Damage degree identification method, system and equipment based on image data and medium

Abstract

The invention relates to the technical field of building damage identification, and provides a damage degree identification method, system and equipment based on image data and a medium, which can realize cross-disaster double-temporal building damage degree evaluation and is used for building damage degree evaluation and building positioning through an end-to-end convolutional neural network. And experiments are carried out on an xBD data set, so that the network provided by the method provided by the invention is proved to have better performance. According to the method, dual-time-phase feature fusion is adopted to effectively fuse pre-disaster and post-disaster feature maps, and feature fusion is carried out on dual-time-phase feature maps to enhance feature representation; and meanwhile, style unification is carried out by adopting fast Fourier transform, so that the problem of different dual-time-phase data imaging styles caused by different imaging conditions is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of building damage identification, and in particular to a damage degree identification method, system, equipment and medium based on image data. Background Art

[0002] With the widespread application of remote sensing satellites and airborne sensors, remote sensing image data has shown an explosive growth in geometric series, not only accumulating rapidly in quantity, but also significantly improving in quality. This significant change has greatly promoted the development of the field of intelligent interpretation of remote sensing images. In particular, deep learning models represented by convolutional neural networks have shown excellent performance in the fields of image segmentation and target detection, which has opened up new possibilities for accurately assessing the degree of damage to buildings by combining neural network models with remote sensing images. This technical fusion not only improves the accuracy and efficiency of the assessment, but also provides strong technical support for disaster management, urban planning and other fields. The existing work of assessing the degree of damage to buildings from satellite images can be roughly divided into the following two types:

[0003] (1) Determination of single-phase image data. The building damage degree recognition method based on single-phase images adopts the idea of ​​semantic segmentation, using post-disaster high-resolution remote sensing image data with detailed texture information and context information as network input. By detecting the spectrum, texture, edge, spatial relationship, structure, shape and shadow of the building in the image that has changed due to damage, the building damage degree can be recognized. However, due to the impact of natural disasters, the originally complete building vector boundaries were destroyed and confused with the surrounding objects, so misclassification and omissions are inevitable.

[0004] (2) Determination of dual-temporal image data before and after disasters. The building damage degree identification and assessment method based on dual-temporal images adopts the idea of ​​change detection to assess the degree of building damage. However, compared with change detection, dual-temporal building damage degree determination requires not only detecting the changed areas, that is, damaged buildings, but also detecting the unchanged areas, that is, undamaged buildings, and distinguishing irrelevant background changes. Different imaging conditions between dual-temporal phases, such as differences in solar altitude angle, atmospheric conditions, seasonal changes, etc., may lead to significant changes in spectral values, thereby affecting the accuracy of the detection results.

[0005] It can be seen that in recent years, natural disasters have occurred frequently. As the main carrier of human social activities and economic development, building areas are often the areas with the most serious losses after natural disasters. Rapid assessment of the damage degree of buildings in disaster areas through remote sensing images can minimize losses. However, most of the existing studies on the assessment of the degree of damage to buildings in natural disasters are aimed at a specific type of natural disaster, and lack a research framework and evaluation criteria for the assessment of the degree of damage to buildings across disasters. At the same time, buildings with different building materials and architectural styles will suffer different degrees of damage after suffering different types of natural disasters; at the same time, the inconsistent styles of dual-phase images with different imaging conditions are also challenges that must be faced in the work of determining the degree of damage to buildings. Summary of the invention

[0006] In view of this, the present invention provides a method, system, device and medium for identifying the degree of damage based on image data, which can realize the dual-temporal assessment of the degree of damage to buildings across disaster types.

[0007] To achieve the above object, the present invention proposes a method for identifying the degree of damage based on image data, comprising the following steps:

[0008] Step 1: Establish a backbone network;

[0009] Step 2: Design the encoder, specifically: perform feature extraction and channel deepening by setting up double-layer convolution, and change the size of the feature map by maximum pooling or convolution;

[0010] Step 3: The bi-phase encoder features are passed into the CE module to perform feature semi-exchange operation;

[0011] Step 4, integrating bi-temporal features;

[0012] Step 5, design the encoder: set the decoder to include two different decoders, one decoder is responsible for building positioning; the other decoder is responsible for determining the degree of damage to the building and evaluating and grading the degree of change in the image.

