A method for detecting craters based on orbit camera imagery

Abstract

The present application relates to the technical field of lunar orbit image analysis, in particular to a crater detection method based on orbit camera images, a training data set is constructed based on lunar orbit camera images, the training data set contains crater edge heat map labels, offset direction labels and radius labels based on multi-interval coding; a deep learning network is trained using the data set, the network output layer is set to three branches, respectively used for predicting edge heat maps, offset directions and radius multi-intervals; input the image to be detected into the network, obtain the prediction result and filter the effective edge pixel set based on the heat map, decode the direction and radius to generate the crater center candidate point; the mean shift clustering algorithm is used to separate the candidate points into instances, and the crater detection result is obtained, the combination of edge prediction and multi-interval coding improves the detection rate and positioning accuracy of weak edges, small scales and adjacent craters.

Description

Technical Field

[0001] This invention relates to the field of lunar orbit image analysis technology, specifically a method for detecting craters based on orbital camera images. Background Technology

[0002] With the advancement of manned lunar landings and lunar-based scientific research / resource development programs, lunar orbiters are generally equipped with high-resolution optical cameras to perform lunar surface imaging, navigation and positioning, and topographic mapping tasks. Using orbital parameters, attitude information, and camera intrinsic and extrinsic parameters, image coordinates can be converted into three-dimensional lunar surface coordinates, enabling precise positioning of the lander, rover, and subsequent construction equipment. To maintain centimeter-level or even sub-pixel-level positioning accuracy over long periods, it is essential to rely on stable lunar surface features (such as craters) as geometric control points for on-orbit calibration of the orbiter's cameras and related sensors.

[0003] Current crater detection technology has evolved through traditional image processing, machine learning, and deep learning. Early template matching and geometric analysis methods were simple to implement, but had a high false detection rate for small-sized, weakly edged craters. Subsequent introductions of machine learning methods such as SVM and Boosting, while offering improvements, were limited by manual feature design. In recent years, CNN models (such as DeepMoon / U-Net) have achieved end-to-end recognition with significantly improved accuracy, but anchor-based methods involve complex anchor box design. The latest anchor-free methods (such as CenterNet) directly predict the center point and size, making them more suitable for dense, small-scale crater detection, demonstrating advantages in targets below 100 meters in size. Despite these advancements, existing methods still have the following shortcomings:

[0004] Due to limitations in size and power consumption, and the limited optical aperture of cameras, images often suffer from blurred edges due to defocusing, point spread function effects, and noise interference. Existing methods operate directly on degraded images, making it difficult to achieve sub-pixel accuracy.

[0005] Small craters are crucial for lunar surface positioning, but existing anchor-based methods require a large number of anchor boxes, which are computationally complex; anchor-free methods are affected by resolution and lighting conditions, and are prone to missed detections; existing detection methods mostly rely on center points or candidate boxes, which are difficult to directly and accurately align with elevation data (DEM), resulting in insufficient control point accuracy.

[0006] Post-processing is not robust and usually relies on non-maximum suppression (NMS). When craters are densely distributed, targets are prone to mutual interference, which reduces the recall rate. At the same time, it only outputs the center and radius and lacks pixel-level edge information, which leads to a significant decrease in the accuracy of subsequent feature clustering and terrain matching, making it difficult to meet the requirements of high-precision geometric control point extraction.

[0007] In view of this, the present invention provides a method for detecting craters based on orbital camera images. Summary of the Invention

[0008] The purpose of this invention is to provide a crater detection method based on orbital camera images. It aims to combine optical sensing modeling, pixel-level crater segmentation and adaptive clustering to overcome the limitations of small-aperture blurry imaging and provide a sub-pixel-level, illumination- and scale-robust automatic extraction method for lunar surface control points to support the urgent need for lunar orbit calibration and higher-precision positioning and navigation.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] This invention provides a method for detecting craters based on orbital camera imagery, comprising the following steps:

[0011] S101. Extract training dataset based on lunar orbit camera images. The training dataset includes heat map labels of crater edges, offset direction labels of edge pixels relative to the center, and radius labels based on multi-interval coding.

