A method and system for extracting lunar boulders based on feature enhancement and light shading polarity constraint

Abstract

The application relates to a moon surface stone block extraction method and system based on feature enhancement and light illumination dark pole constraint, which comprises the following steps: acquiring a lunar orbit image and pre-processing to obtain an overlapping image block; performing parallel feature extraction to obtain shallow detail features, deep semantic features and reference features; fusing the shallow detail features and the deep semantic features to obtain fused features; superimposing residuals of the fused features and the reference features to obtain first enhanced features; fusing the first enhanced features, local contrast and edge strength to obtain a weight map; performing residual re-calibration to obtain second enhanced features; acquiring a solar azimuth angle, and dividing a candidate target frame into a light-encountering half area and a light-encountering half area; further calculating a normalized light-dark pole score; based on the normalized light-dark pole score, judging whether the candidate target frame is a moon surface stone block or a moon surface impact crater to perform false detection post-processing. Compared with the prior art, the application has the advantages of high target extraction accuracy.

Description

Technical Field

[0001] This invention relates to the field of deep space exploration and robotic vision technology, and in particular to a method and system for extracting lunar rocks based on feature enhancement and light illumination dark polarity constraints. Background Technology

[0002] In lunar landing and exploration missions, rocks are a significant obstacle affecting safe landing, path planning, and surface exploration. Existing rock identification methods are ill-suited to the characteristics of lunar images, such as small rock size, weak grayscale differences, and strong interference from small impact craters. They are prone to failure in detecting small-scale rocks and easily misidentify impact craters, shadow patches, and abrupt texture changes as rocks. For landers, flybys, or rovers, the spatial distribution, size, and density of rocks within the landing area not only affect landing safety but also the selection of subsequent routes and scientific exploration targets. Therefore, automatically and accurately extracting rock obstacles from lunar orbital and landing images is a crucial step in autonomous navigation and safe landing on the lunar surface. Currently, automatic rock identification on the lunar and Martian surfaces typically draws on general target detection or instance segmentation methods. However, general methods face two prominent problems in lunar imagery: (1) lunar rocks typically occupy only a very small number of pixels in the imagery, especially in mid-to-high altitude landing or orbital images, where a large number of rocks are tiny targets that are easily missed; (2) small impact craters, slope shadows, noise patches, and targets truncated at slice edges can generate a large number of false detections that are similar to rocks. Therefore, it is necessary to design specialized methods for detecting tiny rocks and suppressing false detections on the lunar surface to improve the efficiency of lunar obstacle detection.

[0003] The invention disclosed in CN122200141A is a method for detecting lunar rocks in small samples based on multi-scale feature alignment, belonging to the fields of deep space exploration and robot vision. This method aims to solve the detection challenges caused by the scarcity of lunar samples, variable illumination, and domain differences. The scheme includes: acquiring the image to be detected, extracting multi-scale feature maps, and generating illumination-adaptive weights based on the global grayscale mean of the image for feature fusion; fusing visual features with semantic features from structured physical semantic descriptions to generate multimodal features and map them to scores; finally, calculating confidence scores and outputting rock detection results that meet the threshold. This invention effectively improves the detection accuracy and robustness under small sample sizes and complex illumination by constructing a hybrid sample library, illumination-adaptive fusion, cross-domain feature alignment, and multimodal information fusion. However, this scheme does not introduce local texture enhancement and edge-guided features, resulting in poor recognition accuracy under complex illumination conditions. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a method and system for extracting lunar rocks based on feature enhancement and light illumination dark polarity constraint.