[0013] Wherein, the step 4 specifically includes:

[0014] The difference between the two feature maps is calculated to identify the areas where the actual surface has changed significantly; the attention weight matrix is ​​multiplied with the difference feature map, and the final preliminary change feature map is obtained through deep separable convolution to perform feature fusion on all feature maps; the aggregated spatial features are passed to two one-dimensional convolutions to determine the dual-temporal channel weights of the input dual-temporal features, and then Softmax is used in the dual-temporal channel weights to make their sum equal to 1, and the three weights are compared to determine the higher value between them, thereby determining the important part between the dual-temporal features in the channel dimension; the dual-temporal space weights are determined in the same way in the spatial branch, and the final attention weight matrix is ​​determined by adding the dual-temporal space weights to the dual-temporal channel weights, thereby calculating the final change difference feature map and determining the important part between the dual-temporal features in the spatial dimension; the dual-temporal channel weights and the dual-temporal space weights are combined to obtain the dual-temporal weights to determine the important part between the dual-temporal features; the dual-temporal weights are multiplied by the dual-temporal features to fuse the dual-temporal features.

[0015] This also includes the use of fast Fourier transform to unify the styles of two-phase images.

[0016] Among them, in the step 2, in the encoder part, the incoming feature map is firstly convolved with a downsampling step size of 2, and Haar wavelet transform downsampling is used as the residual connection part to retain more feature information; in the feature transfer of the model, the channel separation is performed on the incoming feature map, and only half of the feature map is subjected to deep feature extraction.

[0017] Among them, in the step 2, the CBAM attention module is introduced to capture global and local key features and fuse information of channel and spatial dimensions through an adaptive attention weighting mechanism; important feature information is automatically highlighted through residual connections while retaining the details of the original features.

[0018] Among them, a parameter-sharing network encoder is constructed in the encoder part for feature extraction.

[0019] The present invention also provides a damage degree recognition system based on image data, including a backbone network, an encoder, a channel switching module, a time series feature fusion module DFMA and a decoder.

[0020] Among them, the backbone network is the Unet network model; a double-layer convolution is set in the encoder for feature extraction and channel deepening, and the feature map size is changed through maximum pooling or convolution to complete feature extraction; the channel exchange module passes the dual-phase encoder features to the CE module for feature semi-exchange operation; the temporal feature fusion module is used to fuse the dual-temporal features and send them to the decoder; the decoder includes decoder a and decoder b; decoder a is responsible for building positioning, and locates the building on the original T1 and T2 feature maps passed in; the feature map generated at time T2 is used as auxiliary positioning, and is only involved in training, and is not involved in calculation during verification and reasoning; decoder b is responsible for judging and reasoning the degree of building damage, evaluating and grading the degree of change in the image, and multiplying the building positioning mask obtained by the image data at time T1 with the building damage degree map by the principal element, thereby obtaining an instantiated building damage degree positioning map.

[0021] It also includes a dual-phase style transfer module, which uses fast Fourier transform to unify the styles of two-phase images.

[0022] The present invention also provides an electronic device, which includes a processor and a memory for storing executable instructions of the processor; the processor is used to read the executable instructions from the memory and execute the instructions to implement the method for identifying the degree of damage based on image data described in the present invention.

[0023] The present invention also provides a computer-readable storage medium, wherein the storage medium stores a computer program, and the computer program is used to execute the damage degree identification method based on image data described in the present invention.

[0024] Beneficial effects:

[0025] 1. In the method of the present invention, an end-to-end convolutional neural network is used for building damage assessment and building positioning, and experiments on the xBD dataset prove that the network proposed by the method of the present invention has better performance.

