A Method for Identifying Lunar Impact Craters Based on Data Fusion and Transfer Learning

Abstract

The present invention provides a method for lunar crater identification based on data fusion and transfer learning, belonging to the technical field of image recognition, and comprising the following steps: S1, construction of a source domain data set; S2, construction of a network model; S3, construction of multi-source fusion data; S4, construction of a target domain data set: randomly cropping the fused image, annotating the craters, drawing the true edge information of the craters, and dividing into a training set and a test set; S5, transfer learning model; S6, result verification. After fusing lunar remote sensing images and elevation data, the method of the present invention further realizes the detection and identification of craters in lunar remote sensing images. The established network consists of two crater detection channels and one crater identification channel. The present invention proposes a new method for fusing elevation data and image data, and adopts a transfer learning strategy for the newly fused data, which can achieve accurate identification of craters only with small sample data annotation, and obtains a good identification effect.

Description

Technical Field

[0001] The present invention belongs to the technical field of image recognition, and relates to a computer deep learning method, in particular to a method for identifying lunar impact craters based on data fusion and transfer learning. Background Art

[0002] The moon is the celestial body closest to the earth and also the satellite with the largest mass relative to the earth. There are a large number of impact craters on the lunar surface. Impact craters are depressions formed by meteorites hitting celestial bodies and, together with lunar seas, highlands, etc., constitute typical lunar surface features. At present, the Chang'e project has completed the lunar surface survey mission, brought back a large number of lunar surface pictures, and drawn a three-dimensional map of the entire moon, which helps further research on the moon in the future. Through the study of impact craters, not only can we know the relative geological age and surface characteristics of the celestial body surface, but also we can provide data analysis for the positioning and obstacle avoidance of spacecraft navigation in future space exploration by mapping the celestial body surface. Therefore, the extraction and identification of impact craters have important significance in the field of space exploration.

[0003] Currently, the most important method for identifying impact craters is manual identification. The advantage of the manual identification method is that it can accurately identify various large impact craters. However, the manual identification method is not only time-consuming and laborious, but also the limitation of "visibility" will lead to incomplete identification results, and the identification effect for small impact craters is not very ideal. Therefore, how to quickly and accurately identify lunar impact craters remains a difficult problem and focus in the field of lunar exploration.

[0004] In order to improve the recognition efficiency of celestial impact craters, a preliminary attempt was made to use a computer to identify impact craters, and various automatic impact crater recognition algorithms were designed. Most of these algorithms are based on computer learning methods. Impact crater recognition is divided into two categories, one is unsupervised learning, and the other is supervised learning. In the unsupervised learning method, mainly using the geometric features that impact craters are mostly circular or elliptical, methods based on terrain analysis and mathematical morphology or methods based on local gray levels of image regions are used to match the impact craters in the image. In the supervised learning method, the already labeled impact craters are usually divided into three types, training samples, validation samples, and test samples, and then methods such as neural networks, support vector machines, ensemble learning, transfer learning, or continuous scalable template matching algorithms are used to improve the recognition accuracy rate during the repeated learning process, and then the recognition effect of the supervised learning method is demonstrated in the test sample set. The unsupervised learning method has a better recognition effect in specific images, but the illumination and shooting angle have a great impact on the recognition effect, and these reasons make it difficult for this type of method to be put into formal application. In supervised learning, by inputting an object (lunar surface image) to obtain an expected output value (impact crater edge labeled image), and then comparing the output value with the real output value (impact crater real labeled image), continuously adjusting the classifier parameters to minimize the recognition error. Currently, in supervised learning, the impact crater images manually labeled are usually used as the real output. However, the manual labeling is not complete for smaller impact craters, which will have some impacts on the actual application.

[0005] In recent years, deep learning has achieved great success in the field of computer vision. Different from traditional machine learning methods, deep learning will automatically find out the important features required to solve problems, while traditional machine learning methods may need to manually define some features. As the amount of data increases, the performance of deep learning is better than that of traditional machine learning. Algorithms based on convolutional neural networks perform very well in solving problems such as object detection, segmentation, and classification. A convolutional neural network is a multi-layer neural network that reduces the data dimension and gradually extracts data features through operations such as convolution and pooling. Finally, the classification task is completed through the weights of the trained convolutional neural network. In the face recognition task, the recognition accuracy rate of deep learning has exceeded that of human eyes. At the same time, deep learning has also achieved success in tasks such as autonomous driving and satellite image recognition. In addition to the field of computer vision, deep learning has also achieved great success in fields such as speech recognition and natural language processing. Summary of the Invention

[0006] The object of the present invention is to address the above problems existing in the prior art, and propose a lunar impact crater recognition method based on data fusion and transfer learning that combines two types of data to construct a new fused dataset, which can not only accurately judge the existence of impact craters, but also accurately draw the real contour of impact craters, thereby greatly improving the recognition accuracy and precision.