[0005] The objective of this invention can be achieved through the following technical solutions: A method for extracting lunar rocks based on feature enhancement and illumination dark polarity constraints includes: Lunar orbit images are acquired and preprocessed to obtain overlapping image blocks; the overlapping image blocks are then input into a pre-constructed lunar image feature enhancement network for parallel feature extraction to obtain shallow detail features, deep semantic features, and reference features. The shallow detail features and deep semantic features are convolutionally projected and upsampled and then fused to obtain a fused feature; the fused feature is residually superimposed with the reference feature to obtain a first enhanced feature; the local contrast and edge intensity corresponding to the overlapping image block are obtained; The first enhanced feature, local contrast, and edge intensity are fused to obtain a weight map; the first enhanced feature is recalibrated using the weight map to obtain a second enhanced feature; prediction is performed based on the second enhanced feature to output candidate target boxes and confidence scores. Obtain the solar azimuth angle and divide the candidate target box into a sunlit half-zone and a shadowed half-zone; obtain the grayscale mean of the sunlit half-zone and the shadowed half-zone, and calculate the normalized brightness polarity score; based on the normalized brightness polarity score, determine whether the candidate target box is a lunar rock or a lunar impact crater for false positive post-processing.

[0006] Furthermore, the formula for calculating the normalized brightness polarity score includes: in, To normalize the light and dark polarity scores, The average gray value of the illuminated half-area. The grayscale average of the backlight half-region. is the numerical stability constant.

[0007] Furthermore, based on the normalized brightness polarity score, the candidate target box is determined to be a lunar rock or a lunar impact crater. The corresponding determination process includes: Obtain the shadow ratios of the illuminated and shaded halves; when the normalized brightness polarity score is greater than a preset threshold and the shadow ratio of the shaded half is greater than that of the illuminated half, determine that the candidate target box is a lunar rock and retain it as a candidate; when the normalized brightness polarity score is less than a preset threshold and the shadow ratio of the illuminated half is greater than that of the shaded half, determine that the candidate target box is a lunar impact crater and remove it as a candidate.

[0008] Furthermore, the fused feature and the reference feature are residually superimposed to obtain the first enhanced feature, and the corresponding fusion expression is: in, As the first enhancement feature, For reference features, For learnable residual coefficients, For 3×3 convolution fusion, This indicates upsampling to a quarter scale. , This represents a 1×1 convolution projection. These are shallow detail features. It represents deep semantic features.

[0009] Furthermore, the first enhanced feature, local contrast, and edge intensity are fused to obtain a weight map, and the corresponding fusion expression is: in, For weighted graphs, This represents the Sigmoid activation function. This represents a 3×3 convolution operation. As the first enhancement feature, These are the local contrast features after feature mapping. The edge intensity features are obtained after feature mapping. , Feature mapping functions representing local contrast and edge intensity, respectively. For local contrast, This represents the edge strength.

[0010] Furthermore, the first enhanced feature is recalibrated using the weighted graph, and the corresponding residual recalibration expression is: in, As the second enhancement feature, As the first enhancement feature, This is a weighted graph.

[0011] Furthermore, the expression for obtaining local contrast is: in, For local contrast, For pixels in a normalized grayscale image grayscale value at that location It is a local mean. It is the numerical stability constant. Standard deviation, , Represents pixel coordinates.

[0012] Furthermore, the process of obtaining edge strength includes: Using the Sobel operator, the horizontal and vertical gradients of the overlapping image blocks are calculated; further, based on the horizontal and vertical gradients, the edge intensity is calculated, and the corresponding calculation formula is as follows: in, For edge strength, For horizontal gradient, For vertical gradient, The horizontal Sobel convolution kernel is used. The Sobel convolution kernel is in the vertical direction. Represents pixel coordinates, The input image.

[0013] Furthermore, the preprocessing process includes sequentially performing image segmentation, grayscale scaling, and normalization on the lunar orbit image.

[0014] The present invention also provides a lunar rock extraction system based on feature enhancement and light illumination dark polarity constraint, including a memory and a processor. The memory stores a computer program, and the processor calls the computer program to execute the steps of the method described above.

