Moon crest extraction method and system based on double-branch deep learning model, storage medium and electronic equipment

Abstract

The invention discloses a moon wrinkle ridge extraction method and system based on a double-branch deep learning model, a storage medium and electronic equipment. The method comprises the following steps of data preprocessing, DEM data processing and slope direction data generation; extracting features by using a double-branch feature encoder to obtain DEM features and slope features; based on an attention mechanism, fusing the double-branch deep features to obtain deep semantic information in the fused features; performing multi-scale selection in a self-adaptive manner through spatial cavity convolution pooling; fusing shallow layer features, performing jump connection, and providing detail information for up-sampling; and carrying out up-sampling and image size recovery to obtain an extraction result. Through the method, the form and the position of the creases on the lunar surface can be accurately identified in a large range, the labor amount of manual extraction work can be effectively reduced, a basis is provided for moon stress structure research, and subsequent research on the stress state and evolution history of the moon is facilitated.

Description

Technical Field

[0001] The present application relates to the field of deep space exploration and remote sensing, and in particular to a lunar ridge extraction method, system, storage medium and electronic device based on a dual-branch deep learning model. Background Art

[0002] The lunar geological structure has important indicative significance for the study of lunar evolution. According to its geometric characteristics, the lunar geological structure can be roughly divided into linear structure and annular structure. Wrinkle ridges are one of the most common linear structures on the lunar surface. They are related to the stress state inside the moon. Studying wrinkle ridges is of great significance for understanding the lunar stress field and evolution history. To better understand wrinkle ridges, it is first necessary to identify wrinkle ridges and understand their distribution. However, due to the complex morphology of wrinkle ridges and the unclear feature differences, the detection of wrinkle ridges is challenging.

[0003] In order to understand the distribution of lunar wrinkle ridges, researchers have proposed a large number of lunar wrinkle ridge extraction algorithms, which can be divided into manual visual interpretation methods and traditional image processing methods, and all of them can complete the task of wrinkle ridge extraction under certain conditions. However, most of the existing methods have certain limitations. Although the manual visual interpretation method can extract lunar wrinkle ridges more accurately, it is labor-intensive and has low extraction efficiency, making it difficult to perform large-scale lunar wrinkle ridge extraction tasks. Most traditional image processing methods are based on threshold selection, mainly using terrain curvature, phase information, morphological operations and other technologies. Such traditional image processing methods have high extraction efficiency, simple calculations and easy implementation, but such methods can only identify ridge lines that represent the direction of lunar wrinkle ridges, and cannot identify the shape and edges of wrinkle ridges. There is a lack of detailed description of wrinkle ridges, and the extraction effect depends largely on the selection of thresholds. Existing methods are difficult to achieve efficient and accurate lunar wrinkle ridge recognition at the same time.

[0004] In summary, in order to obtain efficient and accurate lunar ridge extraction, it is of great significance to use new deep learning methods and study deep learning models for extracting lunar ridges. Summary of the invention

[0005] The purpose of this application is to provide a lunar ridge extraction method and system based on a dual-branch deep learning model to solve or alleviate the problems existing in the above-mentioned prior art.

[0006] In order to achieve the above objectives, this application provides the following technical solutions:

[0007] The present application provides a lunar wrinkle ridge extraction method based on a dual-branch deep learning model, including: step S101, data preprocessing, processing DEM data and generating aspect data; step S102, using a dual-branch feature encoder to extract features to obtain DEM features and aspect features; step S103, based on an attention mechanism, fusing the dual-branch deep features to obtain deep semantic information in the fused features; step S104, adaptively performing optimal multi-scale selection through spatial hole convolution pooling; step S105, fusing shallow features, performing jump connections, and providing detail information for upsampling; step S106, upsampling and restoring the image size to obtain extraction results, and performing accuracy evaluation.

[0008] Preferably, in step S101, the data is preprocessed to process the DEM data and generate the aspect data, specifically:

[0009] The DEM data is normalized by linear function to speed up the convergence of the deep learning model, and the slope data is calculated based on the DEM data.

[0010] The DEM data uses a linear function to linearly transform the original data to the range of [0,1]. The normalization formula is as follows:

[0011]

[0012] Among them, X represents the pixel value before normalization, X norm Represents the normalized pixel value, X min Represents the minimum value of the image pixel, X max Indicates the maximum value of image pixels.

