A domain adaptation crater detection method, system, device and storage medium
By combining a feature pyramid network and a progressive domain adaptive network with low-level features and a circular boundary enhancement module, the problems of difficult dataset acquisition and inaccurate feature alignment in impact crater detection are solved, achieving high efficiency and accuracy in cross-domain detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- MACAU UNIV OF SCI & TECH
- Filing Date
- 2023-01-09
- Publication Date
- 2026-04-21
AI Technical Summary
Existing impact crater detection methods suffer from difficulties in acquiring datasets, inaccurate feature alignment, and insufficient interpretability of deep models, resulting in poor cross-domain detection performance.
Feature extraction is performed by combining a feature pyramid network with a low-level feature enhancement module and a circular boundary enhancement module. Boundary box detection is performed by a progressive domain adaptive network and a region recommendation network. Similarity weighting is performed by a weighted instance-level alignment network to improve detection accuracy.
It effectively improves the efficiency and accuracy of impact crater detection, especially in cross-domain detection between different planetary terrains.
Smart Images
Figure CN116091905B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing technology, and in particular to a domain-adaptive impact crater detection method, system, device, and storage medium. Background Technology
[0002] With the application of science and technology in outer space exploration, planets such as Mars and the Moon are gradually revealing their mysteries. Impact crater-dominated planetary topographic features are of great significance to researchers; the number and size of impact craters are crucial for analyzing a planet's history and can aid in spacecraft navigation and landing. Therefore, fast and accurate impact crater detection algorithms will contribute to better exploration of other planets.
[0003] However, domain-adaptive detection across impact crater datasets remains a significant challenge in the following three aspects:
[0004] 1) Regarding datasets, well-annotated datasets are expensive and difficult to obtain, and there are currently no real-world cross-domain impact crater detection datasets. Existing methods are mainly used for general natural scenes. Even if these methods may have the potential to be applied to impact crater scenes, they do not effectively utilize the features of the craters. Therefore, it is naturally impossible to achieve high domain generalization performance based on existing data.
[0005] 2) Regarding feature alignment, existing methods can only roughly align two distributions and cannot effectively align samples at the edges of the distributions. Furthermore, minimizing the distance between two domains severely compromises some of the model's distribution awareness. In most cases, looser models perform better than tighter models, but research in this area is limited.
[0006] 3) Regarding the interpretability of deep models, distribution metrics such as maximum mean distance (MMD), graph distance, or prototype distance can be used to align the source and target domains. However, this process is achieved by minimizing the loss function during training, which is agnostic and can usually only be verified through experimental results. Furthermore, there is no theoretical research on which distribution metric is best suited for the current task.
[0007] Therefore, improving the efficiency of impact crater detection is an urgent problem to be solved. Summary of the Invention
[0008] In view of this, embodiments of the present invention provide a domain-adaptive impact crater detection method, system, device, and storage medium, which can effectively improve the detection efficiency of impact craters.
[0009] On one hand, embodiments of the present invention provide a domain-adaptive impact crater detection method, comprising:
[0010] Obtain task samples; whereby task samples include source domain samples with known labels and target domain samples with unknown labels;
[0011] Feature pyramid network is used to extract features from task samples to obtain feature maps;
[0012] The feature extraction network includes a low-level feature enhancement module and a circular boundary enhancement module; the feature map includes a first feature map and a second feature map.
[0013] A linear transformation is performed on the feature map, and intermediate subspace features are obtained through a progressive domain adaptive network.
[0014] Using a region recommendation network, the intermediate subspace features and the second feature map are analyzed to obtain the intermediate bounding box detection results and the target domain bounding box detection results.
[0015] By using a weighted instance-level alignment network, the similarity weights of the intermediate bounding box detection results and the target domain bounding box detection results are applied to obtain the target detection results.
[0016] Optionally, it also includes:
[0017] The overall network is trained by using classification loss and regression loss as loss functions.
[0018] The overall network includes a weighted instance-level aligned network.
[0019] Optionally, it also includes:
[0020] A feature pyramid network is constructed based on a low-level feature enhancement module and a circular boundary enhancement module, combined with a convolutional network.
[0021] The feature pyramid network consists of two branches, each of which includes a low-level feature enhancement module and a circular boundary enhancement module connected side-by-side by a convolutional network.
[0022] Optionally, a feature pyramid network is used to extract features from the task samples to obtain feature maps, including:
[0023] The task samples are processed by convolutional networks;
[0024] The low-level feature enhancement module utilizes an attention mechanism to enhance the low-level semantic information of the task samples after convolution.
