Remote sensing image change detection method based on modular learning network

Through a modular learning network based on Res2Net, the change detection task is decoupled, multi-scale bottleneck modules and difference modules are used, and specific loss function training is combined, the problem of low accuracy of existing change detection methods is solved, and the accuracy and efficiency are improved.

CN118918460AActive Publication Date: 2024-11-08CHONGQING UNIV
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202410912590.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-09
Publication Date
2024-11-08
Estimated Expiration
2044-07-09

AI Technical Summary

Technical Problem

The existing single-stream and double-stream structure change detection methods have low accuracy in practical applications, making it difficult to effectively improve network performance without requiring complex structural design.

Method used

Using a modular learning network based on Res2Net, by decoupling the change detection task into feature extraction and change recognition related to the detection task, a simplified network structure is designed, and a multi-scale bottleneck module and a differential module are used, combining the binary cross-quotient loss function and Dice loss function is trained.

Benefits of technology

Improves the accuracy and efficiency of change detection, simplifies the detection process, improves the performance and accuracy of the model, and the method is model-independent, and existing methods can be ported to a new framework for accuracy improvement.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118918460A_ABST
    Figure CN118918460A_ABST
Patent Text Reader

Abstract

The invention provides a remote sensing image change detection method based on a modular learning network in the field of change detection, and the method comprises the steps: carrying out the decoupling of a change detection method, taking the change detection as an object related to a detection task, and detecting the change of the object. Specifically, in the framework, two specialized components are designed: a task-centric learner and a difference detector, and the task-centric learner is based on an encoder-decoder structure and is designed to identify objects related to tasks. The difference detector combines a difference module and a feature pyramid network for identifying changes between targets. The method can greatly improve the prediction precision of the network.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of deep learning in artificial intelligence, and specifically is a remote sensing image change detection method based on a modular learning network. Background Art

[0002] Change detection is an intersection between computer vision and remote sensing. It is also the basis for some advanced computer vision tasks and is of great significance to ecological protection, environmental monitoring, land resource utilization, etc. Change detection is to effectively predict the area where the previous and next images have changed through a pair of RGB images.

[0003] Although the field of change detection has developed rapidly, the current mainstream single-stream and dual-stream structure methods require the simultaneous implementation of task-related object detection and change area detection in the decoder part, resulting in lower accuracy in practical applications.

[0004] Therefore, how to provide a change detection method to improve the accuracy of the network and improve the performance of the existing model without complex structural design has become an urgent problem to be solved. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a remote sensing image change detection method based on a modular learning network, while improving the accuracy and efficiency of existing change detection.

[0006] The present invention provides a remote sensing image change detection method based on a modular learning network, the method comprising the following steps:

[0007] Step S10: Building a modular learning network based on Res2Net.

[0008] Step S20: Acquire labeled sample images to construct a data set, and use the sample images in the data set to train the modular learning network.

[0009] Step S30: input the image to be tested into the trained modular learning network, and the modular learning network outputs the probability value corresponding to the area. When the probability value is greater than or equal to the threshold, the area is marked as a changed area, otherwise it is marked as an unchanged area.

[0010] Furthermore, the modular learning network in step S10 includes:

[0011] A first convolution kernel of size 3×3, a second convolution kernel of size 3×3, a third convolution kernel of size 3×3, a fourth convolution kernel of size 3×3, a fifth convolution kernel of size 3×3, a sixth convolution kernel of size 3×3, a seventh convolution kernel of size 3×3, an eighth convolution kernel of size 3×3, a ninth convolution kernel of size 1×1, a tenth convolution kernel of size 1×1, an eleventh convolution kernel of size 1×1, a twelfth convolution kernel of size 1×1, a channel A first ReLU activation function with a channel number of 64, a second ReLU activation function with a channel number of 64, a third ReLU activation function with a channel number of 64, a first subnet with a channel number of 256, a second subnet with a channel number of 512, a third subnet with a channel number of 1024, a fourth subnet with a channel number of 2048, a first difference module with a channel number of 256, a second difference module with a channel number of 128, a third difference module with a channel number of 64, and a fourth difference module with a channel number of 32.