[0013] Based on the normalized DEM data, the slope data is calculated. The calculation uses a 3×3 sliding window, assuming e i is the i-th element in the sliding window, then the slope aspect is calculated as follows:

[0014]

[0015] Among them, Slope x Indicates the rate of change of slope in the x direction, Slope y It represents the rate of change of the slope in the y direction, and Aspect represents the calculated aspect value.

[0016] Preferably, in step S102, a dual-branch feature encoder is used to extract features to obtain DEM features and slope features, specifically:

[0017] Using DEM data and aspect data as inputs to the deep learning model, two Resnet50s pre-trained on the ImageNet dataset are used as backbone networks to extract features from the two input data, obtaining feature maps of DEM data and aspect data; the dual-branch feature encoder consists of basic residual blocks, a basic residual block includes a convolutional layer, batch normalization, activation function, and short-circuit connection. Use residual blocks to learn the residual mapping between input and output.

[0018] A dual-branch feature encoder is used to extract two kinds of multi-faceted and multi-level features, which provide semantic information and detail information for target localization and image size recovery.

[0019] Preferably, in step S103, based on the attention mechanism, the dual-branch deep features are fused, specifically:

[0020] For the two feature maps extracted by the dual-branch feature encoder, namely the DEM data feature map and the aspect data feature map, a maximization operation is first performed to combine the most significant features of the two feature maps:

[0021] X=Max(X 1 ,X 2 )

[0022] Among them, X 1 ,X 2 They represent the input DEM data and aspect data respectively. X is the result of maximizing the two data.

[0023] Convolution, batch normalization and activation function are used to map the features, and the local channel features L(X)∈R C ×H×W The following formula can be used for calculation:

[0024] L(X)=γ(β(Conv(X)))

[0025] Among them, Conv represents convolution, β represents batch normalization, and γ represents the ReLU activation function.

[0026] The output of the feature fusion module is converted into a probability value through the Sigmoid function, which is then used for binary classification problems. The probability value output by the feature fusion module represents the possibility that the sample belongs to a certain class. The obtained probability is the attention weight.

[0027] The attention weights obtained by the attention mechanism are used to weight the original feature maps, focusing attention on the content that is significantly expressed in the feature maps of two different data, and the two feature maps after the attention weighting are spliced ​​and fused in the channel dimension. The calculation formula can be expressed as:

[0028]

[0029] Among them, concat represents the concatenation operation. represents multiplication, σ represents Sigmoid function, and F(X) represents fusion feature.

[0030] Preferably, in step S105, shallow features are fused and skip connections are performed to provide detailed information for upsampling, specifically:

[0031] The dual-branch feature encoder extracts the shallow feature maps of the two data, fuses them using the addition operation, and then splices them in the channel dimension through the jump connection and the decoded feature map. The calculation formula can be expressed as follows:

[0032] F(X) = concat(X 1 +X 2 ,D(X))

[0033] X 1 ,X 2 Represents the shallow features of two different data, D(X) represents the decoding feature map, and F(X) represents the fusion of shallow features and decoding features.

[0034] The jump connection part is used to jump-connect the extracted feature map corresponding to the feature extraction module of the feature encoding part and the decoded feature map corresponding to the feature decoding module of the upsampling part to obtain a jump feature map. The jump connection is specifically: splicing the extracted feature map and the decoded feature map in the channel dimension; passing the shallow features to the upsampling process through the jump connection.

[0035] Preferably, in step S106, upsampling and restoring the image size, obtaining the extraction result, and performing accuracy evaluation are performed, specifically:

[0036] For low-resolution feature maps, a linear interpolation method is used to restore the low-resolution feature maps to high-resolution prediction results to obtain the semantically segmented image.

[0037] The accuracy of the dual-branch deep learning model is evaluated based on the unlabeled images in the test set and the labeled images in the test set; wherein the unlabeled images in the test set are obtained by directly intercepting the preprocessed DEM images and aspect images; and the labeled images in the test set are obtained by labeling the unlabeled images in the test set according to preset classification labels.

[0038] The accuracy evaluation indicators are Precision, Recall, F1, and IoU, which evaluate the model's wrinkle ridge detection ability. The calculation formula is as follows:

[0039]

[0040] Among them, TP, FP, FN and TN represent the number of true positive, false positive, false negative and true negative pixels in the prediction results, respectively.