[0025] The circular boundary enhancement module utilizes diagonal convolution to perform domain-invariant feature enhancement on the task samples after convolution processing;
[0026] The features obtained by fusing low-level semantic information enhancement processing and domain-invariant feature enhancement processing are used to obtain a feature map;
[0027] The first feature map is obtained by extracting features from source domain samples with known labels, and the second feature map is obtained by extracting features from target domain samples with unknown labels.
[0028] Optionally, the low-level feature enhancement module includes a first branch, a second branch, and a third branch. The low-level feature enhancement module utilizes an attention mechanism to perform low-level semantic information enhancement processing on the convolutionally processed task samples, including:
[0029] Based on the task samples after convolution processing, the first feature, second feature and third feature are obtained by 1×1 convolution in the first branch, the second branch and the third branch, respectively.
[0030] Based on the first feature, the fourth feature is obtained by passing through Softmax in the first branch;
[0031] Multiply by the fourth and second features, and add to the third feature.
[0032] Optionally, the circular boundary enhancement module includes a fourth branch and a fifth branch. The circular boundary enhancement module utilizes diagonal convolution to perform domain-invariant feature enhancement processing on the convolutionally processed task samples, including:
[0033] Based on the task samples after convolution processing, the fifth feature is obtained by using shift and subtraction operations with two sets of 1×1 convolutions in the first branch.
[0034] Based on the task samples after convolution processing, the sixth feature is obtained by 3×3 convolution in the fifth branch;
[0035] Add the sixth feature to the fifth feature.
[0036] Optionally, a linear transformation is performed on the feature map to obtain intermediate subspace features through a progressive domain adaptive network, including:
[0037] A linear transformation is performed on the feature map to obtain a first subspace and a second subspace; wherein the first subspace is obtained by linear transformation of the first feature map, and the second subspace is obtained by linear transformation of the second feature map;
[0038] Based on the first and second subspaces, multiple intermediate domain features are obtained through adaptive changes in the hyperparameters of the progressive domain adaptive network.
[0039] One of the features from multiple intermediate domain features is used as the intermediate subspace feature.
[0040] On the other hand, embodiments of the present invention provide a domain-adaptive impact crater detection system, comprising:
[0041] The first module is used to obtain task samples; the task samples include source domain samples with known labels and target domain samples with unknown labels.
[0042] The second module is used to extract features from task samples using a feature pyramid network to obtain feature maps.
[0043] The feature extraction network includes a low-level feature enhancement module and a circular boundary enhancement module; the feature map includes a first feature map and a second feature map.
[0044] The third module is used to perform linear transformation on the feature map and obtain intermediate subspace features through a progressive domain adaptive network.
[0045] The fourth module is used to analyze the intermediate subspace features and the second feature map using a region recommendation network to obtain the intermediate bounding box detection results and the target domain bounding box detection results.
[0046] The fifth module is used to perform similarity weighting on the intermediate bounding box detection results and the target domain bounding box detection results using a weighted instance-level alignment network to obtain the target detection results.
[0047] On the other hand, embodiments of the present invention provide a domain-adaptive impact crater detection device, including a processor and a memory;
[0048] Memory is used to store programs;
[0049] The processor executes the program as described above.
[0050] On the other hand, embodiments of the present invention provide a computer-readable storage medium storing a program that is executed by a processor to implement the method described above.
[0051] This invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the aforementioned method.
[0052] This invention first acquires task samples, which include source domain samples with known labels and target domain samples with unknown labels. A feature pyramid network is used to extract features from the task samples, obtaining feature maps. The feature extraction network includes a low-level feature enhancement module and a circular boundary enhancement module. The feature maps include a first feature map and a second feature map. A linear transformation is performed on the feature maps, and an intermediate subspace feature is obtained through a progressive domain adaptive network. A region recommendation network is used to analyze the intermediate subspace feature and the second feature map, respectively, to obtain intermediate bounding box detection results and target domain bounding box detection results. A weighted instance-level alignment network is used to weight the intermediate bounding box detection results and the target domain bounding box detection results based on similarity, resulting in the target detection result. This invention can effectively improve the detection efficiency of impact craters. Attached Figure Description
[0053] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0054] Figure 1 A schematic flowchart of a domain-adaptive impact crater detection method provided in an embodiment of the present invention;
[0055] Figure 2 A schematic diagram of a framework structure for domain-adaptive impact crater detection provided in an embodiment of the present invention;
[0056] Figure 3 A schematic diagram of the architecture of the low-level feature enhancement module provided in an embodiment of the present invention;
[0057] Figure 4 This is a schematic diagram of the architecture of the circular boundary enhancement module provided in an embodiment of the present invention;
[0058] Figure 5 This is a schematic diagram showing the effect of diagonal convolution in extracting circular features according to an embodiment of the present invention;
[0059] Figure 6 A schematic diagram illustrating an algorithm example provided in an embodiment of the present invention;
[0060] Figure 7 A schematic diagram of a dataset instance provided in an embodiment of the present invention;
[0061] Figure 8 This is a schematic diagram of the visualized detection results of DCD provided in an embodiment of the present invention;
[0062] Figure 9This is a schematic diagram showing the effect comparison of the visualization detection results provided in the embodiments of the present invention. Detailed Implementation
[0063] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0064] On the one hand, refer to Figure 1 The present invention provides a domain-adaptive impact crater detection method, comprising:
[0065] S100, Obtain task samples;
[0066] It should be noted that the task samples include source domain samples with known labels and target domain samples with unknown labels.