[0012] The first subnetwork includes three multi-scale bottleneck modules arranged in parallel and connected horizontally, respectively. 1,1 , S 1,2 , S 1,3 The second subnetwork includes four multi-scale bottleneck modules that are arranged in parallel and connected horizontally, and are represented by S 2,1 , S 2,2 ,……,S 2,4 The third subnetwork includes six multi-scale bottleneck modules that are arranged in parallel and connected horizontally, and are represented by S 3,1 , S 3,2 ,……,S 3,6 The fourth subnetwork includes three multi-scale bottleneck modules arranged in parallel and connected horizontally, which are represented by S 4,1 , S 4,2 , S 4,3 express.

[0013] The output end of the first convolution kernel is connected to the input end of the first ReLU activation function, the output end of the first ReLU activation function is connected to the input end of the second convolution kernel, and the output end of the second convolution kernel is connected to the input end of the second ReLU activation function, so as to change the number of channels of the image feature map; the output end of the second ReLU activation function is connected to the input end of the first subnet, so as to extract image features; the first subnet, the second subnet, the third subnet and the fourth subnet are arranged from left to right in sequence.

[0014] The first subnet outputs a feature map S out,1 , the second subnet outputs the feature map S out,2 , the third subnet outputs a feature map S out,3, the fourth subnet outputs a feature map S out,4 , the feature map S out,4 After a bilinear interpolation upsampling, the resolution of the feature map is increased to the same level as the feature map S out,3 Consistent with the feature map S out,3 Stack the elements to form the feature map F out,3 In the feature map F out,3 Then, the third convolution kernel and the fourth convolution kernel are convolved to obtain the feature map F out,up3 ; The feature map F out,up3 After a bilinear interpolation upsampling, the resolution of the feature map is increased to the same level as the feature map S out,2 Consistent with the feature map S out,2 Stack the elements to get the feature map F out,2 In the feature map F out,2 Then, the fifth convolution kernel and the sixth convolution kernel are convolved to obtain the feature map F out,up2 ; The feature map F out,up2 After a bilinear interpolation upsampling, the resolution of the feature map is increased to the same level as the feature map S out,1 Consistent with the feature map S out,1 Stack the elements to get the feature map F out,1 In the feature map F out,1 Then, the seventh convolution kernel and the eighth convolution kernel are convolved to obtain the feature map F out,up1 .

[0015] The feature map S out,4 Send it to the first difference module to get the feature map F diff,4 ; The characteristic graph F out,up3 The output after inputting the second difference module is convolved with the ninth convolution kernel to obtain the feature map F f,up3 ; The feature map F diff,4 After a bilinear interpolation upsampling, the resolution of the feature map is increased to the same level as the feature map F f,up3 Consistent with the feature map F f,up3 Add the elements to get the feature map F p,3 In the feature map F out,up2 The output after inputting the third difference module is convolved with the tenth convolution kernel to obtain the feature map F f,up2 In the feature map F p,3 After a bilinear interpolation upsampling, the resolution of the feature map is increased to the same level as the feature map F f,up2 Consistent with the feature map F f,up2 Add the elements to get the feature map F p,2 In the feature map F out,up1After inputting the fourth difference module, the output is convolved with the eleventh convolution kernel to obtain the feature map F f,up1 In the feature map F p,2 After a bilinear interpolation upsampling, the resolution of the feature map is increased to the same level as the feature map F f,up1 Consistent with the feature map F f,up1 Add the elements to get the feature map F p,1 In the feature map F p,1 Then the twelfth convolution kernel operation is performed to obtain the feature map F out .

[0016] Furthermore, in step S20, the process of training the modular learning network includes:

[0017] Step S21, obtain a sample image in the data set, adjust the sample image to an RGB image with a size of 256×256, and input it into the constructed modular learning network.

[0018] Step S22: The modular learning network outputs a feature map F out .

[0019] Step S23: Calculate the L of the modular learning network using the binary cross quotient loss function and the true label bce Loss value:

[0020]

[0021] Where N represents all the pixels in the sample image; y i Represents the true label value, y i ∈{0,1}, represents the probability value of the change area predicted by the modular learning network,

[0022] The Dice loss function and the true label are used to calculate the Dice loss value of the modular learning network:

[0023]

[0024] Where N represents all the pixels in the sample image; y i Represents the true label value, y i ∈{0,1}, Represents the probability value predicted by the network, y i ∈[0,1], ∈ represents a constant.