[0041] The embodiment of the present application also provides a lunar ridge extraction system based on a dual-branch deep learning model, comprising:

[0042] A data processing unit is configured to segment the pre-processed DEM image and calculate and derive a slope aspect image;

[0043] A feature encoding unit is configured to use a dual-branch feature encoder to extract features from the input data to obtain features of the two types of data;

[0044] The feature fusion unit is configured to obtain the attention weight of the salient feature map based on the attention mechanism, perform attention weighting, and fuse the two features in a cascade manner;

[0045] The feature decoding unit is configured to perform semantic segmentation on the input data based on the fused features, restore the image size through upsampling, and obtain the extraction result after semantic segmentation.

[0046] An embodiment of the present application also provides a computer-readable storage medium on which a computer program is stored, wherein the program is a lunar ridge extraction method based on a dual-branch deep learning model as described above.

[0047] An embodiment of the present application also provides an electronic device, comprising: a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the lunar ridge extraction system based on the dual-branch deep learning model as described above is implemented.

[0048] Beneficial effects:

[0049] In the lunar wrinkle ridge extraction method based on the dual-branch deep learning model provided in the present application, first, the DEM data and aspect data are processed, and the two types of data are simultaneously input into the deep learning model to obtain a richer feature expression. Then, a dual-branch feature encoder is used to extract features from the two types of data, and the two feature maps are fused through a feature fusion module based on an attention mechanism. Finally, the feature map is upsampled, the image size is restored, and the final semantic segmentation result is obtained. The method of the present application simultaneously utilizes the main features in the DEM data and the edge features in the aspect data. By fusing the main features and the edge features, the feature expression ability of the model is improved, and the morphology and edges of the lunar wrinkle ridges can be accurately located and described, achieving efficient and accurate lunar wrinkle ridge extraction effects, which is conducive to subsequent research on the stress state and evolution history of the moon. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The drawings constituting part of the present application are used to provide a further understanding of the present application. The exemplary embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. Among them:

[0051] Figure 1 A schematic diagram of a process according to the present invention;

[0052] Figure 2 A structural diagram of a dual-branch deep learning model according to the present invention;

[0053] Figure 3 A display image for data preprocessing provided according to some embodiments of the present application;

[0054] Figure 4 A result diagram of extracting lunar ridges by semantically segmenting preprocessed data based on the present application according to some embodiments of the present application;

[0055] Figure 5 It is a schematic diagram of the structure according to the present invention. DETAILED DESCRIPTION

[0056] The present application will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments. Each example is provided by way of explanation of the present application and does not limit the present application. In fact, it will be clear to those skilled in the art that modifications and variations may be made in the present application without departing from the scope or spirit of the present application. For example, a feature shown or described as a part of an embodiment may be used in another embodiment to produce yet another embodiment. Therefore, it is desired that the present application includes such modifications and variations within the scope of the appended claims and their equivalents.

[0057] Exemplary Methods

[0058] like Figure 1 As shown, the lunar ridge extraction method based on the dual-branch deep learning model includes:

[0059] Step S101, data preprocessing, processing DEM data and generating aspect data;

[0060] Specifically, the DEM data is normalized by a linear function to accelerate the convergence of the deep learning model, and the slope aspect data is calculated based on the DEM data.

[0061] In order to obtain an input image that meets the requirements of the dual-branch deep learning model, the acquired DEM image needs to be preprocessed and the slope image needs to be calculated based on the DEM image.

[0062] In the embodiments of the present application, an existing product generated by merging the data provided by LRO LOLA and SELENE TC is taken as an example for description, and the resolution of the product is 59m / pixel.

[0063] In the embodiment of the present application, the DEM product provides elevation information in the form of floating-point data, and uses a linear function normalization to convert the original data to the range of [0,1]. The normalization formula is as follows:

[0064]

[0065] Among them, X represents the pixel value before normalization, X norm Represents the normalized pixel value, X min Represents the minimum value of the image pixel, X max Indicates the maximum value of image pixels.

[0066] Based on the normalized DEM data, the slope data is calculated. The calculation uses a 3×3 sliding window, assuming e i is the i-th element in the sliding window, then the slope aspect is calculated as follows:

[0067]

[0068]

[0069] Among them, Slope x Indicates the rate of change of slope in the x direction, Slope y It represents the rate of change of the slope in the y direction, and Aspect represents the calculated aspect value.