[0026] 2. In the method of the present invention, dual-phase feature fusion is adopted to effectively fuse the pre-disaster and post-disaster feature maps, and the feature representation is enhanced by fusing the dual-phase feature maps. At the same time, fast Fourier transform is used for style unification to reduce the problem of different imaging styles of dual-phase data caused by different imaging conditions.

[0027] 3. In the method of the present invention, the xBD dataset is used as the basic dataset. This dataset is the largest and highest quality public building damage assessment dataset to date, covering 6 types and 19 natural disaster events in different regions, and the image data covers more than 45,000 square kilometers.

[0028] 4. The system of the present invention is used to implement the method of the present invention, which is used for building damage assessment and building positioning through an end-to-end convolutional neural network. Experiments on the xBD dataset have proved that the network proposed in the present invention has better performance.

[0029] 5. In the system of the present invention, a dual-phase feature fusion module DFMA is proposed for establishing a mapping relationship model in dual-phase image data, and fast Fourier transform is used for style unification to reduce the problem of different imaging styles of dual-phase data caused by different imaging conditions.

[0030] 6. In the device of the present invention, an end-to-end convolutional neural network is used for building damage assessment and building positioning, and experiments on the xBD data set prove that the network proposed in the present invention has better performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 The figure is a flow chart of a method according to an embodiment of the present invention.

[0032] Figure 2 Schematic diagram of bi-temporal style conversion according to an embodiment of the present invention.

[0033] Figure 3 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0034] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0035] The present invention provides a method for identifying the degree of damage based on image data, which can quickly and accurately evaluate the degree of damage to buildings from remote sensing image data, and has important guiding significance for post-disaster emergency rescue work, post-disaster reconstruction and academic research. Figure 1 As shown, the following steps are included:

[0036] Step 1: Establish a backbone network.

[0037] Specifically, this embodiment selects the Unet network model as the backbone network. Unet has shown significant advantages in image segmentation tasks through its encoder-decoder architecture and jump connection design. It can effectively capture the contextual information in the image and accurately restore the details of the image through feature fusion, thereby achieving high-quality image segmentation results. In addition, Unet has less demand for training data and has strong scalability and adaptability, making it highly valuable in various image segmentation tasks.

[0038] Step 2: Design the encoder.

[0039] Specifically, in this embodiment, the original network structure of Unet is relatively simple. The feature extraction in the encoder part is mainly performed by setting a double-layer convolution for feature extraction and channel deepening, and the feature map size is changed by maximum pooling or convolution to complete the feature extraction. Although the double convolution layer can gradually extract rich feature information and enhance nonlinear modeling capabilities, it also causes the loss of low-level features and increases computational complexity and overfitting risks.

[0040] In the encoder part, the incoming feature map is first convolved with a downsampling step of 2, and Haar wavelet transform downsampling is used as the residual connection part to retain more feature information. In the feature transfer of the model, in order to reduce parameter operations, the incoming feature map is channel separated, and only half of the feature map is subjected to deep feature extraction. The CBAM attention module is introduced, and the global and local key features are accurately captured through the adaptive attention weighting mechanism, and the information of the channel and spatial dimensions is integrated. Important feature information is also automatically highlighted through residual connections, while retaining the details of the original features. This strategy effectively realizes the automatic selection and adaptive enhancement of features, greatly improving the model's ability to model the target area.

[0041] In order to enhance the network model's ability to extract the vector boundaries of damaged buildings after disasters, a parameter-sharing network encoder is constructed in the encoder part for feature extraction.

[0042] Step 3, the bi-temporal encoder features are passed to the CE module for feature semi-exchange operation, so that each branch contains bi-temporal features. This means that the bi-temporal features are connected, and each branch can determine the area of ​​the building by itself. Based on the change area, the features with only single time information before channel exchange contain rich spatial features, which can be used to refine the change area and accurately locate the change objects in the same time phase.

