F2(X2), the probability map matrix P is used u The model is as follows: Among them, P u is a probability map matrix, the value of each pixel represents the probability that the pixel position has not changed, and its size is the input image size, and λ is the regularization factor; 2) When F1(X1)-F2(X2) is fixed, the probability map matrix P u The solution is: P u This is the difference map for change detection.

4. The method for constructing a local and global deep contrast feature learning network for inter-image change detection according to claim 1, characterized in that: The local neighborhood information fuzzy C-means clustering method is used to analyze the difference map and obtain the final binary change map.

Images