[0007] The object of the present invention can be achieved by the following technical solutions: A method for identifying lunar craters based on data fusion and transfer learning, comprising the following steps:

[0008] S1. Construction of source domain dataset:

[0009] Randomly crop lunar image data, mark and draw crater pictures according to the diameter and coordinate information of existing craters, and divide the training set and test set;

[0010] S2. Construction of Mask R-CNN network model: Set network training parameters; use image pictures and marked pictures as the input and output of the convolutional neural network for training, adjust the network parameters to obtain the best training model, and output and save the pre-trained model and its training weight information;

[0011] S3. Construction of multi-source fusion data: Crop LOLA_DEM elevation data of the same area, generate a contour map of the area after realizing geographic information registration, fuse the contour map and image data, mark the craters in the area according to the elevation value change of the contour map, construct the training set and test set, and generate the target domain dataset;

[0012] S4. Construction of target domain dataset: Randomly crop and mark the fused image, draw the true edge information of the crater, and divide the training set and test set;

[0013] S5. Transfer learning model: Fine-tune the parameters and model structure of the pre-trained model that has achieved good training results on the source domain dataset, and then apply it to the new target domain dataset to obtain the crater edge recognition image, and draw the crater recognition image according to the crater position matching information;

[0014] S6. Result verification: Calculate the accuracy and precision of crater recognition, and save the new crater information; if the recognition accuracy and precision of the model meet the requirements, save the model for crater recognition; if the recognition accuracy and precision of the model do not meet the requirements, adjust the network model training parameters and retrain.

[0015] In the above method for identifying lunar craters based on data fusion and transfer learning, in step S2, the implementation method of the convolutional neural network is as follows:

[0016] (1) Feature extraction process: First, input the picture into the feature extraction layer of the model. The extraction layer is a backbone framework obtained by removing the last fully connected layer from an image classification residual network, which is used to obtain a feature map with high semantic information in the original picture. The information of each object is contained in this feature map;

[0017] (2) Prior selection process: There is a multiple relationship in size between the feature map obtained by feature extraction and the original image. According to this multiple relationship, the original image is divided into several prior boxes at regular intervals of several pixels. Then, the feature map is fed into 3*3 and 1*1 convolutions. The predicted offset is calculated based on the size of the output size. Then, the prior boxes are offset according to the predicted offset output and it is determined whether they are the background. Then, the invalid prior boxes are removed according to the probability ranking, and the remaining prior boxes are taken out in the order of the offset magnitude;

[0018] (3) Mapping process: The ROI Align method is used to find the position of the prior box on the original image to complete the mapping from the original image to the feature map;

[0019] (4) Prediction process: The prediction process includes a reconfirmation branch and a mask branch; Reconfirmation means that the obtained prior box is transformed into a one-dimensional vector through the average pooling algorithm, and the coordinate offset and class probability are obtained through the fully connected layer respectively, and then fine-tuning is performed to remove redundant boxes to improve the accuracy; The mask branch is to predict the pixel class inside the prior box.

[0020] In the above lunar crater recognition method based on data fusion and transfer learning, the residual unit of the residual network contains two convolutional layers. After the operation of the first convolutional layer, a batch normalization operation and a rectified linear unit operation are performed; After the operation of the second convolutional layer, the output feature map is superimposed with the input feature map, that is, the corresponding elements of the feature map are added; During the feature extraction process, a convolutional operation is performed once before each residual unit to double the number of filters, and a downsampling is performed once after each residual unit; During the image restoration process, a convolutional operation is performed once before each residual unit to halve or double the number of filters, and an upsampling is performed once after the residual unit.

[0021] In the above lunar crater recognition method based on data fusion and transfer learning, in the convolutional operation, the convolution kernel in the last output convolutional layer in the network is 1×1, and then a sigmoid operation is performed; The convolution kernel size in other convolutional layers in the network is 3×3, the stride is 1, and the padding strategy is zero-padding to ensure that the sizes of the input and output images are the same.