[0012] S102. Train a deep learning network using the training dataset. The network includes at least three output branches, which are used to predict the heat map of the crater edge, predict the offset direction, and predict the radius in multiple intervals, respectively.

[0013] S103. Input the image to be detected and obtain the prediction results of the network. Based on the predicted heatmap, filter the set of effective edge pixels and decode the set by offset direction and radius to generate candidate points for the center of the crater.

[0014] S104. The candidate points are clustered and separated using mean-shift clustering to obtain the corresponding crater detection results.

[0015] As a preferred embodiment of the present invention, the method for labeling craters includes:

[0016] The visible edge arc segments of the crater are labeled using polylines or spline curves along high-contrast regions. During labeling, the line width of the edge polylines is limited, and the centroid of the pixel set of the arc segments is used as the approximate center of the crater.

[0017] Each crater instance records a unique ID, edge polyline coordinate sequence, approximate center, and optional arc segment confidence and curvature statistics.

[0018] As a preferred embodiment of the present invention, the steps for generating the heat map of the crater edge include:

[0019] Establish a normal coordinate system at the marked points on the edge of the crater, and mark the pixel value of the edge line as 1 and the pixel value of the background as 0;

[0020] In the normal coordinate system, a Gaussian function is used to weight the transition region between the edge and the non-edge to obtain the thermal value of the pixel in the buffer. The thermal value of any point gradually decreases as its normal distance to the edge point increases.

[0021] When the same pixel is within multiple normal buffers, the maximum thermal value of the pixel under all normal directions is taken as the final thermal value of the pixel, thereby forming a thermal map with a smooth transition between the boundary and the background.

[0022] As a preferred embodiment of the present invention, the offset encoding method includes:

[0023] The offset of the edge pixels of the crater from its approximate center is represented as direction and radius, where:

[0024] The direction is obtained by performing arctangent calculation on the components of the offset in the x and y directions to obtain the angle, which is encoded as ( ), , ;

[0025] The radius is obtained by calculating the square root of the sum of the squares of the components of the offset in the x and y directions, and the minimum and maximum ranges are determined based on the 5% and 95% quantiles of the radius distribution in the dataset. Within this range, the data is divided into N intervals using a geometric series, and each interval is represented by its geometric mean.

[0026] For each radius value, determine its interval and output the interval index. At the same time, perform residual regression within the interval to limit the residual to a preset range, thereby completing the multibin encoding of the radius.

[0027] As a preferred embodiment of the present invention, the loss function combination for network training includes:

[0028] The segmentation loss used for edge heatmap prediction is measured by weighted cross-entropy or Dice loss to determine the difference between the predicted heatmap and the labeled heatmap.

[0029] The direction loss used for offset direction prediction uses cosine similarity combined with regularization terms to constrain the consistency between the predicted direction vector and the true direction vector, and determines the unit length of the direction vector.

[0030] The classification and residual loss used for radius prediction obtain the prediction probability of each bin through Softmax classification loss and combine it with regression residual loss to improve the accuracy of radius estimation; wherein, the orientation loss and radius loss are calculated within the effective edge region.

[0031] As a preferred embodiment of the present invention, the output layer of the deep learning network consists of three branches: wherein,

[0032] The crater edge heatmap prediction branch consists of several convolutional layers, outputting a single-channel feature map, which is then activated by the Sigmoid function to obtain the crater edge heatmap.

[0033] The offset orientation prediction branch consists of several convolutional layers, and the output contains a feature map with two channels, where the first channel represents the orientation component Dx and the second channel represents the orientation component Dy.

[0034] The offset radius prediction branch consists of several convolutional layers, which output feature maps of 2N channels. The first N channels are processed by the Softmax function to obtain the classification probability of each bin, and the last N channels output the residual code of the corresponding bin, thereby realizing multi-interval prediction of the radius of the crater.