[0025] Weighted L bce The loss value and Dice loss value give the total loss L of the modular learning network:

[0026] L=αLbce +βL dice

[0027] Where α and β are weights, α=β=1.

[0028] Step S24: According to the total loss L, the parameters of the modular learning network are reversely updated using the gradient descent method until the total loss L reaches a preset threshold, and the training ends.

[0029] Furthermore, the steps of step S30 include:

[0030] After the image to be tested is resized to an RGB image of size 256×256, it is input into the trained modular learning network to calculate the probability value of the area in the RGB image; the area with a probability value greater than or equal to 0.5 is marked as a changed area, and the area with a probability value less than 0.5 is marked as an unchanged area.

[0031] Compared with the prior art, the beneficial effects of the present invention are:

[0032] By decoupling the change detection task into the features related to the detection task and the identification of changes, it is ensured that each module in the network is only for one subtask. The network designed according to this principle simplifies the entire change detection process and improves the performance and accuracy of the model. In addition, the present invention is a model-independent invention, and the existing methods can be transplanted to the framework proposed by the present invention, thereby improving the accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The present invention will be further described below in conjunction with embodiments with reference to the accompanying drawings.

[0034] Figure 1 Flow chart of the method of the present invention.

[0035] Figure 2 Schematic diagram of the structure of the modular learning network.

[0036] Figure 3 Schematic diagram of the structure of the multi-scale bottleneck module.

[0037] Figure 4 It is a structural diagram of the difference module. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only partial embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] In order to better understand the technical solution of the present invention, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0040] Embodiment 1: A remote sensing image change detection method based on a modular learning network, such as Figure 1 The method comprises the following steps:

[0041] Step S10: Building a modular learning network based on Res2Net.

[0042] Step S20: Acquire labeled sample images to construct a data set, and use the sample images in the data set to train the modular learning network.

[0043] That is, a large number of sample images are obtained in advance, and after the change area of ​​each sample image is marked, they are divided into training set, verification set and test set. The sample images in the training set are input into the modular learning network for training, and then the trained modular learning network is verified using the verification set to determine whether the loss value reaches the preset threshold; the test set is input into the trained modular learning network to finally determine the capability achieved by the modular learning network.

[0044] Step S30: input the image to be tested into the trained modular learning network to calculate the final changed area.

[0045] For better, please refer to Figure 2 , the modular learning network includes:

[0046] A first convolution kernel of size 3×3, a second convolution kernel of size 3×3, a third convolution kernel of size 3×3, a fourth convolution kernel of size 3×3, a fifth convolution kernel of size 3×3, a sixth convolution kernel of size 3×3, a seventh convolution kernel of size 3×3, an eighth convolution kernel of size 3×3, a ninth convolution kernel of size 1×1, a tenth convolution kernel of size 1×1, an eleventh convolution kernel of size 1×1, a twelfth convolution kernel of size 1×1, a channel A first ReLU activation function with a channel number of 64, a second ReLU activation function with a channel number of 64, a third ReLU activation function with a channel number of 64, a first subnet with a channel number of 256, a second subnet with a channel number of 512, a third subnet with a channel number of 1024, a fourth subnet with a channel number of 2048, a first difference module with a channel number of 256, a second difference module with a channel number of 128, a third difference module with a channel number of 64, and a fourth difference module with a channel number of 32.

[0047] The first subnetwork includes three multi-scale bottleneck modules arranged in parallel and connected horizontally, respectively. 1,1 , S 1,2, S 1,3 The second subnetwork includes four multi-scale bottleneck modules that are arranged in parallel and connected horizontally, and are represented by S 2,1 , S 2,2 ,……,S 2,4 The third subnetwork includes six multi-scale bottleneck modules that are arranged in parallel and connected horizontally, and are represented by S 3,1 , S 3,2 ,……,S 3,6 The fourth subnetwork includes three multi-scale bottleneck modules arranged in parallel and connected horizontally, which are represented by S 4,1 , S 4,2 , S 4,3 express.

[0048] The output end of the first convolution kernel is connected to the input end of the first ReLU activation function, the output end of the first ReLU activation function is connected to the input end of the second convolution kernel, and the output end of the second convolution kernel is connected to the input end of the second ReLU activation function, so as to change the number of channels of the image feature map; the output end of the second ReLU activation function is connected to the input end of the first subnet, so as to extract image features; the first subnet, the second subnet, the third subnet and the fourth subnet are arranged from left to right in sequence.