[0022] In the above lunar crater recognition method based on data fusion and transfer learning, the neural network parameters are set including the number of filters and the probability value of dropout. The initial number of filters in the network is set to 112, and the probability of dropout is set to 0.15; The optimizer of the network is the Adam optimizer; The loss calculation function used by the network is the binary cross-entropy.

[0023] In the above lunar crater recognition method based on data fusion and transfer learning, in step S3, the image data fusion method is as follows:

[0024] (1) Geometric registration: Extract images and elevation raster datasets of the same area on the lunar surface, and use the control point alignment algorithm based on the GIS platform to achieve geometric registration of multi-source lunar data;

[0025] (2) Contour map drawing: Extract DEM vector data of the specified area on the lunar surface to draw a contour map, set the contour interval to 2m, and then perform contour simplification. Use the Douglas-Peucker algorithm to thin linear features, exclude a large number of redundant geometric data points, and then perform contour smoothing. Use the Bezier curve fitting algorithm to obtain relatively smooth contour curves. Finally, perform data cleaning, remove contours with short length values, and label the corresponding elevation values for the contours;

[0026] (3) Match the contour map to the image map to achieve image data fusion.

[0027] In the above lunar crater recognition method based on data fusion and transfer learning, the geometric registration process includes identifying a series of ground control points and connecting the positions of the raster dataset with the positions in the spatial reference data coordinate system.

[0028] In the above lunar crater recognition method based on data fusion and transfer learning, in step S3, the object detection and semantic segmentation of the fused data are achieved by fine-tuning the pre-trained model Mask R-CNN. The process mainly includes:

[0029] (1) Define your own dataset: Label the new fused image data, draw the true boundaries of the crater images, construct the target domain dataset, and divide the training set and the test set;

[0030] (2) Model fine-tuning: Fine-tune the previously obtained Mask R-CNN network model, retain the first several layers of the pre-trained model, reconstruct the last multi-classifier layer, use the extreme learning machine to replace the fully connected layer of the network, overcome the overfitting phenomenon caused by insufficient sample quantity, and use the truncated gradient method to assist the learning of the deep neural network model parameters in the target domain during transfer learning, thereby implicitly helping the activated neurons to learn and restricting the learning of the unactivated neurons.

[0031] In the above lunar crater recognition method based on data fusion and transfer learning, in step S6, the specific process of result verification is as follows:

[0032] (1) Use the trained network to identify the crater edge, and then use the template matching method to obtain the positions of similar craters in the crater edge image;

[0033] (2) Compare the crater with the existing crater information;

[0034] (3) Calculate the accuracy and precision of recognition:

[0035] Precision: P = Tp / (Tp + Fp);

[0036] Recall: R = Tp / (Tp + Fn);

[0037] Model score calculation formula: F2 = 5 × P × R / (4 × P + R);

[0038] Impact crater discovery rate: DR1 = Fp / (Tp + Fp), DR2 = Fp / (Tp + Fn + Fp);

[0039] (4) Convert the newly discovered impact craters from pixel positions to lunar longitude and latitude, and store them in a record file for subsequent manual verification; if the performance evaluation of the network model meets the set requirements, then the network model is used for the lunar impact crater recognition task; otherwise, adjust the network training parameters and retrain.

[0040] Compared with the prior art, the lunar impact crater recognition method based on data fusion and transfer learning has the following beneficial effects:

[0041] 1. After fusing lunar remote sensing images and elevation data, the method of the present invention further realizes the detection and recognition of impact craters in lunar remote sensing images. The network established by this method is named TransCrater, and the established network consists of two impact crater detection channels and one impact crater recognition channel. Aiming at the problem of inaccurate recognition of a single data source in impact crater detection, the present invention proposes a new method of fusing elevation data and image data, and adopts a transfer learning strategy for the newly fused data, which can achieve accurate recognition of impact craters only with small sample data annotation, and obtain good recognition effects.

[0042] 2. The method of the present invention uses a convolutional neural network to identify the edges of impact craters on the lunar surface image, reducing the complexity of impact crater recognition, and improving the speed and accuracy of impact crater recognition.

[0043] 3. Compared with other classification methods, this method has a fast classification speed, a high probability of impact crater recognition, and a high impact discovery rate, and is applicable to the recognition of impact craters on various celestial bodies. Description of the Drawings

[0044] Figure 1 is the step block diagram of the lunar impact crater recognition method based on data fusion and transfer learning.