[0070] Aspect identifies the downslope direction of the greatest rate of change in value from each cell to its neighbors. Aspect can be thought of as the direction of the slope. The value of each cell in the output raster indicates the compass direction of the surface at that cell location. It is measured clockwise, with angles ranging from 0 (due north) to 360 (still due north), a full circle. Flat areas with no downslope direction are assigned a value of -1.

[0071] Figure 3 To show the result of data preprocessing, the value of each pixel in the aspect dataset can indicate the slope direction of the pixel. The input data of the embodiment of the present application is specified as an image size of 512×512.

[0072] This step processes the DEM data and aspect data through traditional image processing operations to facilitate the effective use of the data by subsequent models.

[0073] Step S102: extract features using a dual-branch feature encoder to obtain DEM features and slope features;

[0074] Specifically, DEM data and aspect data are used as inputs of the deep learning model, and two Resnet50s pre-trained on the ImageNet dataset are used as backbone networks to extract features from the two input data to obtain feature maps of DEM data and aspect data; the dual-branch feature encoder consists of basic residual blocks, and a basic residual block contains convolutional layers, batch normalization, activation functions, and short-circuit connections. The residual block is used to learn the residual mapping between input and output.

[0075] A dual-branch feature encoder is used to extract two kinds of multi-faceted and multi-level features, which provide semantic information and detail information for target localization and image size recovery.

[0076] In the embodiment of the present application, the dual-branch feature encoder includes a 4-layer structure, and each layer includes residual blocks of {3, 4, 6, 3}. The feature map size of the shallow feature is 128×128, which is generated by the first layer. The feature map size of the deep feature is 32×32, which is generated by the fourth layer.

[0077] like Figure 2 As shown, it is a structural diagram of the deep learning model for lunar ridge lifting based on dual branches provided in an embodiment of the present application.

[0078] Step S103: Based on the attention mechanism, the deep features of the two branches are fused;

[0079] Specifically, for the two feature maps extracted by the dual-branch feature encoder, namely the DEM data feature map and the aspect data feature map, a maximization operation is first performed to combine the most significant features of the two feature maps:

[0080] X=Max(X 1 ,X 2 )

[0081] Among them, X 1 ,X 2 They represent the input DEM data and aspect data respectively. X is the result of maximizing the two data.

[0082] Take two 4×4 matrices as an example to illustrate the process of maximization operation. 1 and S 2 as follows:

[0083]

[0084] Assuming c>a, d>b, g>e, h>f, the maximization operation obtains the maximum value at the same position in the two matrices. The matrix after the maximization operation is:

[0085]

[0086] Convolution, batch normalization and activation function are used to map the features, and the local channel features L(X)∈R C ×H×W The following formula can be used for calculation:

[0087] L(X)=γ(β(Conv(X)))

[0088] Among them, Conv represents convolution, β represents batch normalization, and γ represents the ReLU activation function.

[0089] The output of the feature fusion module is converted into a probability value through the Sigmoid function, which is then used for binary classification problems. The probability value output by the feature fusion module represents the possibility that the sample belongs to a certain class. The obtained probability is the attention weight.

[0090] The attention weights obtained by the attention mechanism are used to weight the original feature maps, focusing attention on the content that is significantly expressed in the feature maps of two different data, and the two feature maps after the attention weighting are spliced ​​and fused in the channel dimension. The calculation formula can be expressed as:

[0091]

[0092] Among them, concat represents the concatenation operation. represents multiplication, σ represents Sigmoid function, and F(X) represents fusion feature.

[0093] Step S104, adaptively performing multi-scale selection through spatial atrous convolution pooling pyramid;

[0094] Specifically, the dilated convolutional pooling pyramid consists of a 1×1 ordinary convolution, three 3×3 spatial dilated convolutions (with dilation rates of 6, 12, and 18, respectively), and a global average pooling. The difference between the dilated convolution layer and the ordinary convolution is the dilation rate. Different dilation rates can be used to obtain receptive fields of different sizes and extract multi-scale information. The dilated convolutional pooling pyramid combines feature maps of multiple receptive fields and adaptively selects the best multi-scale.

[0095] In the embodiment of the present application, the configurations of the three spatial dilated convolutions are respectively:

[0096] Kernel size=3×3, stride=1, padding=6, rate=6;

[0097] kernel size=3×3, stride=1, padding=12, rate=12;

[0098] kernel size=3×3, stride=1, padding=18, rate=18;

[0099] This setting ensures that the size of the feature map remains consistent while the convolution receptive field expands.