[0067] Specifically, such as Figure 2 As shown in the Input section, it is assumed that the data in the source domain (e.g., Mars Craters) and the target domain (e.g., Luna Craters) are similar, meaning their geodesic distances in the manifold space are absolutely close. However, due to the different distributions, a model trained on the source domain cannot be directly applied to the target domain. The dataset in the source domain Ds contains labeled samples, but the dataset in the target domain Dt contains only unlabeled samples. These samples are extracted from two different distributions. The goal of this embodiment is to train a deep model using only the labels of Ds that perform well on Dt.
[0068] S200. Using the feature pyramid network, feature extraction is performed on the task samples to obtain feature maps;
[0069] The feature extraction network includes a low-level feature enhancement module and a circular boundary enhancement module; the feature map includes a first feature map and a second feature map.
[0070] It should be noted that in some embodiments, the feature pyramid network is constructed by combining a low-level feature enhancement module and a circular boundary enhancement module with a convolutional network. The feature pyramid network includes two branches, each of which includes a low-level feature enhancement module and a circular boundary enhancement module connected in parallel by a convolutional network.
[0071] In some embodiments, the task samples are convolved using a convolutional network; a low-level feature enhancement module uses an attention mechanism to enhance the low-level semantic information of the convolved task samples; a circular boundary enhancement module uses diagonal convolution to enhance the domain-invariant features of the convolved task samples; and the features obtained from the low-level semantic information enhancement and the domain-invariant feature enhancement are fused to obtain a feature map; wherein, the first feature map is obtained by feature extraction from source domain samples with known labels, and the second feature map is obtained by feature extraction from target domain samples with unknown labels.
[0072] In some embodiments, such as Figure 3 As shown, the low-level feature enhancement module includes a first branch, a second branch, and a third branch. This module utilizes an attention mechanism to perform low-level semantic information enhancement on the convolutionally processed task samples. This includes: based on the convolutionally processed task sample (i.e., Feature Map A), performing 1×1 convolutions in the first, second, and third branches to obtain the first feature (i.e., Feature Map B), the second feature, and the third feature (not shown in the figure); based on the first feature, performing a softmax operation in the first branch to obtain the fourth feature (i.e., Attention Map S); multiplying by the fourth and second features and adding them to the third feature. The final Feature Map C is obtained.
[0073] In some embodiments, such as Figure 4 As shown, the circular boundary enhancement module includes a fourth branch and a fifth branch. This module utilizes diagonal convolution to perform domain-invariant feature enhancement on the convolutionally processed task samples. This includes: using two sets of 1×1 convolutions with shift subtraction operations in the first branch to obtain the fifth feature (not shown in the figure); using a 3×3 convolution in the fifth branch to obtain the sixth feature (not shown in the figure); and adding the sixth feature to the fifth feature to obtain the final output feature.
[0074] Specifically, such as Figure 2 As shown in the Feature Extraction section, the input consists of source domain samples with known labels (e.g., Mars Craters) and target domain samples with unknown labels (e.g., Luna Craters), both set to the same size. The two samples are passed through the same convolutional network and augmented using a Circular Boundary Enhancement Module (CEEM) and a Low-Level Feature Enhancement Module (LFEM) respectively, yielding fused and enhanced features. These features are then subjected to several convolutional operations to obtain feature maps F1 and F2. Finally, a feature pyramid network with shared weights is used to extract image features from the source and target domains respectively, resulting in deep features F1 and F2, which are the feature maps of the last layer of the network.