[0045] Figure 2 is the schematic diagram of the convolutional network in the lunar impact crater recognition method based on data fusion and transfer learning. Detailed Embodiments

[0046] The following further describes the specific implementation manners of the present invention in conjunction with the accompanying drawings and specific embodiments:

[0047] As Figure 1 and Figure 2 shown, the lunar crater recognition method based on data fusion and transfer learning includes the following steps:

[0048] S1. Construction of source domain dataset:

[0049] Randomly crop lunar image data, mark and draw crater pictures according to the diameter and coordinate information of existing craters, and divide the training set and test set;

[0050] S2. Construction of Mask R-CNN network model: Set network training parameters; use image pictures and annotation pictures as the input and output of the convolutional neural network for training, adjust network parameters to obtain the best training model, and output and save the pre-trained model and its training weight information;

[0051] S3. Construction of multi-source fusion data: Crop LOLA_DEM elevation data in the same area, generate a contour map of the area after realizing geographic information registration, fuse the contour map and image data, mark the craters in the area according to the elevation value change of the contour map, construct the training set and test set, and generate the target domain dataset; used to exclude concave craters and convex hills that are easily confused on single image data.

[0052] S4. Construction of target domain dataset: Randomly crop and mark the fused image, draw the true edge information of the crater, and divide the training set and test set;

[0053] S5. Transfer learning model: Fine-tune the parameters and model structure of the pre-trained model that has achieved good training results on the source domain dataset, and then apply it to the new target domain dataset to obtain the crater edge recognition image, and draw the crater recognition image according to the crater position matching information;

[0054] S6. Result verification: Calculate the accuracy and precision of crater recognition, and save the new crater information; if the recognition accuracy and precision of the model meet the requirements, save the model for crater recognition; if the recognition accuracy and precision of the model do not meet the requirements, adjust the network model training parameters and retrain.

[0055] In step S2, the implementation method of the convolutional neural network is as follows:

[0056] (1) Feature extraction process: First, input an image with a size of 512*512 into the feature extraction layer (backbone) of the model. The extraction layer is a backbone framework obtained by removing the last fully connected layer from an image classification residual network, which is used to obtain a feature map with high semantic information in the original image. The information of each object is contained in this feature map.

[0057] (2) Prior selection process: There is a multiple relationship in size between the feature map obtained by feature extraction and the original image. According to this multiple relationship, the original image is divided into several prior boxes at regular intervals of several pixels. Then, the feature map is passed into 3*3 and 1*1 convolutions. The predicted offset is calculated according to the size of the output. Then, the prior boxes are offset according to the predicted offset obtained and it is determined whether they are the background. Then, the invalid prior boxes are removed according to the probability ranking, and the remaining prior boxes are taken out in the order of the offset size. The function of this step is to let the model learn by itself which places in the original image have no objects and which places have objects.

[0058] (3) Mapping process: The multiple prior boxes obtained in the previous stage are not accurate enough and need to be further refined. The ROIAlign method is used to find the positions of the prior boxes on the original image to complete the mapping from the original image to the feature map.

[0059] (4) Prediction process: The prediction process includes a reconfirmation branch and a mask branch. Reconfirmation means that the obtained prior boxes are transformed into a one-dimensional vector through the average pooling algorithm, and the coordinate offset and class probability are obtained through the fully connected layer respectively, and then fine-tuning is performed to remove redundant boxes to improve the accuracy. The mask branch is to predict the pixel classes inside the prior boxes.

[0060] The residual unit of the residual network contains two convolutional layers. After the operation of the first convolutional layer, a batch normalization operation and a rectified linear unit operation are performed. After the operation of the second convolutional layer, the output feature map is superimposed on the input feature map, that is, the corresponding elements of the feature map are added. During the feature extraction process, a convolutional operation is performed once before each residual unit to double the number of filters, and a downsampling is performed once after each residual unit. During the image restoration process, a convolutional operation is performed once before each residual unit to halve or double the number of filters, and an upsampling is performed once after the residual unit.

[0061] In the convolutional operation, the convolution kernel in the last output convolutional layer of the network is 1×1, and then a sigmoid operation is performed. The convolution kernel size in other convolutional layers of the network is 3×3, the stride is 1, and the padding strategy is zero padding to ensure that the sizes of the input and output images are the same.