[0049] The first subnet outputs a feature map S out,1 , the second subnet outputs the feature map S out,2 , the third subnet outputs a feature map S out,3 , the fourth subnet outputs a feature map S out,4 , the feature map S out,4 After a bilinear interpolation upsampling, the resolution of the feature map is increased to the same level as the feature map S out,3 Consistent with the feature map S out,3 Stack the elements to form the feature map F out,3 In the feature map F out,3 Then, the third convolution kernel and the fourth convolution kernel are convolved to obtain the feature map F out,up3 ; The feature map F out,up3 After a bilinear interpolation upsampling, the resolution of the feature map is increased to the same level as the feature map S out,2 Consistent with the feature map S out,2 Stack the elements to get the feature map F out,2 In the feature map F out,2 Then, the fifth convolution kernel and the sixth convolution kernel are convolved to obtain the feature map F out,up2 ; The feature map F out,up2 After a bilinear interpolation upsampling, the resolution of the feature map is increased to the same level as the feature map S out,1 Consistent with the feature map Sout,1 Stack the elements to get the feature map F out,1 In the feature map F out,1 Then, the seventh convolution kernel and the eighth convolution kernel are convolved to obtain the feature map F out,up1 .

[0050] The feature map S out,4 Send it to the first difference module to get the feature map F diff,4 ; The characteristic graph F out,up3 The output after inputting the second difference module is convolved with the ninth convolution kernel to obtain the feature map F f,up3 ; The feature map F diff,4 After a bilinear interpolation upsampling, the resolution of the feature map is increased to the same level as the feature map F f,up3 Consistent with the feature map F f,up3 Add the elements to get the feature map F p,3 In the feature map F out,up2 The output after inputting the third difference module is convolved with the tenth convolution kernel to obtain the feature map F f,up2 In the feature map F p,3 After a bilinear interpolation upsampling, the resolution of the feature map is increased to the same level as the feature map F f,up2 Consistent with the feature map F f,up2 Add the elements to get the feature map F p,2 In the feature map F out,up1 After inputting the fourth difference module, the output is convolved with the eleventh convolution kernel to obtain the feature map F f,up1 In the feature map F p,2 After a bilinear interpolation upsampling, the resolution of the feature map is increased to the same level as the feature map F f,up1 Consistent with the feature map F f,up1 Add the elements to get the feature map F p,1 In the feature map F p,1 Then the twelfth convolution kernel operation is performed to obtain the feature map F out .

[0051] Preferably, in step S21, a sample image in the data set is obtained, the sample image is adjusted to an RGB image of a size of 256×256, and then input into the constructed modular learning network;

[0052] Step S22: The modular learning network outputs a feature map F out ;

[0053] Step S23: Calculate the L of the modular learning network using the binary cross quotient loss function and the true label bce Loss value:

[0054]

[0055] Where N represents all the pixels in the sample image; y i Represents the true label value, y i ∈{0,1}, represents the probability value of the change area predicted by the modular learning network,

[0056] The Dice loss function and the true label are used to calculate the Dice loss value of the modular learning network:

[0057]

[0058] Where N represents all the pixels in the sample image; y i Represents the true label value, y i ∈{0,1}, Represents the probability value predicted by the network, y i ∈[0,1], ∈ represents a constant. To prevent the denominator from being 0, the value of the present invention is 0.000001.

[0059] Weighted L bce The loss value and Dice loss value give the total loss L of the modular learning network:

[0060] L=αL bce +βL dice

[0061] Where α and β are weights, α=β=1.

[0062] Step S24: According to the total loss L, the parameters of the modular learning network are reversely updated using the gradient descent method until the total loss L reaches a preset threshold, and the training ends.

[0063] Preferably, step S30 includes:

[0064] After adjusting the image to be tested to an RGB image of size 256×256, it is input into the trained modular learning network;

[0065] Calculate the probability value of the region in the RGB image; mark the region with a probability value greater than or equal to 0.5 as a changed region, and mark the region with a probability value less than 0.5 as an unchanged region.