[0075] Among them, such as Figure 3 The diagram shows the structure of the Low-Level Feature Enhancement Module (LFEM). Based on extensive experiments, it was found that neural networks exhibit a certain robustness to object features across different domains, especially in shallow networks. This is because neural networks can extract low-level features such as edges, colors, and textures. These low-level features show almost no inter-domain differences. When a model trained in one domain is directly applied to another, the heatmap of shallow CNN features shows that most impact crater features useful for classification are activated. Based on this observation, to further enhance these domain-invariant information, a method is introduced such as... Figure 3 The attention mechanism shown is illustrated, and a Low-Level Feature Enhancement Module (LFEM) is proposed. LFEM uses the attention mechanism to enhance low-level semantic information such as texture and color.
[0076] Since shallow feature layers are typically large, the original self-attention algorithm requires significant computational resources. Therefore, a simplified self-attention module is used by computing a global attention map and sharing it across all query locations. This effectively reduces the number of triggers while achieving the same performance. In LFEM, feature map A is first processed through a 1×1 convolutional layer to obtain B, and then the attention map S is obtained through the Softmax function. Next, the feature map obtained by multiplying feature map A (obtained through a 1×1 convolution) with S is multiplied to obtain the resulting feature map, which is then added to the feature map obtained by convolving feature map A through a 1×1 convolution, resulting in the final feature map C, defined as:
[0077]
[0078] Among them, W q W k and W v Let C represent the linear transformation matrix. i Representing feature maps C and A i Representing feature maps A and B j and B m Representing feature maps B and N p The number of feature maps is indicated by the subscripts i, j, and m, which represent the indices of the feature maps.
[0079] Among them, such as Figure 4 The diagram shows the structure of the Circular Edge Enhancement Module (CEEM). Circular edge information is crucial for understanding impact craters. Inspired by the Robert operator, a Circular Edge Enhancement Module was proposed to extract this information. Experiments revealed that diagonal convolution operations outperform vertical and horizontal convolution operations when extracting circular features; this is determined by the characteristics of circular objects. Figure 5As shown, when using diagonal convolution to extract circular features, almost no boundary features are lost. However, using convolution in the vertical and horizontal directions leads to the loss of some important boundary information, such as... Figure 5 As shown in red in the image. Furthermore, the circular boundary features obtained through diagonal convolution are evident (larger values).
[0080] Based on this, a circular boundary enhancement module is proposed to simulate diagonal convolution and better extract the directional features of circular boundaries in deep networks. CEEM enhances domain-invariant features through diagonal convolution based on the shape features of the impact crater. Figure 4 As shown, firstly, two sets of 1×1 convolutions are used in the features along different diagonal directions, followed by shift subtraction. Finally, circular boundary enhancement features and the original features are added to obtain the final output features.
[0081] S300. Perform a linear transformation on the feature map and obtain intermediate subspace features through a progressive domain adaptive network.
[0082] It should be noted that in some embodiments, a linear transformation is performed on the feature map to obtain a first subspace and a second subspace; wherein, the first subspace is obtained by linear transformation of the first feature map, and the second subspace is obtained by linear transformation of the second feature map; based on the first subspace and the second subspace, multiple intermediate domain features are obtained by adaptively changing the hyperparameters of the progressive domain adaptive network; one of the multiple intermediate domain features is sampled as an intermediate subspace feature.
[0083] Specifically, such as Figure 2 As shown in the ProgressiveDpmain Adaption section, after linear transformation of feature maps F1 and F2, points S1 and S2 on the manifold can be obtained. Then, intermediate subspace features are obtained by controlling the hyperparameter t. These features are controlled by the hyperparameter; that is, the smaller the hyperparameter, the closer the feature is to the source domain data, and the larger the hyperparameter, the closer the feature is to the target domain data. Multiple intermediate domain features are obtained by gradually increasing the value of the hyperparameter.
[0084] S400. Using a region recommendation network, the intermediate subspace features and the second feature map are analyzed respectively to obtain the intermediate bounding box detection results and the target domain bounding box detection results.
[0085] Specifically, such as Figure 2 As shown, by sampling one feature from multiple intermediate domain features (i.e., intermediate subspace features) and feeding it into the subsequent Region Recommendation Network (RPN), multiple instance-level bounding box detection results (i.e., intermediate bounding box detection results) are obtained. Correspondingly, the F2 obtained by extracting the target domain sample features is also fed into the Region Recommendation Network (RPN) to obtain the target domain bounding box detection results.