[0043] Step 4, by fusing the dual-time features, all changing targets can be accurately located, which can be expressed as the following formula:

[0044] T′1,T′2=M*T1+(1-M)*T2

[0045] Among them, T′1 and T′2 represent the fusion of dual-time features. T1 is the pre-disaster image data. The buildings in the image are not damaged and have complete and regular building vector boundaries. The spatial information of the feature map is enhanced through the spatial attention mechanism, which prompts the model to be more interested in the significant features of the local area at time T1, focusing on the building area; for the image data at time T2, due to damage caused by natural disasters, the complete building boundaries are destroyed, and the spatial information position is relatively chaotic and complex. At this time, the channel attention mechanism is performed on the feature map at time T2. By highlighting certain feature channels, the feature change type of the object in phase B can be better captured; M represents the weight of T1.

[0046] How to establish a mapping relationship model between dual-phase image data is also a problem that must be considered. There are three main fusion methods: simple fusion directly merges the features of the two phases, but may ignore the complex relationship between them; convolution enhancement uses convolution operations to capture the local relationship and spatial context information between features to generate richer fusion features; and attention enhancement introduces an attention mechanism to perform weighted processing according to the importance of the features to highlight key features and improve detection accuracy and robustness. The choice of fusion method depends on the specific application scenario and requirements. The specific time series feature fusion is as follows:

[0047] In the change detection model, the fusion methods from the bi-temporal feature maps can be divided into: simple fusion, convolution enhancement and attention enhancement.

[0048] The simple fusion method generates the change region by performing algebraic operations on the feature maps and splicing them together. It does not require complex network structures or parameter learning and is easy to understand and implement. However, a single fusion operation may not fully utilize the complex relationship between the features of two time points, resulting in information loss; it lacks universality and cannot adaptively adjust the fusion strategy according to the specific characteristics of the data.

[0049] The convolution operation can learn the complex relationship between the features of two time points, optimize the fusion effect by training network parameters, and adaptively adjust the fusion strategy according to the specific characteristics of the data, improve the generalization ability of the model, and retain the local detail information of the features. However, the convolution operation usually requires more computing resources, which may slow down the model training and reasoning. At the same time, the convolution enhancement method needs to learn a large number of network parameters, which may cause the model to overfit or unstable training.

[0050] Attention enhancement can form a focal focus, allowing the model to focus on important areas or elements in the features of two time points and ignore irrelevant information. The focus can also be adaptively adjusted according to the specific characteristics of the data to improve the generalization ability of the model. Attention weights can provide explainability for change detection and help understand how the model makes decisions. However, the attention mechanism usually requires additional network structures to calculate the attention weights, which may increase the complexity and computational complexity of the model. The attention mechanism may be sensitive to noise and outliers in the data, which may cause the model to perform poorly in some cases.

[0051] Therefore, in the field of change detection, the core challenge is how to effectively extract and fuse temporal information from bi-temporal features (i.e., data features at two different time points) to generate more powerful feature maps. These feature maps are crucial for capturing and quantifying the differences and changes between two time points.

[0052] Assume that the feature map R1 at time T1 and T2 is passed in b×c×h×w , R2 b×c×h×w , where b is the batchsize, c is the number of channels, and h and w are the feature map sizes. Multiply the obtained feature matrices to obtain the attention weight matrix; make a difference between the feature maps at time T1 and T2, and quantitatively analyze the changes in the surface by comparing the remote sensing images or feature maps at two different time points. The key to this step is to calculate the difference between the two feature maps, so as to identify the areas where the actual surface has changed significantly; multiply the attention weight matrix with the difference feature map, and obtain the final preliminary change feature map Tm through depth-separable convolution. The calculation process is as follows:

[0053] I s =Conv2D(Concat(Avg(T1),Max(T1)))

[0054] I C =Conv1D(Concat(Avg(T2),Max(T2)))

[0055] T m =ConvDeep([GELU(IS)·Signeid(Ic)]·(T1-T2))

[0056] At this time, the feature maps T1, T m And T2 for feature fusion, the calculation method is as follows:

[0057] S c =Concat(Avg(T1),Max(T1),Avg(T m ), Max(T m ), Avg(T2), Max(T2))

[0058] Where Sc represents the aggregated spatial features, Avg(·) and Max(·) represent the global average pooling and global maximum pooling across spatial dimensions, respectively.