[0035] As a preferred embodiment of the present invention, the post-processing steps include:

[0036] Threshold filtering is performed on the heat map of the crater edge to extract the set of valid edge pixels. Within the set, the direction prediction results are normalized to obtain the corresponding unit direction vector.

[0037] The radius of a pixel is recovered from the bin classification results based on radius prediction and the residual regression value.

[0038] Candidate points for the center of the crater are generated based on the edge pixel coordinates, the radius, and the direction vector.

[0039] The candidate points are clustered using mean-shift clustering, with each cluster center serving as an approximate center of a crater, and the set of edge pixels of its cluster members serving as the boundary of the crater, thus completing the final detection of the crater.

[0040] As a preferred embodiment of the present invention, the decoding process in the post-processing step includes:

[0041] A threshold is set based on the edge heatmap of the crater, and a set of valid edge pixels is selected from it. The pixels in the set are then decoded to reduce noise interference.

[0042] The direction prediction results are normalized to obtain the unit direction vector corresponding to each valid pixel;

[0043] Based on the bin classification results of radius prediction, the bin with the highest probability is selected, and the geometric mean of the corresponding bin is taken. At the same time, the radius of the pixel is recovered by combining the residual regression value.

[0044] Candidate points for the center of the crater are generated by decoding the edge pixel coordinates, the radius, and the unit direction vector.

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] This invention transforms the crater detection problem into a three-branch joint learning approach: edge heatmap prediction, offset direction prediction, and radius multibin prediction. This effectively addresses the challenges of large scale spans, severe edge occlusion, and easy confusion between neighboring targets in lunar orbiter images. Compared to traditional geometric detection methods based on circle / ellipse fitting, this invention maintains high detection rates and accuracy even under conditions of incomplete edges, shadow interference, and terrain noise. Separating the offset direction and radius predictions reduces the impact of long-tailed distributions on regression accuracy, while multibin encoding improves the stability of radius regression under large-scale variations. Using clustering instead of NMS post-processing effectively separates nearby crater instances, reducing false negatives and false positives. The overall method maintains high accuracy on high-resolution images while maintaining high computational efficiency, making it suitable for batch processing and online inference. Attached Figure Description

[0047] Figure 1 This is a rendering of the edge marking method for the crater of the present invention;

[0048] Figure 2 This is the output layer design logic diagram of the present invention;

[0049] Figure 3 This is a flowchart of the crater detection method based on orbital camera images according to the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] In the description of this invention, it should be noted that the terms "vertical," "upper," "lower," "horizontal," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0052] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0053] Example 1

[0054] Please see Figure 3 This invention provides a technical solution: a method for detecting craters based on orbital camera images, comprising the following steps:

[0055] Step S101: Ring mountain labeling, the ring mountain is represented as the boundary line of the ring mountain, the ring mountain is represented as the edge arc segment, and the centroid of the arc segment is used as the approximate center point;

[0056] It should be noted that traditional complete closed boundary annotation is difficult to implement due to issues such as blurred edges, shadow occlusion, and large scale spans in crater imagery. Therefore, an edge arc segment annotation method is proposed, drawing polylines or splines only in high-contrast, clearly visible areas, without forcing the formation of closed contours. In image science and processing, contrast typically refers to the difference in grayscale values ​​between adjacent regions or pixels. High-contrast areas are those with clear edges and significant differences in brightness, such as the boundary between bright and shadow areas at the edge of a crater. For example, the high-contrast area refers to a region where the local pixel grayscale difference is greater than a preset threshold, or a region determined as an edge after edge detection operator processing, used to represent areas with significant changes in brightness and distinct edge features at the edge of a crater.