[0086] S500: Using a weighted instance-level alignment network, the similarity weights of the intermediate bounding box detection results and the target domain bounding box detection results are applied to obtain the target detection results.
[0087] It should be noted that in some embodiments, the method further includes: training the overall network using classification loss and regression loss as loss functions; wherein the overall network includes a weighted instance-level alignment network. Optionally, the overall network may include a progressive domain adaptation network, a region recommendation network, and a weighted instance-level alignment network.
[0088] Specifically, such as Figure 2 As shown in the Crater Detection section, a gradient inverse layer is used to shorten the distribution difference between the two branches, and a weighted instance-level attention network (i.e., a weighted instance-level alignment network, WIAN) is used to better guide the target domain features to obtain accurate results. This network first calculates the similarity between the source domain and the target domain, and then weights them according to the similarity. During training, a weighted attention loss function is used, and the classification loss and regression loss of conventional object detection are combined as the overall loss function of the entire network.
[0089] In a preferred embodiment, the loss function of this invention follows the standard Fast R-CNN framework for region-adaptive crater detection. This framework consists of two stages: a Region Proposal Network (RPN) stage and an R-CNN stage. Each stage has two sub-tasks: crater classification and location regression. In the RPN stage, the network predicts preliminary crater class probabilities and obtains approximate bounding box results. In the R-CNN stage, the network predicts final crater probabilities and precise bounding box regression results. Therefore, the multi-task loss function is defined as:
[0090]
[0091] In the formula, L det＝L (·) represents the overall loss function, p i t represents the output class probability. i N represents the bounding box position. cls L represents the number of samples for classification. cls (·) represents the classification loss function. Indicates the label category, N reg L represents the number of regression samples. reg (·) represents the regression loss function. This indicates the bounding box label.
[0092] The loss function for a weighted instance-level attention network (weighted instance-level alignment network, WIAN) is as follows:
[0093]
[0094] In the formula, Let D represent the loss function of WIAN. i p represents the hyperparameter. i,j This represents the output of the domain classifier for the j-th proposed image at the instance level.
[0095] The technical solution of the present invention will be further described below with reference to specific embodiments. It should be understood that the following is an explanation of the present invention and should not be regarded as a limitation of the present invention.
[0096] The specific algorithm implementation of this invention is as follows: Figure 6 As shown, let X n1 ∈R N Let X represent the input of the source domain. n2 ∈R N The input represents the target domain. Image features from the source and target domains are extracted using a feature pyramid network with shared weights, and deep features F1 and F2 are obtained as feature maps of the last layer of the network. Here, S1 and S2 represent the n-dimensional general subspaces obtained by linear transformation of F1 and F2, respectively. To obtain an intermediate subspace between S1 and S2, enabling the network to incrementally learn cross-domain features, S1 and S2 are identified as Grassmann manifolds G. N，d The two points on top.
[0097] In this embodiment of the invention, it is assumed that the subspaces of the source domain and the target domain are similar, and the geodesic distance between the source data subspace S1 and the target data subspace S2 is sufficiently small. Therefore, the Euclidean distance can be used instead of the geodesic distance of the isotropic curve on the Grassmann manifold. Based on the similarity between the data features of the source domain and the target domain, to obtain the intermediate subspace, G is first... N Considered SO (N) The entropy space is given, and the geodesic path starting from S1 is a single-parameter exponential flow: ψ(t')=Q exp(t'B)J, where ψ(·) represents the exponential flow, t' represents the control factor, Q represents the projection matrix, B represents the transformation matrix, and J represents the original feature. Detailed calculations are as follows... Figure 6 As shown in the algorithm, matrix A represents the direction and geodesic velocity. We can obtain intermediate subspaces using the exponential flow in row 10 of the algorithm by changing the value of t' from 0 to 1. During training, we obtain multiple subspaces by gradually increasing the value of t' and projecting the original data into these subspaces respectively to obtain multiple inter-domain features. Then, we send the obtained intermediate domain features to the Region Proposal Network (RPN) for training. During training, we gradually increase t' to complete the domain adaptation from the source domain to the target domain.
[0098] Instance-level feature alignment helps reduce local differences between instances, such as appearance, size, and viewpoint. However, in the early stages of training, deep networks cannot effectively learn instance information from the target domain, leading to inaccurate detection locations and thus hindering effective alignment. To address this issue, this invention proposes a Weighted Instance Alignment Module (WIAM), which adaptively enhances useful instance-level feature alignment based on the similarity of instance-level information from two domains. To calculate the similarity of proposals from different domains, an affinity matrix is used to establish a soft correspondence between the proposed features of the source and target domains.