[0059] The aggregated spatial features are passed to two 1D convolutions to determine the bi-temporal channel weights of the input bi-temporal features. The three channel weights can be formulated as:

[0060] W c1 , W cm , W c2 =Conv1(S c )Conv m (S c ), Conv2(S c )

[0061] Where W c1 , W cm , W c2 Denote channel weights respectively, Conv1(·), Convm(·), and Conv2(·) denote one-dimensional convolutions. Softmax is then used in the bi-temporal channel weights to make their sum equal to 1, which means comparing the three weights to determine the higher value among them, thereby determining the important part between the bi-temporal features in the channel dimension. The softmax method can be formulated as

[0062]

[0063] Where W′ c0 , W′ c1 and W′ c2 Represents the output dual time channel weights.

[0064] In the spatial branch, the dual temporal spatial weight W′ is determined in the same way s1 , W′ sm , W′ s2 By adding the dual-time spatial weights and the dual-time channel weights, the final attention weight matrix W′1,W′ is determined m ,W′2. Thus the final change difference feature map is calculated.

[0065] Thus, the important parts between the dual-temporal features in the spatial dimension are determined. The dual-temporal channel weights and the dual-temporal spatial weights are combined to obtain the dual-temporal weights to determine the important parts between the dual-temporal features. Finally, the dual-temporal weights are multiplied by the dual-temporal features and summarized to effectively fuse the dual-temporal features. The output can be formulated as:

[0066] Output=GELU(W′1)·Signeid(T1)+GELU(W′ m)·Signeid(T m )+GELU(W′2)·Signeid(T2)

[0067] Step 5, design the decoder:

[0068] There are two different decoders in the decoder part. Decoder a is responsible for building positioning, and it performs building positioning on the original feature maps T1 and T2 that are passed in. The feature map generated at time T2 is used as auxiliary positioning, and is only involved in training, and is not involved in calculations during verification and reasoning; decoder b is responsible for determining and reasoning the degree of damage to the building, and evaluating and grading the degree of change in the image. The building positioning mask obtained from the image data at time T1 is multiplied by the main elements of the building damage degree map to obtain an instantiated building damage degree positioning map.

[0069] Assume the output of the building positioning part is P b ∈R 2×H×W , where H and W represent the height and width of the output respectively, the building positioning result can be expressed as formula xxxx. B ∈{0,1} 1×H×W , 1 represents the building, and 0 represents the background. Assume that the output of damage assessment is P d ∈R C×H×W , where C represents the number of damage levels. To guide the building location, the final output P is expressed as. where · represents element-wise multiplication.

[0070] P B = argmax(P b )

[0071] P = arg max (P B ·P d )

[0072] Furthermore, in order to reduce the difference in the unchanged area in the dual-phase image data and highlight the characteristics of the changed area, this embodiment uses fast Fourier transform to unify the style of the two-phase images. Through frequency domain conversion, the high-frequency component reflects the edge and contour of the image, and the low-frequency component comprehensively measures the image intensity. By replacing the low-frequency part of the B image with the low-frequency part of the A image and inversely transforming it back to the spatial domain, the style of the two images can be unified, such as Figure 2 shown.

[0073] By conducting experiments on the neural network constructed in the method of the present invention on the xBD dataset, it is proved that the network proposed by the method of the present invention has better performance. In the method of the present invention, the xBD dataset is used as the basic dataset. This dataset is the largest and highest quality public building damage assessment dataset to date, covering 6 types and 19 natural disaster events in different regions, and the image data covers more than 45,000 square kilometers.

[0074] The present invention also provides a damage degree recognition system based on image data, which can realize the recognition method of the present invention. The system of the present invention can be realized by software and / or hardware, and can generally be integrated in electronic devices. The system includes a backbone network, an encoder, a channel switching module, a time series feature fusion module DFMA and a decoder.

[0075] Among them, the backbone network is the Unet network model.