[0057] When labeling, only label the portion of the crater boundary line with clearly defined edges. The labeling method is as follows: Figure 1 As shown, the labeled objects are the visible edge arcs of the crater. Only areas with obvious edges are labeled; shadowed or low-contrast segments are not forced to close. Polylines or splines are used to draw along the high-contrast edges, with a line width of 1 pixel. The centroid of the arc segment's pixel set is calculated as an approximate center, and this approximate center is defined as the centroid of the arc segment's pixel set, denoted as . Each instance records a unique ID, an edge polyline coordinate sequence, an approximate center, and optional arc segment confidence and curvature statistics.

[0058] Step S102: Generate training data for the crater, construct an edge heatmap, generate a heatmap supervision signal based on the annotation results of step S101, and use multibin encoding to generate offset labels of crater pixels relative to the centroid, wherein the offset is decoupled as direction and radius; solve the problem that the above arc segment annotations are insufficient as training sample information.

[0059] Specifically, the construction principle of the edge heatmap is as follows:

[0060] The crater samples labeled in step S101 are lines, which is insufficient for network training. To reduce the learning difficulty at the intersection of boundary lines and non-boundary lines, a buffer zone between the boundary and non-boundary areas is set. We label the crater boundary line value as 1 and the background value as 0. A coordinate system is established with the boundary normal direction, and point p on the boundary line is the origin of the coordinate system. The value of coordinate r in the normal coordinate system is... for:

[0061]

[0062] in: Indicates the output thermal value; Indicates the edge point (reference point) of the crater; Indicates the normal distance to the edge point (input variable); This represents the buffer control parameters (hyperparameters); This represents an exponential function (decaying form); for a curve, points within the buffer lie on multiple normals, and the ambiguity is eliminated by taking the maximum value.

[0063] Offset label: Offset is the amount of distance a pixel on a crater is offset from its approximate centroid. Offset labels only apply to h> The resolution of a lunar orbiter is around 8K pixels. Taking this as an example, the minimum and maximum radii of detected craters differ by two orders of magnitude. The network's accuracy is not high when directly predicting offsets. Therefore, a multibin bounding box prediction method is used:

[0064] Represent the offset as direction And the radius r, specifically:

[0065]

[0066]

[0067] Direction encoding: Encode the direction as ( ), ,

[0068] Radius encoding: Calculate the 5th and 95th quantiles of all values ​​of r in the entire dataset, denoted as _____. and .

[0069] Let bin number (Recommendations 7–9) Use geometric series to divide:

[0070]

[0071] bin interval , represents the geometric median of the scale. .

[0072] Given Determine the bin index Encode the residual (range) within this bin. ):

[0073]

[0074] The training labels are: class labels residual With direction .

[0075] Step S103: Prediction method, by configuring the output layer of the deep learning network as a three-branch prediction channel: the three-branch prediction channels output the edge heatmap of the crater, the offset direction and the multibin radius respectively;

[0076] Specifically, such as Figure 2 As shown, the network output consists of three branches: the crater edge heatmap prediction branch, the offset direction prediction branch, and the offset radius prediction branch.

[0077] The heatmap prediction branch for the crater edge consists of a c1-layer CNN, which outputs a 1-channel feature map, and then connects to a sigmoid layer to output the heatmap.

[0078] The offset direction prediction branch consists of a c2-layer CNN, outputting a 2-channel feature map. The first channel represents the direction Dx, and the second channel represents the direction Dy.

[0079] The offset radius prediction branch consists of a c3-layer CNN, outputting feature maps with 2N channels. The first N channels are connected to a softmax layer to output the score for each bin, and the last N channels are the encoded residuals for each bin.

[0080] During training, this embodiment constructs a joint loss function, including edge segmentation loss, direction prediction loss (cosine similarity combined with a regularization term), and radius classification and residual regression loss, which is calculated only within the effective edge region to improve the targeting and stability of training. The training method is as follows:

[0081] Section 2 provides training data suitable for the crater prediction algorithm, which is then used to train the network. No restrictions are placed on the hardware or hyperparameters used for network training. Only the design of the loss function is protected; the loss function is as follows:

[0082] Only in the effective edge area Calculate the directional and radius losses.