[0015] Compared with the prior art, the present invention has the following advantages: (1) This invention constructs a lunar image feature enhancement network to extract features in parallel from overlapping image blocks, extracting shallow detail features, deep semantic features and reference features. The shallow detail features retain the edges, textures and small rock positions, while the deep semantic features provide background landforms and category semantics. The first enhancement feature is then generated through residual superposition, which effectively solves the problems of blurred rock edges and missed detection of small targets under complex lunar illumination. By introducing local contrast, edge intensity and fusing with the first enhancement feature and recalibrating the residual, the feature expression of the unique bright spots, shadows and edge combination patterns of the rocks is further enhanced. By using the solar azimuth angle to divide candidate target boxes into illuminated and shaded half-regions, calculating normalized brightness polarity scores, and eliminating false detections based on the opposite brightness distribution characteristics of rocks and impact craters, the combination of deep learning feature extraction and lunar surface illumination geometry significantly improves the accuracy and robustness of lunar rock extraction.

[0016] (2) By extracting high-resolution enhancement features, further performing local texture enhancement, edge-guided feature fusion, and impact crater false detection suppression based on sunlight direction, this invention realizes the automatic extraction of rock targets in lunar orbit images, which can improve the recall rate of small rocks and reduce the false detection rate of impact craters and shadow patches, providing reliable data support for lunar landing area safety assessment, rover path planning, and lunar surface topography analysis. Attached Figure Description

[0017] Figure 1 This is a flowchart of a lunar rock extraction method based on feature enhancement and illumination dark polarity constraint provided in an embodiment of the present invention; Figure 2 This is a simplified flowchart of a lunar rock extraction method based on feature enhancement and illumination dark polarity constraint provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the illumination principle for false detection removal of impact craters in a lunar rock extraction method based on feature enhancement and illumination dark polarity constraint provided in an embodiment of the present invention. Figure 4 This is a schematic diagram of the baseline network detection results without small target feature enhancement, for a lunar rock extraction method based on feature enhancement and illumination dark polarity constraint provided in an embodiment of the present invention. Figure 5 This is a schematic diagram of the lunar rock extraction method based on feature enhancement and light illumination dark polarity constraint provided in this embodiment of the invention, showing the rock detection results. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0019] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0020] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0021] Definitions: Sobel operator: a classic discrete differential operator used for edge detection in digital image processing. It extracts edge information by calculating the approximate gradients of the image grayscale function in the horizontal and vertical directions.

[0022] Example 1 like Figure 1 and Figure 2 As shown, this embodiment provides a method for extracting lunar rocks based on feature enhancement and illumination dark polarity constraints. The method includes the following steps: S1: Acquire lunar orbit images and preprocess them to obtain overlapping image patches; input the overlapping image patches into a pre-constructed lunar image feature enhancement network for parallel feature extraction to obtain shallow detail features, deep semantic features, and reference features; Preferred, The preprocessing process includes sequentially dividing the lunar orbit image into segments, scaling the grayscale image, and normalizing it.

[0023] S2: After convolutional projection and upsampling of shallow detail features and deep semantic features, the fused features are obtained; the fused features are residually superimposed with the reference features to obtain the first enhanced features; the local contrast and edge intensity corresponding to the overlapping image blocks are obtained. S3: Fuse the first enhanced feature, local contrast, and edge intensity to obtain a weight map; use the weight map to recalibrate the residual of the first enhanced feature to obtain the second enhanced feature; make predictions based on the second enhanced feature and output candidate target boxes and confidence scores; Specifically, The construction process of the lunar image small target feature enhancement network includes: 1. Construction of high-resolution feature layers First, the lunar orbit image is segmented into overlapping image blocks and input into the backbone network, selecting shallower features. F s and deeper features F d . F s Preserve the edges, textures, and the positions of the small stones. F d Provides background topography and category semantics. These are then stitched together after 1×1 convolutional projection, upsampled to one-quarter scale of the input image, and fused using 3×3 convolution; reference features are also included. F ref The features are obtained from the original multi-scale projection module of the basic detection network: Original projection features corresponding to the high-resolution branch are selected; if their spatial size differs from the quarter-scale, upsampling or downsampling alignment is performed, and the number of channels is unified through a 1×1 convolution. Subsequently, the fused features are... F refResidual stacking is performed to obtain enhanced high-resolution features. F P2 : in, φ s , φ d For 1×1 convolutional projection, Up 4 This indicates upsampling to a quarter scale. ψ For 3×3 convolution fusion, λ These are the learnable residual coefficients. F P2 It is not the original shallow features directly output by the backbone network, but the enhanced features after the fusion of shallow details and deep semantics; it is input together with the projection features of the remaining scales of the original detection network into the subsequent Transformer encoder and detection head.