[0076] A double-layer convolution is set in the encoder for feature extraction and channel deepening, and the feature map size is changed by maximum pooling or convolution to complete feature extraction; specifically, in the encoder part, the incoming feature map is first subjected to downsampling convolution with a step size of 2, and Haar wavelet transform downsampling is used as the residual connection part to retain more feature information. In the feature transfer of the model, in order to reduce parameter calculations, the incoming feature map is channel separated, and only half of the feature map is subjected to deep feature extraction. The CBAM attention module is introduced to accurately capture global and local key features through an adaptive attention weighting mechanism, and to fuse information from channel and spatial dimensions. Important feature information is also automatically highlighted through residual connections, while retaining the details of the original features. In order to enhance the network model's ability to extract vector boundaries of damaged buildings after disasters, the encoder includes a parameter-sharing network encoder for feature extraction.

[0077] The channel exchange module passes the bi-phase encoder features to the CE module for feature semi-exchange operation.

[0078] The temporal feature fusion module is used to fuse the bi-temporal features and send them to the decoder.

[0079] The decoder includes decoder a and decoder b; decoder a is responsible for building positioning, and performs building positioning on the original T1 and T2 feature maps passed in; the feature map generated at time T2 is used as auxiliary positioning, and is only involved in training, and is not involved in calculations during verification and inference; decoder b is responsible for determining and reasoning the degree of damage to the building, evaluating and grading the degree of change in the image, and multiplying the building positioning mask obtained through the image data at time T1 with the building damage degree map by the principal element, thereby obtaining an instantiated building damage degree positioning map.

[0080] The system of the present invention also includes a dual-phase style conversion module. In order to reduce the differences in the unchanged areas in the dual-phase image data and highlight the characteristics of the changed areas, the fast Fourier transform is used to unify the styles of the two-phase images. Through frequency domain conversion, the high-frequency components reflect the image edges and contours, and the low-frequency components comprehensively measure the image intensity. By replacing the low-frequency part of the B image with the low-frequency part of the A image and inversely transforming back to the spatial domain, the styles of the two images are unified.

[0081] The damage degree identification system based on image data provided by the embodiment of the present invention can execute the damage degree identification method based on image data provided by any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0082] The embodiment of the present application also provides an electronic device, Figure 3 The structure of the electronic device provided by the embodiment of the present invention is shown. For example, the electronic device 30 may include a processor 31, a memory 32 and a transmission device 33. The processor 31 is used to execute the damage degree identification method based on image data mentioned in the above embodiment, wherein the processor and the memory may be connected by a bus or other means, taking the bus connection as an example. The transmission device may be connected to the processor and the memory by wire or wirelessly. The memory, as a non-transitory computer-readable storage medium, may be used to store non-transitory software programs, non-transitory computer executable programs and modules, such as program instructions / modules corresponding to the damage degree identification method based on image data in the embodiment of the present application. The processor executes various functional applications and data processing of the processor by running the non-transitory software programs, instructions and modules stored in the memory, that is, the damage degree identification method based on image data in the above method embodiment is implemented. The memory may include a program storage area and a data storage area, wherein the program storage area may store an operating system and an application required by at least one function; the data storage area may store data created by the processor, etc. In addition, the memory may include a high-speed random access memory, and may also include a non-transitory memory, such as at least one disk storage device, a flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include a memory remotely arranged relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof. The one or more modules are stored in the memory, and when executed by the processor, the damage degree identification method based on image data in the embodiment is executed.

[0083] As another aspect, the present application also provides a computer-readable storage medium, which may be the computer-readable storage medium included in the device described in the above embodiment; or it may be a computer-readable storage medium that exists independently and is not assembled into the device. The computer-readable storage medium may be a tangible storage medium, such as a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a floppy disk, a hard disk, a removable storage disk, a CD-ROM, or any other form of storage medium known in the technical field. The computer-readable storage medium stores one or more programs, and the programs are used by one or more processors to execute the method for identifying the degree of damage based on image data described in the present application.