[0066] Experimental analysis

[0067] Table 1 Performance comparison on the SYSU test set

[0068] Comparison Method F1 IoU OA BIT 78.15 64.13 90.18 HCGMNet 79.76 66.33 91.12 DMINet 80.30 67.09 91.30 The present invention 82.39 70.05 91.94

[0069] The experimental data in Table 1 are the comparison results on the SYSU test set. The table compares the performance of classic change detection methods on three evaluation indicators: F1, IoU, and OA. The present invention has achieved the most advanced results on multiple evaluation indicators. Compared with the BIT method, the present invention has improved the F1 and IoU evaluation indicators by 5.4% and 9.2%, respectively. Compared with the HCGMNet method, the present invention has improved the F1 and IoU evaluation indicators by 2.63 and 2.72 percentage points, respectively. Compared with the most advanced DMINet method, the present invention has surpassed the DMINet method in all indicators, especially in the F1 evaluation indicator, which has improved by 2.09 percentage points. The experimental results show that the performance of the present invention is superior to that of existing change detection algorithms. These substantial improvements can be attributed to the modular learning relationship proposed by the present invention. These improvements verify the effectiveness and superiority of the method of the present invention.

[0070] The core idea of ​​the present invention is a modular learning method. Specifically, change detection is considered as a combined task of locating task-related objects and detecting changes. A task-related detector is built based on the Res2Net network, and then the difference detector proposed in the present invention is used to detect specific changes. A large number of experiments on very challenging datasets have verified that the proposed method has good performance and advantages.

[0071] Although the specific implementation modes of the present invention are described above, those skilled in the art should understand that the specific implementation modes described are only illustrative and are not intended to limit the scope of the present invention. Equivalent modifications and changes made by those skilled in the art in accordance with the spirit of the present invention should be included in the scope of protection of the claims of the present invention.

Claims

1. A remote sensing image change detection method based on modular learning network, characterized by: The method comprises the following steps: Step S10, building a modular learning network based on Res2Net; Step S20, obtaining labeled sample images to construct a data set, and using the sample images in the data set to train the modular learning network; Step S30: input the image to be tested into the trained modular learning network, and the modular learning network outputs the probability value corresponding to the area. When the probability value is greater than or equal to the threshold, the area is marked as a changed area, otherwise it is marked as an unchanged area.