[0062] Set the neural network parameters including the number of filters and the probability value of dropout. The initial number of filters in the network is set to 112, and the probability of dropout is set to 0.15. The optimizer of the network is the Adam optimizer. The loss calculation function used by the network is binary cross-entropy.

[0063] In step S3, the method for fusing image data is as follows:

[0064] (1) Geometric registration: Extract images and elevation raster datasets of the same area on the lunar surface, and use the control point alignment algorithm based on the GIS platform to achieve geometric registration of lunar multi-source data.

[0065] (2) Draw a contour map: Extract the DEM vector data of the specified area on the lunar surface to draw a contour map. The contour interval is set to 2m. Then, perform contour simplification. Use the Douglas-Peucker algorithm to thin the linear features, excluding a large number of redundant geometric data points. Then, perform contour smoothing. Use the Bezier curve fitting algorithm to obtain relatively smooth contour curves. Finally, perform data cleaning, removing the contours with short length values, and annotating the corresponding elevation values for the contours.

[0066] (3) Match the contour map to the image map to achieve image data fusion.

[0067] The geometric registration process includes identifying a series of ground control points and connecting the positions of the raster dataset with the positions in the spatial reference data coordinate system. The control points refer to the positions that can be accurately identified in the raster dataset and the actual coordinates, such as rock outcrops on the lunar surface, crater walls, rilles, and fault structures.

[0068] In step S3, fine-tune the pre-trained model Mask R-CNN to achieve object detection and semantic segmentation of the fused data. The process mainly includes:

[0069] (1) Define your own dataset: Annotate the new fused image data, draw the true boundaries of the crater images, construct the target domain dataset, and divide it into a training set and a test set.

[0070] (2) Model fine-tuning: Fine-tune the previously obtained Mask R-CNN network model. Retain the first several layers of the pre-trained model, reconstruct the last multi-classifier layer, use the extreme learning machine to replace the fully connected layer of the network, overcome the overfitting phenomenon caused by insufficient sample quantity, and use the truncated gradient method (dropout) to assist the learning of the deep neural network model parameters in the target domain during transfer learning, thereby implicitly helping the activated neurons to learn and restricting the learning of the unactivated neurons.

[0071] In step S6, the specific process of result verification is as follows:

[0072] (1) Use the trained network to identify the edges of impact craters, and then use the method of template matching to obtain the positions of similar impact craters in the edge images of impact craters;

[0073] (2) Compare the impact craters with the existing impact crater information;

[0074] (3) Calculate the recognition accuracy and precision:

[0075] Precision: P = Tp / (Tp + Fp);

[0076] Recall: R = Tp / (Tp + Fn);

[0077] Model score calculation formula: F2 = 5 × P × R / (4 × P + R);

[0078] Impact crater discovery rate: DR1 = Fp / (Tp + Fp), DR2 = Fp / (Tp + Fn + Fp);

[0079] (4) Convert the newly discovered impact craters from pixel positions to lunar longitude and latitude, and store them in a record file for subsequent manual verification; if the performance evaluation of the network model meets the set requirements, then this network model is used for the lunar impact crater identification task; otherwise, adjust the network training parameters and retrain.

[0080] The principle of this lunar impact crater identification method based on data fusion and transfer learning:

[0081] There are many deep learning models for lunar impact crater identification, but most of them use a single data source for identification. Using a single image to identify impact craters, the advantage is that the edge information of the impact crater can be obtained, so as to draw the true contour of the impact crater. The disadvantage is that there are other raised landforms on the lunar surface, which are very similar to the sunken impact craters in the image data and are difficult to identify; using a single elevation data to identify impact craters, the advantage is that the change of elevation difference can be used to accurately judge whether it is an impact crater, but the disadvantage is that the elevation data generally has a low resolution and the true contour information of the impact crater cannot be obtained. Then, by combining the two types of data to construct a new fusion dataset, the existence of impact craters can be accurately judged, and the true contour of the impact crater can be accurately drawn, thus greatly improving the accuracy and precision of identification.

[0082] For the new fused data, if a corresponding brand-new neural network model is established, first of all, the workload will increase significantly, and the number of samples of the fused data is insufficient. The newly established model may not achieve good results. Therefore, a transfer learning strategy based on a pre-trained model is adopted. A Mask R-CNN model that has achieved good training results in the source domain (i.e., lunar image data) is used. And there are some labeled data available for learning in the target domain itself. The model can be directly applied to the target domain and then fine-tuned to construct the new model Trans Crater established by the present invention.