[0084] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for identifying the degree of damage based on image data, characterized in that: The steps include: Step 1: Establish a backbone network; Step 2: Design the encoder, specifically: perform feature extraction and channel deepening by setting up double-layer convolution, and change the size of the feature map by maximum pooling or convolution; Step 3: The bi-phase encoder features are passed into the CE module to perform feature semi-exchange operation; Step 4, integrating bi-temporal features; Step 5, design the encoder: set the decoder to include two different decoders, one decoder is responsible for building positioning; the other decoder is responsible for determining the degree of damage to the building and evaluating and grading the degree of change in the image.

2. The method according to claim 1, characterized in that The step 4 specifically includes: The difference between the two feature maps is calculated to identify the areas where the actual surface has changed significantly; the attention weight matrix is ​​multiplied with the difference feature map, and the final preliminary change feature map is obtained through deep separable convolution to perform feature fusion on all feature maps; the aggregated spatial features are passed to two one-dimensional convolutions to determine the dual-temporal channel weights of the input dual-temporal features, and then Softmax is used in the dual-temporal channel weights to make their sum equal to 1, and the three weights are compared to determine the higher value between them, thereby determining the important part between the dual-temporal features in the channel dimension; the dual-temporal space weights are determined in the same way in the spatial branch, and the final attention weight matrix is ​​determined by adding the dual-temporal space weights to the dual-temporal channel weights, thereby calculating the final change difference feature map and determining the important part between the dual-temporal features in the spatial dimension; the dual-temporal channel weights and the dual-temporal space weights are combined to obtain the dual-temporal weights to determine the important part between the dual-temporal features; the dual-temporal weights are multiplied by the dual-temporal features to fuse the dual-temporal features.

3. The method according to claim 2, characterized in that It also includes the use of fast Fourier transform to unify the styles of two-phase images.

4. The method according to any one of claims 1 to 3, characterized in that In the step 2, in the encoder part, the incoming feature map is first convolved with a downsampling step of 2, and Haar wavelet transform downsampling is used as the residual connection part to retain more feature information; in the feature transfer of the model, the channel separation is performed on the incoming feature map, and only half of the feature map is subjected to deep feature extraction.

5. The method according to claim 4, characterized in that In step 2, the CBAM attention module is introduced to capture global and local key features and fuse information of channel and spatial dimensions through an adaptive attention weighting mechanism; important feature information is automatically highlighted through residual connections while retaining the details of the original features.

6. The method according to claim 4 or 5, characterized in that In the encoder part, a parameter-sharing network encoder is constructed for feature extraction.

7. A damage degree identification system based on image data, characterized in that: It includes a backbone network, an encoder, a channel switching module, a timing feature fusion module DFMA and a decoder. Among them, the backbone network is the Unet network model; a double-layer convolution is set in the encoder for feature extraction and channel deepening, and the feature map size is changed through maximum pooling or convolution to complete feature extraction; the channel exchange module passes the dual-phase encoder features to the CE module for feature semi-exchange operation; the temporal feature fusion module is used to fuse the dual-temporal features and send them to the decoder; the decoder includes decoder a and decoder b; decoder a is responsible for building positioning, and locates the building on the original T1 and T2 feature maps passed in; the feature map generated at time T2 is used as auxiliary positioning, and is only involved in training, and is not involved in calculation during verification and reasoning; decoder b is responsible for judging and reasoning the degree of building damage, evaluating and grading the degree of change in the image, and multiplying the building positioning mask obtained by the image data at time T1 with the building damage degree map by the principal element, thereby obtaining an instantiated building damage degree positioning map.

8. The system according to claim 7, characterized in that It also includes a dual-phase style transfer module, which uses fast Fourier transform to unify the styles of two-phase images.

9. An electronic device, characterized in that: The electronic device includes a processor and a memory for storing executable instructions of the processor; the processor is used to read the executable instructions from the memory and execute the instructions to implement the method for identifying the degree of damage based on image data as described in any one of claims 1 to 6 above.

10. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and the computer program is used to execute the damage degree identification method based on image data as described in any one of claims 1 to 6 above.

Images