[0100] Step S105: Fusing shallow features and performing skip connections to provide detail information for upsampling;

[0101] Specifically, the shallow feature map is extracted by the dual-branch feature encoder, fused using the addition operation, and then spliced ​​in the channel dimension through the jump connection and the decoded feature map. The calculation formula can be expressed as follows:

[0102] F(X) = concat(X 1 +X 2 ,D(X))

[0103] X 1 ,X 2 Represents the shallow features of two different data, D(X) represents the decoding feature map, and F(X) represents the fusion of shallow features and decoding features.

[0104] The jump connection part is used to jump-connect the extracted feature map corresponding to the feature extraction module of the feature encoding part and the decoded feature map corresponding to the feature decoding module of the upsampling part to obtain a jump feature map. The jump connection is specifically: splicing the extracted feature map and the decoded feature map in the channel dimension; passing the shallow features to the upsampling process through the jump connection.

[0105] Step S106, upsampling and restoring the image size, obtaining the extraction result, and performing accuracy evaluation;

[0106] Specifically, the accuracy of the dual-branch deep learning model is evaluated based on the unlabeled images of the test set and the labeled images of the test set; wherein the unlabeled images of the test set are obtained by directly intercepting the preprocessed DEM images and slope aspect images; and the labeled images of the test set are obtained by classifying the unlabeled images of the test set according to preset classification labels.

[0107] The accuracy evaluation indicators are precision, recall, F1, and IoU, which evaluate the model's wrinkle ridge detection ability. The calculation formula is as follows:

[0108]

[0109] Among them, TP, FP, FN and TN represent the number of true positive, false positive, false negative and true negative pixels in the prediction results, respectively.

[0110] Figure 4 The extraction results of the embodiment of the present application based on the dual-branch deep learning model are shown, and the results are compared with the manual visual interpretation, which is consistent with the actual situation. The extraction accuracy is high and the morphological description is more accurate. Compared with the existing traditional methods, the extraction accuracy of this method is significantly improved.

[0111] Exemplary Systems

[0112] Figure 5 The present invention is a schematic diagram of the structure of a lunar wrinkle ridge extraction system based on a dual-branch deep learning model provided according to an embodiment of the present application. It includes: a data processing unit, configured to segment the preprocessed DEM image and calculate and derive the slope image; a feature encoding unit, configured to use a dual-branch feature encoder to extract features from the input data and obtain the features of the two data; a feature fusion unit, configured to obtain the attention weight of the significant feature map based on the attention mechanism, perform attention weighting, and fuse the two features in a cascade manner; a feature decoding unit, configured to perform semantic segmentation on the input data based on the fused features, restore the image size after upsampling, and obtain the extraction result after semantic segmentation.

[0113] The lunar ridge extraction system based on the dual-branch deep learning model provided in the embodiment of the present application can implement any of the above-mentioned lunar ridge extraction steps and processes based on the dual-branch deep learning model, and achieve the same technical effect, which will not be repeated here one by one.

[0114] Exemplary Devices

[0115] The present application provides an electronic device, including a storage medium and a processor, the processor is suitable for executing various programs; the memory is used to store multiple programs; it is characterized in that when the memory executes the program on the processor, the lunar ridge extraction method based on the dual-branch deep learning model is implemented.

[0116] Since the lunar ridge extraction steps based on the dual-branch deep learning model have been introduced in detail in the specific implementation method example, they will not be repeated here.

[0117] The processor includes a central processing unit (CPU), a network processor (NP), etc. It can also be a digital signal processor, an application-specific integrated circuit, a readily available programmable gate array or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component. The disclosed methods, steps and logic block diagrams in the embodiments of the present application can be implemented or executed. The general processor can be a microprocessor or the processor can also be any conventional processor, etc.

[0118] The processor can be specifically configured as follows: first preprocess the data, process the DEM data and generate the aspect data; use the dual-branch feature encoder to extract features to obtain DEM features and aspect features; based on the attention mechanism, fuse the dual-branch deep features to obtain the deep semantic information in the fused features; adaptively perform multi-scale selection through spatial hole convolution pooling; fuse shallow features and perform jump connections to provide detail information for upsampling; upsample and restore the image size to obtain the extraction result.