[0024] 2. Fusion of local texture enhancement and edge-guided features For the combination of highlights, shadows, and edges of the stones, in F P2 Set a texture edge enhancement branch. Scale the grayscale image to... F P2 Spatial dimensions and normalization, in k×k Calculate normalized local contrast within the window: in, μ k and σ k These are the local mean and standard deviation, respectively. ε This is used to prevent division by zero. The Sobel operator is then used to calculate the horizontal and vertical gradients and edge strength, and the local contrast and edge strength are then compared with... F P2 Fusion to generate weight graph W te . W te These are modulation coefficients with values ​​ranging from 0 to 1, not detection results; they are recalibrated using residual methods. F P2 Enhanced features are obtained F̃ P2 This feature replaces the original F P2 Proceed to subsequent encoding and detection. The corresponding calculation formula is as follows: 3. Multi-scale candidate box prediction Enhanced high-resolution features F̃ P2 The original multi-scale projection features are input into the encoder and decoder. The detection head outputs the bounding boxes and confidence scores of candidate stones, retaining candidates with confidence scores not lower than the threshold; non-maximum suppression is used for deduplication only when multiple queries correspond to the same target.

[0025] S4: Obtain the solar azimuth angle and divide the candidate target box into a sunlit half-zone and a shadowed half-zone; obtain the grayscale mean of the sunlit half-zone and the shadowed half-zone, and calculate the normalized brightness polarity score; based on the normalized brightness polarity score, determine whether the candidate target box is a lunar rock or a lunar impact crater for false positive post-processing.

[0026] Preferred, Based on the normalized brightness polarity score, the candidate target box is determined to be a lunar rock or a lunar impact crater. The corresponding determination process includes: obtaining the shadow ratio of the illuminated half and the shaded half; when the normalized brightness polarity score is greater than a preset threshold and the shadow ratio of the shaded half is greater than that of the illuminated half, the candidate target box is determined to be a lunar rock and retained; when the normalized brightness polarity score is less than a preset threshold and the shadow ratio of the illuminated half is greater than that of the shaded half, the candidate target box is determined to be a lunar impact crater and removed.

[0027] Specifically, the post-processing for false detection suppression based on illumination direction includes: 1. Calculation of light and dark features of candidate targets A local coordinate system is established for the center of the candidate frame and its neighborhood. The solar azimuth angle is converted to the image coordinate system, and the area is divided into a half-zone facing the sun according to the direction of sunlight. Ω + and backlight half-zone Ω - The mean grayscale value, shadow ratio, and edge gradient of each half of the area were statistically analyzed. Protruding rocks are typically brighter on the side facing the light and have shadows on the side facing away from the light; the order of light and dark in concave impact craters is usually the opposite.

[0028] 2. False detection and removal of impact craters Calculate the normalized brightness polarity score based on the mean gray values ​​of the illuminated and shaded halves: when ρ q ≥ τ p Candidates are retained when the shadow is mainly located in the backlight half-zone; when ρq ≤- τ p Furthermore, candidates are eliminated when the brightness order is reversed; the remaining candidates retain their original detection confidence scores. This step is used as post-processing and does not alter the main structure of the detection network. Figure 3 As shown, along the same lighting direction, the detection image of the impact crater is displayed as dark on the left and bright on the right, while the detection image of the rock is displayed as bright on the left and dark on the right.

[0029] Experiments and Analysis: This experiment selected high-resolution lunar orbital images, and after radiometric correction, map projection, and slicing, constructed a rock dataset. The results of manual interpretation were used as validation data.