[0083] The segmentation loss formula is:

[0084]

[0085] The formula for directional loss is:

[0086]

[0087] Add unit length regularity .

[0088] The radius classification and residual formula are as follows:

[0089]

[0090] The total loss formula is:

[0091]

[0092] Step S104: Post-processing method for detection results: Select effective edge pixels based on the heatmap, generate candidate center points by combining direction and radius prediction, and separate crater instances through MeanShift clustering to obtain the final detection result of the crater.

[0093] It should be noted that the predicted output is still at the pixel level and contains noise. Therefore, a threshold-based heatmap is used to extract effective edge pixels. The directional branch output is normalized to obtain a unit direction vector. The radius value is restored using bin classification results and residuals. Center candidate points are generated through voting based on edge points and their corresponding directional radius offsets. Mean-shift clustering is used to cluster the center candidates, with each cluster center serving as an approximate crater center, and the set of edge pixels of the cluster members serving as crater instances. Extreme points are searched along different angles in the heatmap with the cluster center as the axis to extract the crater edge contour, thus achieving the detection of complete craters.

[0094] Specifically, in the post-processing method for detection results, regarding the decoding of predicted data, assume that the deep learning network outputs the following three sets of prediction results:

[0095] Heat map of the crater rim

[0096] Offset direction prediction

[0097] Radius prediction: bin classification score ,satisfy ;Residual within bin

[0098] The decoding steps are as follows:

[0099] Heatmap decoding, effective edge pixel filtering with threshold setting From the heat map Select the set of valid edges:

[0100]

[0101] Only for The pixels in the image are then decoded to reduce noise interference.

[0102] Directional decoding involves normalizing the two channels output by the directional prediction branch.

[0103]

[0104] at this time That is, the predicted unit direction vector.

[0105] Radius decoding determines the bin index in each pixel. Choose the bin with the highest probability:

[0106]

[0107] Take the geometric mean of the bin. :

[0108]

[0109] in , .

[0110] The decoding residual is obtained from the residual channel of the corresponding bin:

[0111]

[0112] Recovery radius:

[0113] Crater center candidate generation: for each valid edge pixel

[0114]

[0115] in and The coordinate units are the same scale (feature map or original image).

[0116] Clustering and instance separation: all Voting is directed to the center point space, and automatic clustering is performed using MeanShift.

[0117] Each cluster center serves as an approximate center for a candidate crater, and the set of edge pixels of the cluster members represents the detection result of that crater. Using the approximate center as an axis, the maximum value in the corresponding heatmap at each angle is found and used as the edge of the crater; thus, the crater detection process is completed.

[0118] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0119] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0120] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art 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 appended claims and their equivalents.

Claims

1. A method for detecting craters based on orbital camera imagery, the method comprising: The method comprises the following steps: ​ S101, extracting a training data set based on a lunar orbit camera image, the training data set containing crater edge thermal map labels, edge pixel relative center offset direction labels, and radius labels based on multi-interval coding; S102, training a deep learning network using the training data set, the deep learning network comprising at least three output branches for predicting a crater edge thermal map, predicting an offset direction, and predicting a radius multi-interval, respectively; S103, inputting an image to be detected and obtaining a prediction result of the deep learning network, screening an effective edge pixel set based on the predicted thermal map, and performing offset direction and radius decoding on the effective edge pixel set to generate a crater center candidate point, wherein, based on a bin classification result of the radius prediction, a bin with the maximum probability is selected, and a geometric mean corresponding to the bin is taken, and a residual error regression value is combined to restore the radius of the pixel; S104, using mean shift clustering to separate the crater center candidate points to obtain corresponding crater detection results.

2. The method of claim 1, wherein: The crater labeling method comprises: annotating the visible edge arc segment of the crater, and using a polyline or a spline curve to depict along a high-contrast region; in the labeling process, the line width of the edge polyline is limited, and the centroid of the arc segment pixel set is taken as the approximate center of the crater, wherein, each crater instance records a unique ID, edge polyline coordinate sequence, approximate center, and optional arc segment confidence and curvature statistics.