[0099] The specific algorithm for asymptotic domain adaptation is as follows: Figure 6 As shown, the network weights are first initialized and the corresponding hyperparameters are set. The input consists of data from two domains: the source domain has labels, while the target domain does not. The two data points are input into the network to obtain features F1 (i.e., F1) and F2 (i.e., F2), respectively. After linear transformation, S1 (i.e., S1) and S2 (i.e., S2) are obtained, which are points on the Grassmann manifold. Then, the N×N orthogonal complement Q of S1 is calculated. T The CS decomposition of S2 is performed, then θ is calculated based on the diagonal elements, the fast diagonal matrix B is calculated, singular value decomposition is performed on B to obtain the diagonal matrix, and finally the expression of the final subspace elements is obtained.
[0100]
[0101] In the formula, F′ represents the self-space characteristic, Q represents the projection matrix, V1 represents the source domain velocity, Γ represents the source domain diagonal matrix, t' represents the adjustment factor, V2 represents the target domain velocity, Σ represents the target domain diagonal matrix, and X n1 Represents source domain characteristics.
[0102] To evaluate the effectiveness of the embodiments of the present invention, a Domain Adaptive Crater Detection (DACD) dataset is proposed based on the publicly available Daytime Mars Impact Crater Dataset (MDCD). This dataset contains 1000 images and more than 20,000 impact craters in two domains. Figure 7 The image shown is a publicly available dataset example. Figure 8 The image shows the visualized detection results of DCD and Figure 9 To visualize the migration of detection results from the DCD dataset to the LRONAC dataset, Table 1 shows a comparison of the results from this invention's embodiment with other Mars observation methods based on the DCD dataset, and Table 2 shows a comparison of the results from this invention's embodiment with other methods migrating the results from the DCD dataset to the LRONAC dataset.
[0103] Table 1
[0104] Methods Recall (%) Precision (%) DAFaster 55.15±0.13 50.91±0.20 SWFaster 56.21±0.16 48.87±0.19 HTCN 46.22±0.17 41.31±0.11 GPA 58.3±0.08 51.26±0.21 EPM 55.69±0.22 48.09±0.15 DIDN 59.92±0.26 51.81±0.25 DSS 60.12±0.37 53.51±0.12 SDA 59.85±0.16 52.34±0.29 CFFA 58.74±0.23 51.13±0.19 RPNPA 59.43±0.26 52.24±0.24 KTNet 60.16±0.15 51.11±0.22 SSAL 59.55±0.23 52.44±0.31 ICCR-VDD 58.77±0.09 51.67±0.14 UMT 59.79±0.14 52.88±0.20 MeGA 60.22±0.17 52.81±0.21 PDAN(ours) 62.31±0.13 54.83±0.06
[0105] Table 2
[0106]
[0107]
[0108] Compared to state-of-the-art (SOTA) methods based on the DAD dataset (from day to night on Mars), the proposed PDAN achieves competitive performance in Table I, achieving 62.31% recall and 54.83% precision. Compared to the best methods MeGA and UMT, recall and precision are improved by 2% and 2%, respectively. Furthermore, PDAN surpasses DAFaster and SWFaster by 7% and 6% in recall, and by 4% and 6% in precision, respectively.
[0109] Tables 1 and 2 show the visualization results of the proposed method and mainstream methods in the DAD dataset. Clearly, due to the superiority of this method, it can detect more impact craters, thus achieving higher accuracy. It is worth noting that, compared to other methods, the method proposed in this embodiment is applicable to small impact craters because information about small objects is usually located in the shallow layer, and the proposed LFEM can enhance the features of small impact craters.
[0110] To further demonstrate the effectiveness of the algorithm, the data distribution differences between domains were increased. DCD was used as the source domain data, and LRONAC data was used as the target domain data. Table 2 shows a comparison between the proposed method and other state-of-the-art (SOTA) methods. The detection results show that, compared to Mars, the features of craters are not very obvious, and other methods typically reduce detection precision and recall. However, the proposed method exhibits better performance because it can learn a looser model and has better generalization performance. Compared to the best method, MeGA, the method proposed in this embodiment improves recall and precision by approximately 2% and 2.5%, respectively.
[0111] Extensive experiments based on DCD and other publicly available datasets demonstrate that the network proposed in this embodiment can effectively eliminate the effects of domain migration and achieve superior results.