[0119] It should be pointed out that, according to the needs of implementation, the various components / steps described in the embodiments of the present application can be split into more components / steps, or two or more components / steps or partial operations of components / steps can be combined into new components / steps to achieve the purpose of the embodiments of the present application.

[0120] The above-mentioned method according to the embodiment of the present application can be implemented in hardware, firmware, or implemented as software or computer code that can be stored in a recording medium (such as a CD ROM, RAM, floppy disk, hard disk or magneto-optical disk), or implemented as computer code originally stored in a remote recording medium or a non-temporary machine storage medium downloaded through a network and to be stored in a local recording medium, so that the method described herein can be stored in such software processing on a recording medium using a general-purpose computer, a dedicated processor or programmable or dedicated hardware (such as an ASIC or FPGA). It can be understood that a computer, a processor, a microprocessor controller or programmable hardware includes a storage component (e.g., RAM, ROM, flash memory, etc.) that can store or receive software or computer code. When the software or computer code is accessed and executed by a computer, a processor or hardware, the lunar wrinkle ridge extraction method based on a dual-branch deep learning model described herein is implemented. In addition, when a general-purpose computer accesses the code for implementing the method shown herein, the execution of the code converts the general-purpose computer into a dedicated computer for executing the method shown herein.

[0121] Those of ordinary skill in the art will appreciate that the units and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and constraints involved in the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the embodiments of the present application.

[0122] It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device and system embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.

[0123] The device and system embodiments described above are merely illustrative, wherein the units not shown as separate may or may not be physically separated, and the units not shown as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Those of ordinary skill in the art may understand and implement it without creative work.

[0124] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A lunar ridge extraction method based on a dual-branch deep learning model, characterized in that: include: Step S101, data preprocessing: normalizing the DEM data, and then calculating and generating the slope aspect data according to the DEM data; Step S102: extract features using a dual-branch feature encoder to obtain DEM features and slope features, i.e., dual-branch deep features, which are used to provide semantic information and detail information for target positioning and image size recovery; Step S103: Based on the attention mechanism, the deep features of the two branches are fused to obtain the deep semantic information in the fused features; Step S104: Adaptively select the best multi-scale by using a spatial atrous convolution pooling pyramid, thereby ensuring that the size of the feature map remains consistent while the convolution receptive field is expanded; Step S105: Fusing shallow features and performing skip connections to provide detail information for upsampling; Step S106: perform upsampling and restore the image size to obtain the extraction result and perform accuracy evaluation.

2. The lunar ridge extraction method based on a dual-branch deep learning model according to claim 1 is characterized in that: The step S101 is specifically: performing linear function normalization on the DEM data to accelerate the convergence speed of the deep learning model, and calculating and generating slope aspect data according to the DEM data; The DEM data uses a linear function to linearly transform the original data to the range of [0,1]. The normalization formula is as follows: Among them, X represents the pixel value before normalization, X norm Represents the normalized pixel value, X min Represents the minimum value of the image pixel, X max Indicates the maximum value of image pixels; According to the normalized DEM data, the slope data is calculated; the calculation uses a 3×3 sliding window, assuming e i is the i-th element in the sliding window, then the slope aspect is calculated as follows: Among them, Slope x Indicates the rate of change of slope in the x direction, Slope y It represents the rate of change of the slope in the y direction, and Aspect represents the calculated aspect value.

3. The lunar ridge extraction method based on a dual-branch deep learning model according to claim 1 is characterized in that: The step S102 is specifically as follows: Using DEM data and aspect data as inputs of a deep learning model, using two Resnet50s pre-trained on the ImageNet dataset as backbone networks to extract features of the two input data, and obtaining feature maps of the DEM data and feature maps of the aspect data; the dual-branch feature encoder is composed of basic residual blocks, and a basic residual block includes a convolution layer, batch normalization, an activation function, and a short-circuit connection; A residual block is used to learn the residual mapping between input and output.