3. The method of claim 1, wherein: The generation step of the crater edge thermal map comprises: establishing a normal coordinate system at the crater edge labeling point, and marking the edge line pixel value as 1 and the background pixel value as 0; in the normal coordinate system, a Gaussian function is used to weight the transition region between the edge and the non-edge to obtain the thermal value of the pixel in the buffer region, wherein the thermal value of any point gradually decays as the normal distance to the edge point increases; when the same pixel is within the range of multiple normal buffer regions, the maximum thermal value of the pixel in all normal directions is taken as the final thermal value of the pixel, thereby forming a thermal map with smooth transition between the boundary and the background.

4. The method of claim 1, wherein: The offset amount coding method comprises: the offset amount of the crater edge pixel point to its approximate center is represented as a direction and a radius, wherein: The direction is obtained by an inverse tangent operation on the components of the offset in the x, y directions to obtain an angle, which is encoded as (angle) mod 360 ), , ; the radius is obtained by calculating the square root of the sum of squares of the x and y direction components of the offset amount, and the minimum and maximum ranges are determined based on the 5% and 95% quantiles of the radius distribution in the data set, and the range is divided into N intervals using a geometric progression, and each interval is represented by its geometric mean; for each radius value, determine the interval it belongs to and output the interval index, and perform residual error regression within the interval to limit the residual error to a preset range, thereby completing the multibin coding of the radius.

5. The method of claim 1, wherein: The loss function combination for training the deep learning network comprises: a segmentation loss for edge thermal map prediction, which uses weighted cross-entropy or Dice loss to measure the difference between the predicted thermal map and the labeled thermal map; A direction loss for offset direction prediction, which combines cosine similarity with a regularization term to constrain the consistency between the predicted direction vector and the true direction vector, determines the unit length of the direction vector; A classification and residual loss for radius prediction, which obtains the prediction probability of each bin through a Softmax classification loss, and combines a regression residual loss to improve the accuracy of radius estimation; wherein the direction loss and the radius loss are calculated within the effective edge region.

6. The method of impact crater detection from orbital camera imagery of claim 1, wherein: The output layer of the deep learning network is composed of three branches: wherein, The crater edge heat map prediction branch is composed of several convolution layers, outputs a single-channel feature map, and obtains a crater edge heat map through Sigmoid function activation; The offset direction prediction branch is composed of several convolution layers, and outputs a feature map containing two channels, wherein the first channel represents the direction component Dx, and the second channel represents the direction component Dy; The offset radius prediction branch is composed of several convolution layers, and outputs a feature map of 2N channels, the first N channels are processed by a Softmax function to obtain the classification probability of each bin, and the last N channels output the residual coding corresponding to the bin, thereby realizing multi-interval prediction of the crater radius.

7. The method of impact crater detection from orbital camera imagery as claimed in claim 1, wherein: The post-processing steps include: Threshold screening is performed on the crater edge heat map to extract an effective edge pixel set, and within the effective edge pixel set, the direction prediction result is normalized to obtain the corresponding direction vector; The radius of the pixel is recovered based on the bin classification result and the residual regression value of the radius prediction; The crater center candidate points are generated according to the edge pixel coordinates, the radius and the direction vector; Mean shift clustering is used to cluster the crater center candidate points, and each clustering center is taken as an approximate center of a crater, and the edge pixel set of the clustering members is taken as the boundary of the crater, thereby completing the final detection of the crater.

8. The method of claim 7, wherein: The decoding process in the post-processing step includes: Based on the crater edge heat map, a threshold is set to filter out an effective edge pixel set, and the pixels in the effective edge pixel set are decoded to reduce noise interference; The direction prediction result is unitized to obtain the direction vector corresponding to each effective pixel; The crater center candidate points are generated through the edge pixel coordinates, the radius and the direction vector decoding.

Images