[0112] In summary, the domain-adaptive impact crater detection method of this invention provides a domain-adaptive impact crater detection algorithm, including: a Progressive Domain Adaptive Network (PDAN), which progressively learns knowledge from the source domain to the target domain by generating intermediate domain features from projected features along geodesics on the Grassmann manifold. Secondly, to further extract domain-invariant features and utilize impact crater features, a Low-Level Feature Enhancement Module (LFEM) and a Circular Boundary Enhancement Module (CEEM) are proposed. Thirdly, a Weighted Instance-Level Alignment Network (WIAN) is proposed to adaptively align instance-level features, further reducing the difference between the two domains and improving performance. The beneficial effects of this invention include: progressively learning features from the source domain to the target domain through the Progressive Domain Adaptive Network (PDAN); a Weighted Instance-Level Alignment Network (WIAN) is proposed to further eliminate instance-level bias in cross-domain information; and, based on the characteristics of impact craters, a Low-Level Feature Enhancement Module (LFEM) and a Circular Boundary Enhancement Module (CEEM) are proposed to further enhance the extraction capability of region-invariant information. Among them, PDAN: The network learns features from the source domain to the target domain step by step through the projection features of the subspace on the Grassman manifold, and gradually realizes the domain adaptation process; LFEM uses the attention mechanism to enhance low-level semantic information such as texture and color; CEEM enhances domain-invariant features through diagonal convolution based on the shape features of the impact crater; WIAN: To align the source domain and target domain features at the instance level.
[0113] On the other hand, embodiments of the present invention provide a domain-adaptive impact crater detection system, comprising: a first module for acquiring task samples; wherein the task samples include source domain samples with known labels and target domain samples with unknown labels; a second module for extracting features from the task samples using a feature pyramid network to obtain feature maps; wherein the feature extraction network includes a low-level feature enhancement module and a circular boundary enhancement module; the feature maps include a first feature map and a second feature map; a third module for performing linear transformation on the feature maps and obtaining intermediate subspace features through a progressive domain-adaptive network; a fourth module for analyzing the intermediate subspace features and the second feature map using a region recommendation network to obtain intermediate bounding box detection results and target domain bounding box detection results; and a fifth module for weighting the intermediate bounding box detection results and target domain bounding box detection results using a weighted instance-level alignment network to obtain target detection results.
[0114] The content of the method embodiments of the present invention is applicable to the system embodiments. The specific functions implemented in the system embodiments are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above methods.
[0115] Another aspect of the present invention provides a domain-adaptive impact crater detection device, including a processor and a memory;
[0116] Memory is used to store programs;
[0117] The processor executes the program as described above.
[0118] The content of the method embodiments of the present invention is applicable to the device embodiments. The specific functions implemented by the device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above methods.
[0119] Another aspect of this invention provides a computer-readable storage medium storing a program that is executed by a processor to implement the methods described above.
[0120] The content of the method embodiments of the present invention is applicable to the computer-readable storage medium embodiments. The specific functions implemented by the computer-readable storage medium embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above methods.
[0121] This invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the aforementioned method.
[0122] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this invention are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is altered and sub-operations described as part of a larger operation are executed independently.
[0123] Furthermore, although the invention has been described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the described functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding the invention. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of conventional skill of an engineer. Therefore, those skilled in the art can implement the invention as set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and not intended to limit the scope of the invention, which is determined by the full scope of the appended claims and their equivalents.
[0124] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0125] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution means, apparatus, or device (such as a computer-based device, a processor-including device, or other means that can fetch and execute instructions from, or in conjunction with, an instruction execution means, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution means, apparatus, or device.
[0126] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.
[0127] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution device. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.
[0128] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
[0129] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.
[0130] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of the present invention.
Claims
1. A domain adaptation crater detection method, characterized in that, include: Obtain task samples; wherein, the task samples include source domain samples with known labels and target domain samples with unknown labels; Feature pyramid network is used to extract features from the task samples to obtain feature maps; The feature pyramid network includes a low-level feature enhancement module and a circular boundary enhancement module; the feature map includes a first feature map and a second feature map. The feature map is linearly transformed, and intermediate subspace features are obtained through a progressive domain adaptive network. The step of performing a linear transformation on the feature map and obtaining intermediate subspace features through a progressive domain adaptive network includes: A linear transformation is performed on the feature map to obtain a first subspace and a second subspace; wherein the first subspace is obtained by linear transformation of the first feature map, and the second subspace is obtained by linear transformation of the second feature map; Based on the first subspace and the second subspace, multiple intermediate domain features are obtained through the adaptive change of the hyperparameters of the progressive domain adaptive network. By sampling one of the features of the multiple intermediate domain features as an intermediate subspace feature; Using a region recommendation network, the intermediate subspace features and the second feature map are analyzed respectively to obtain the intermediate bounding box detection results and the target domain bounding box detection results; By using a weighted instance-level alignment network, the detection results of the intermediate bounding box and the target domain bounding box are weighted by similarity to obtain the target detection result.