4. The lunar ridge extraction method based on a dual-branch deep learning model according to claim 1 is characterized in that: The step S103 is specifically as follows: For the two feature maps extracted by the dual-branch feature encoder, namely the DEM data feature map and the aspect data feature map, a maximization operation is first performed to combine the most significant features of the two feature maps: X=Max(X1,X2) Among them, X1 and X2 represent the input DEM data and slope aspect data respectively; X is the result of maximizing the two data; Convolution, batch normalization and activation function are used to map the features, and the local channel features L(X)∈R C×H×W The following formula can be used for calculation: L(X)=γ(β(Conv(X))) Among them, Conv represents convolution, β represents batch normalization, and γ represents the ReLU activation function. The output of the feature fusion module is converted into a probability value through the Sigmoid function, which is then used for binary classification problems. The probability value output by the feature fusion module represents the possibility that the sample belongs to a certain category, and the obtained probability is the attention weight. The original feature map is weighted by the attention weight obtained by the attention mechanism, focusing attention on the content that is significantly expressed in the feature maps of two different data, and the two feature maps after the attention weighting are spliced ​​and fused in the channel dimension; the calculation formula can be expressed as: Among them, concat represents the concatenation operation. represents multiplication, σ represents Sigmoid function, and F(X) represents fusion feature.

5. The lunar ridge extraction method based on a dual-branch deep learning model according to claim 1 is characterized in that: The atrous convolution pooling pyramid in step S104 is composed of a 1×1 ordinary convolution, three 3×3 spatial atrous convolutions and a global average pooling. The difference between the atrous convolution layer and the ordinary convolution is the expansion rate. Different sizes of receptive fields can be obtained through different expansion rates to extract multi-scale information.

6. The lunar ridge extraction method based on a dual-branch deep learning model according to claim 1 is characterized in that: The step S105 is specifically as follows: The dual-branch feature encoder extracts the shallow feature maps of the two data, fuses them using the addition operation, and then splices them in the channel dimension through the jump connection and the decoded feature map. The calculation formula can be expressed as follows: F(X)=concat(X1+X2,D(X)) X1, X2 represent the shallow features of two different data, D(X) represents the decoded feature map, and F(X) represents the fusion of shallow features and decoded features; The jump connection is used to jump-connect the extracted feature map corresponding to the feature extraction module of the feature encoding part and the decoded feature map corresponding to the feature decoding module of the upsampling part to obtain a jump feature map; the jump connection is specifically: splicing the extracted feature map and the decoded feature map in the channel dimension; through the jump connection, the shallow feature is passed to the upsampling process.

7. The lunar ridge extraction method based on a dual-branch deep learning model according to claim 1 is characterized in that: The step S106 is specifically as follows: For low-resolution feature maps, a linear interpolation method is used to restore the low-resolution feature maps to high-resolution prediction results to obtain the semantically segmented image; The accuracy of the dual-branch deep learning model is evaluated based on the unlabeled images in the test set and the labeled images in the test set; wherein the unlabeled images in the test set are obtained by directly intercepting the preprocessed DEM images and aspect images; and the labeled images in the test set are obtained by labeling the unlabeled images in the test set according to preset classification labels. The accuracy evaluation indicators are precision, recall, F1, and IoU, which evaluate the model's wrinkle ridge detection ability; the calculation formula is as follows: Among them, TP, FP, FN and TN represent the number of true positive, false positive, false negative and true negative pixels in the prediction results, respectively.

8. A lunar ridge extraction system based on a dual-branch deep learning model, characterized in that: include: The data processing unit is configured to: segment the pre-processed DEM image and calculate and derive the slope aspect image; Obtaining an unlabeled image of a target training set and an labeled image of a target training set, wherein the labeled image of the target training set is obtained by labeling the unlabeled image of the target training set according to a preset classification label; The feature encoding unit is configured to: use a dual-branch feature encoder to extract features from the input data to obtain features of the two types of data; The feature fusion unit is configured as follows: based on the attention mechanism, the attention weight of the salient feature map is obtained, attention weighting is performed, and two features are fused in a cascade manner; The feature decoding unit is configured to: perform semantic segmentation on the input data based on the fused features, restore the image size through upsampling, and obtain the extraction result after semantic segmentation.

9. A storage medium storing a plurality of programs, characterized in that: The program application is loaded and executed by a processor to implement the lunar ridge extraction method based on a dual-branch deep learning model as described in any one of claims 1-7.

10. An electronic device, comprising a storage medium and a processor; the processor is adapted to execute various programs; and the memory is adapted to store a plurality of programs; characterized in that: When the memory executes the program on the processor, it implements the lunar ridge extraction method based on a dual-branch deep learning model as described in any one of claims 1-7.

Images