[0083] Obtain the image data of the area to be recognized on the lunar surface, generate training samples and test samples by manual annotation according to the existing impact crater position information and lunar image data, and establish a source domain dataset; construct a Mask R-CNN convolutional neural network model and set network parameters; use the LRO wide-angle camera data as the network input and the impact crater recognition image as the network output to train the model to obtain a pre-trained model with strong generalization ability and the weight information of the pre-trained model; obtain the DEM elevation data of the corresponding area on the lunar surface and realize the fusion with the image data of this area; make impact crater annotations on the fused data to generate a target domain dataset; establish a transfer learning model TransCrater based on the pre-trained model to realize the recognition of impact craters in the new fused data.

[0084] Compared with the prior art, the lunar impact crater recognition method based on data fusion and transfer learning has the following beneficial effects:

[0085] 1. After the method of the present invention fuses lunar remote sensing images and elevation data, it further realizes the detection and recognition of impact craters in lunar remote sensing images. The network established by this method is named TransCrater, and the established network consists of two impact crater detection channels and one impact crater recognition channel. Aiming at the problem of inaccurate recognition of a single data source in impact crater detection, the present invention proposes a new method of fusing elevation data and image data, and adopts a transfer learning strategy for the newly fused data, and only needs small sample data annotation to achieve accurate recognition of impact craters and obtains good recognition results.

[0086] 2. The method of the present invention uses a convolutional neural network to identify the edges of impact craters in lunar surface images, reducing the complexity of impact crater recognition, improving the speed and accuracy of impact crater recognition.

[0087] 3. Compared with other classification methods, this method has a fast classification speed, a high probability of impact crater recognition, and a high impact discovery rate, and is applicable to the recognition of impact craters on various celestial bodies.

[0088] Certainly, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by those skilled in the art within the scope of the essence of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for lunar crater identification based on data fusion and transfer learning, characterized in that, It includes the following steps: S1. Construction of source domain dataset: Randomly crop lunar image data, mark and draw crater pictures according to the diameter and coordinate information of existing craters, and divide the training set and test set; S2. Construction of Mask R-CNN network model: Set network training parameters; use image pictures and annotation pictures as the input and output of the convolutional neural network for training, adjust the network parameters to obtain the best training model, and output and save the pre-trained model and its training weight information; S3. Construction of multi-source fusion data: Crop LOLA_DEM elevation data in the same area, generate a contour map of this area after realizing geographic information registration, fuse the contour map and image data, and label the craters in this area according to the elevation value change of the contour map, construct the training set and test set, and generate the target domain dataset. The image data fusion method is as follows: (1) Geographic registration: Extract the image and elevation raster dataset of the same area on the lunar surface, and use the control point alignment algorithm based on the GIS platform to realize the geographic registration of lunar multi-source data; (2) Drawing contour map: Extract the DEM vector data of the specified area on the lunar surface to draw the contour map, set the contour interval to 2m, then perform the simplification operation of the contour, use the Douglas-Peucker algorithm to thin the linear elements, exclude a large amount of redundant geometric data points, and then perform the smoothing process of the contour, use the Bezier curve fitting algorithm to obtain a relatively smooth contour curve, and finally perform data cleaning, remove the contours with shorter length values, and label the corresponding elevation values for the contours; (3) Match the contour map to the image map to realize image data fusion; S4. Construction of target domain dataset: Randomly crop and label the craters on the fused image, draw the true edge information of the craters, and divide the training set and test set; S5. Transfer learning model: Fine-tune the parameters and model structure of the pre-trained model that has achieved good training results on the source domain dataset, and then apply it to the new dataset in the target domain to obtain the crater edge recognition image, and draw the crater recognition image according to the crater position matching information; S6. Result verification: Calculate the accuracy and precision of crater recognition, and save the new crater information; if the recognition accuracy and precision of the model meet the requirements, save the model for crater recognition; if the recognition accuracy and precision of the model do not meet the requirements, adjust the network model training parameters and retrain.