2. The method of claim 1, wherein, Also includes: The overall network is trained using classification loss and regression loss as loss functions. The overall network includes the weighted instance-level alignment network.
3. The method of claim 1, wherein, Also includes: A feature pyramid network is constructed based on a low-level feature enhancement module and a circular boundary enhancement module, combined with a convolutional network. The feature pyramid network includes two branches, each of which includes the low-level feature enhancement module and the circular boundary enhancement module connected side-by-side by the convolutional network.
4. The domain adaptation crater detection method of claim 3, wherein, The step of using a feature pyramid network to extract features from the task samples and obtain feature maps includes: The task samples are processed by convolution through the convolutional network. The low-level feature enhancement module utilizes an attention mechanism to perform low-level semantic information enhancement processing on the convolutional task samples; The circular boundary enhancement module utilizes diagonal convolution to perform domain-invariant feature enhancement on the convolution-processed task samples; The features obtained by fusing the low-level semantic information enhancement processing and the domain-invariant feature enhancement processing are used to obtain a feature map; The first feature map is obtained by extracting features from source domain samples with known labels, and the second feature map is obtained by extracting features from target domain samples with unknown labels.
5. The domain adaptation crater detection method of claim 4, wherein, The low-level feature enhancement module includes a first branch, a second branch, and a third branch. The step of using the attention mechanism within the low-level feature enhancement module to perform low-level semantic information enhancement processing on the convolutionally processed task samples includes: Based on the task sample after convolution processing, the first feature, the second feature and the third feature are obtained by 1×1 convolution in the first branch, the second branch and the third branch, respectively. Based on the first feature, a fourth feature is obtained by applying Softmax to the first branch; Multiply by the fourth feature and the second feature, and add to the third feature.
6. The domain adaptation crater detection method of claim 4, wherein, The low-level feature enhancement module includes a first branch, a second branch, and a third branch; the circular boundary enhancement module includes a fourth branch and a fifth branch; and the process of performing domain-invariant feature enhancement on the convolutionally processed task samples using the circular boundary enhancement module via diagonal convolution includes: Based on the task sample after convolution processing, the fifth feature is obtained by using shift and subtraction operations with two sets of 1×1 convolutions in the first branch. Based on the task sample after convolution processing, the sixth feature is obtained by performing a 3×3 convolution in the fifth branch; Add the sixth feature to the fifth feature.
7. A domain adaptation crater detection system, comprising: include: The first module is used to acquire task samples; wherein, the task samples include source domain samples with known labels and target domain samples with unknown labels; The second module is used to extract features from the task samples using a feature pyramid network to obtain feature maps. The feature pyramid network includes a low-level feature enhancement module and a circular boundary enhancement module; the feature map includes a first feature map and a second feature map. The third module is used to perform a linear transformation on the feature map and obtain intermediate subspace features through a progressive domain adaptive network. The step of performing a linear transformation on the feature map and obtaining intermediate subspace features through a progressive domain adaptive network includes: A linear transformation is performed on the feature map to obtain a first subspace and a second subspace; wherein the first subspace is obtained by linear transformation of the first feature map, and the second subspace is obtained by linear transformation of the second feature map; Based on the first subspace and the second subspace, multiple intermediate domain features are obtained through the adaptive change of the hyperparameters of the progressive domain adaptive network. By sampling one of the features of the multiple intermediate domain features as an intermediate subspace feature; The fourth module is used to analyze the intermediate subspace features and the second feature map using a region recommendation network to obtain the intermediate bounding box detection results and the target domain bounding box detection results. The fifth module is used to perform similarity weighting on the intermediate bounding box detection results and the target domain bounding box detection results using a weighted instance-level alignment network to obtain the target detection results.
8. A domain-adaptive impact crater detection device, comprising a processor and a memory; The memory is used to store programs; The processor executes the program to implement the method as described in any one of claims 1 to 6.
9. A computer-readable storage medium, characterized in that, The storage medium stores a program that is executed by a processor to implement the method as described in any one of claims 1 to 6.