2. The remote sensing image change detection method based on modular learning network as claimed in claim 1, characterized in that: The modular learning network in step S10 includes: A first convolution kernel of size 3×3, a second convolution kernel of size 3×3, a third convolution kernel of size 3×3, a fourth convolution kernel of size 3×3, a fifth convolution kernel of size 3×3, a sixth convolution kernel of size 3×3, a seventh convolution kernel of size 3×3, an eighth convolution kernel of size 3×3, a ninth convolution kernel of size 1×1, a tenth convolution kernel of size 1×1, an eleventh convolution kernel of size 1×1, a twelfth convolution kernel of size 1×1, a channel A first ReLU activation function with a channel number of 64, a second ReLU activation function with a channel number of 64, a third ReLU activation function with a channel number of 64, a first subnet with a channel number of 256, a second subnet with a channel number of 512, a third subnet with a channel number of 1024, a fourth subnet with a channel number of 2048, a first difference module with a channel number of 256, a second difference module with a channel number of 128, a third difference module with a channel number of 64, and a fourth difference module with a channel number of 32; The first subnetwork includes three multi-scale bottleneck modules arranged in parallel and connected horizontally, respectively. 1,1 , S 1,2 , S 1,3 The second subnetwork includes four multi-scale bottleneck modules that are arranged in parallel and connected horizontally, and are represented by S 2,1 , S 2,2 ,……,S 2,4 The third subnetwork includes six multi-scale bottleneck modules that are arranged in parallel and connected horizontally, and are represented by S 3,1 , S 3,2 ,……,S 3,6 The fourth subnetwork includes three multi-scale bottleneck modules arranged in parallel and connected horizontally, which are represented by S 4,1 , S 4,2 , S 4,3 express; The output end of the first convolution kernel is connected to the input end of the first ReLU activation function, the output end of the first ReLU activation function is connected to the input end of the second convolution kernel, and the output end of the second convolution kernel is connected to the input end of the second ReLU activation function, so as to change the number of channels of the image feature map; the output end of the second ReLU activation function is connected to the input end of the first subnet, so as to extract image features; the first subnet, the second subnet, the third subnet and the fourth subnet are arranged from left to right in sequence; The first subnet outputs a feature map S out,1 , the second subnet outputs the feature map S out,2 , the third subnet outputs a feature map S out,3 , the fourth subnet outputs a feature map S out,4 , the feature map S out,4 After a bilinear interpolation upsampling, the resolution of the feature map is increased to the same level as the feature map S out,3 Consistent with the feature map S out,3 Stack the elements to form the feature map F out,3 In the feature map F out,3 Then, the third convolution kernel and the fourth convolution kernel are convolved to obtain the feature map F out,up3 ; The feature map F out,up3 After a bilinear interpolation upsampling, the resolution of the feature map is increased to the same level as the feature map S out,2 Consistent with the feature map S out,2 Stack the elements to get the feature map F out,2 In the feature map F out,2 Then, the fifth convolution kernel and the sixth convolution kernel are convolved to obtain the feature map F out,up2 ; The feature map F out,up2 After a bilinear interpolation upsampling, the resolution of the feature map is increased to the same level as the feature map S out,1 Consistent with the feature map S out,1 Stack the elements to get the feature map F out,1 In the feature map F out,1 Then, the seventh convolution kernel and the eighth convolution kernel are convolved to obtain the feature map F out,up1 ; The feature map S out,4 Send it to the first difference module to get the feature map F diff,4 ; The characteristic graph F out,up3 The output after inputting the second difference module is convolved with the ninth convolution kernel to obtain the feature map F f,up3 ; The feature map F diff,4 After a bilinear interpolation upsampling, the resolution of the feature map is increased to the same level as the feature map F f,up3 Consistent with the feature map F f,up3 Add the elements to get the feature map F p,3 In the feature map F out,up2 The output after inputting the third difference module is convolved with the tenth convolution kernel to obtain the feature map F f,up2 In the feature map F p,3 After a bilinear interpolation upsampling, the resolution of the feature map is increased to the same level as the feature map F f,up2 Consistent with the feature map F f,up2 Add the elements to get the feature map F p,2 In the feature map F out,up1 After inputting the fourth difference module, the output is convolved with the eleventh convolution kernel to obtain the feature map F f,up1 In the feature map F p,2 After a bilinear interpolation upsampling, the resolution of the feature map is increased to the same level as the feature map F f,up1 Consistent with the feature map F f,up1 Add the elements to get the feature map F p,1 In the feature map F p,1 Then the twelfth convolution kernel operation is performed to obtain the feature map F out .

3. A remote sensing image change detection method based on a modular learning network as claimed in claim 1 or 2, characterized in that: In step S20, the process of training the modular learning network includes: Step S21, obtaining a sample image in the data set, adjusting the sample image to an RGB image of size 256×256, and inputting the constructed modular learning network; Step S22: The modular learning network outputs a feature map F out ; Step S23: Calculate the L of the modular learning network using the binary cross quotient loss function and the true label bce Loss value: Where N represents all the pixels in the sample image; y i Represents the true label value, y i ∈{0,1}, represents the probability value of the change area predicted by the modular learning network, The Dice loss function and the true label are used to calculate the Dice loss value of the modular learning network: Where N represents all the pixels in the sample image; y i Represents the true label value, y i ∈{0,1}, Represents the probability value predicted by the network, y i ∈[0,1], ∈ represents a constant; Weighted L bce The loss value and Dice loss value give the total loss L of the modular learning network: L=αL bce +βL dice Where α and β are weights, α=β=1. Step S24: According to the total loss L, the parameters of the modular learning network are reversely updated using the gradient descent method until the total loss L reaches a preset threshold, and the training ends.

4. The remote sensing image change detection method based on modular learning network as claimed in claim 1, characterized in that: The steps of step S30 include: After adjusting the image to be tested to an RGB image of size 256×256, it is input into the trained modular learning network; Calculate the probability value of the region in the RGB image; mark the region with a probability value greater than or equal to 0.5 as a changed region, and mark the region with a probability value less than 0.5 as an unchanged region.

Citation Information

Patent Citations

  • Twin network change detection model based on deep learning

    CN114419464A

  • Multi-model fusion remote sensing image classification method

    CN116486275A

  • Anti-attention guava semantic segmentation detection equipment

    CN116597439A

  • Medical image segmentation method based on u-network

    US20220398737A1