2. The lunar crater identification method based on data fusion and transfer learning according to claim 1, wherein In step S2, the implementation method of the convolutional neural network is as follows: (1) Feature extraction process: First, input the picture into the feature extraction layer of the model. The extraction layer is the backbone framework obtained by removing the last fully connected layer from an image classification residual network, which is used to obtain the feature map with high semantic information in the original picture. The information of each object is contained in this feature map; (2) Prior selection process: There is a multiple relationship in size between the feature map obtained by feature extraction and the original image. According to the multiple relationship, the original image is divided into several prior boxes at intervals of several pixels. Then the feature map is fed into 3*3 and 1*1 convolutions. The predicted offset is calculated according to the size of the output size. Then, the prior boxes are offset according to the predicted offset output and it is determined whether they are the background. Then, the invalid prior boxes are removed according to the probability ranking, and the remaining prior boxes are taken out in the order of the offset size; (3) Mapping process: Use the ROIAlign method to find the position of the prior box on the original image to complete the mapping from the original image to the feature map; (4) Prediction process: The prediction process includes a reconfirmation branch and a mask branch; Reconfirmation means that the obtained prior box is transformed into a one-dimensional vector through the average pooling algorithm, and the coordinate offset and class probability are obtained through the fully connected layer respectively, and then fine-tuning is performed to delete redundant boxes to improve the accuracy; The mask branch means predicting the pixel categories inside the prior box.

3. The lunar impact crater recognition method based on data fusion and transfer learning according to claim 2, wherein The residual unit of the residual network contains two convolutional layers. After the operation of the first convolutional layer, a batch normalization operation and a rectified linear unit operation are performed; After the operation of the second convolutional layer, the output feature map is superimposed with the input feature map, that is, the corresponding elements of the feature map are added; During the feature extraction process, the number of filters is doubled by performing a convolutional operation before each residual unit, and downsampling is performed after each residual unit; During the image restoration process, the number of filters is halved or doubled by performing a convolutional operation before each residual unit, and upsampling is performed after the residual unit.

4. The method for lunar crater recognition based on data fusion and transfer learning according to claim 2, wherein, In the convolutional operation, the convolution kernel in the last output convolutional layer of the network is 1×1, and then a sigmoid operation is performed; In other convolutional layers of the network, the convolution kernel size is 3×3, the stride is 1, and the padding strategy is zero-padding to ensure that the sizes of the input and output images are the same.

5. The lunar crater identification method based on data fusion and transfer learning according to claim 2, wherein, Set the neural network parameters including the number of filters and the probability value of dropout. The initial number of filters in the network is set to 112, and the probability of dropout is set to 0.15; The optimizer of the network is the Adam optimizer; The loss calculation function used by the network is the binary cross entropy.

6. The method for lunar crater recognition based on data fusion and transfer learning according to claim 1, wherein The georegistration process includes identifying a series of ground control points to connect the position of the raster dataset with the position of the spatial reference data coordinate system.

7. The lunar crater identification method based on data fusion and transfer learning according to claim 1, characterized in that In step S3, the object detection and semantic segmentation of the fused data are realized by fine-tuning the pre-trained model Mask R-CNN. The process mainly includes: (1) Define your own dataset: Annotate the new fused image data, draw the true boundary of the crater image, construct the target domain dataset, and divide the training set and the test set; (2) Model fine-tuning: Fine-tune the previously obtained Mask R-CNN network model. Retain the first several layers of the pre-trained model, reconstruct the last multi-classifier layer, replace the fully connected layer of the network with an extreme learning machine to overcome the overfitting phenomenon caused by insufficient sample quantity, and use the truncated gradient method to assist in the learning of the parameters of the deep neural network model in the target domain during transfer learning, thereby implicitly helping the activated neurons to learn and restricting the learning of unactivated neurons.

8. The lunar crater recognition method based on data fusion and transfer learning according to claim 1, wherein In step S6, the specific process of result verification is as follows: (1) Use the trained network to identify the crater edge, and then use the template matching method to obtain the positions of similar craters in the crater edge image; (2) Compare the crater with the existing crater information; (3) Calculate the recognition accuracy and precision: Precision: P = Tp / (Tp + Fp); Recall: R = Tp / (Tp + Fn); Model score calculation formula: F2 = 5 × P × R / (4 × P + R); Crater discovery rate: DR1 = Fp / (Tp + Fp), DR2 = Fp / (Tp + Fn + Fp); (4) Convert the newly discovered crater from pixel positions to lunar longitude and latitude, and store it in a record file for subsequent manual verification; If the performance evaluation of the network model meets the set requirements, then the network model is used for the lunar crater recognition task; otherwise, adjust the network training parameters and retrain.

Images