[0030] Ablation groups were set up, including baseline network, high-resolution feature enhancement, texture edge enhancement, and illumination false detection suppression. Training and evaluation conditions were kept consistent across groups. Recall and precision for small-scale stones were compared. AP small Boundary positioning error and impact crater false detection rate. (Next) Figure 4 , Figure 5 The qualitative comparison results are presented.

[0031] like Figure 4 and Figure 5 The figures show the detection results of the baseline network without small target feature enhancement and the stone detection results of the proposed method, respectively. As can be seen from the figures, compared to the baseline network, the proposed method detected several new small-sized stone targets; simultaneously, Figure 4 There is a false positive target with an impact crater. After processing using the illumination-dark polarity constraint method of this approach, the false positive target is... Figure 5 The stones were removed from the list. This demonstrates that this method can effectively suppress false detections of some impact craters while enhancing the detection of small-sized stones.

[0032] In summary, this invention proposes an automatic lunar rock extraction method based on small target feature enhancement. It strengthens high-resolution detail representation in the target detection network and eliminates false detections in the result stage based on the opposite light and dark distribution characteristics of the rocks and impact craters, thereby achieving fast, stable and automated extraction of lunar rocks.

[0033] Example 2 This embodiment provides a lunar rock extraction system based on feature enhancement and light illumination dark polarity constraint, including a memory and a processor. The memory stores a computer program, and the processor calls the computer program to execute the steps of the method as described in Embodiment 1.

[0034] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for extracting lunar rocks based on feature enhancement and illumination dark polarity constraints, characterized in that, include: Lunar orbit images are acquired and preprocessed to obtain overlapping image blocks; the overlapping image blocks are then input into a pre-constructed lunar image feature enhancement network for parallel feature extraction to obtain shallow detail features, deep semantic features, and reference features. The shallow detail features and deep semantic features are convolutionally projected and upsampled and then fused to obtain a fused feature; the fused feature is residually superimposed with the reference feature to obtain a first enhanced feature; the local contrast and edge intensity corresponding to the overlapping image block are obtained; The first enhanced feature, local contrast, and edge intensity are fused to obtain a weight map; the first enhanced feature is then recalibrated using the weight map to obtain a second enhanced feature. Based on the second enhanced feature, a prediction is made, and candidate target boxes and confidence scores are output. Obtain the solar azimuth angle and divide the candidate target box into a sunlit half-zone and a shadowed half-zone; obtain the grayscale mean of the sunlit half-zone and the shadowed half-zone, and calculate the normalized brightness polarity score; Based on the normalized brightness polarity score, the candidate target box is determined to be a lunar rock or a lunar impact crater for false detection post-processing.

2. The lunar rock extraction method based on feature enhancement and illumination dark polarity constraint according to claim 1, characterized in that, The formula for calculating the normalized brightness polarity score includes: in, To normalize the light and dark polarity scores, The average gray value of the illuminated half-area. The grayscale average of the backlight half-region. is the numerical stability constant.

3. The lunar rock extraction method based on feature enhancement and illumination dark polarity constraint according to claim 2, characterized in that, Based on the normalized brightness polarity score, the candidate target box is determined to be a lunar rock or a lunar impact crater. The corresponding determination process includes: Obtain the shadow ratios of the illuminated and shaded halves; when the normalized brightness polarity score is greater than a preset threshold and the shadow ratio of the shaded half is greater than that of the illuminated half, determine that the candidate target box is a lunar rock and retain it as a candidate; when the normalized brightness polarity score is less than a preset threshold and the shadow ratio of the illuminated half is greater than that of the shaded half, determine that the candidate target box is a lunar impact crater and remove it as a candidate.

4. The lunar rock extraction method based on feature enhancement and illumination dark polarity constraint according to claim 1, characterized in that, The fused feature is residually superimposed with the reference feature to obtain the first enhanced feature, and the corresponding fusion expression is: in, As the first enhancement feature, For reference features, For learnable residual coefficients, For 3×3 convolution fusion, This indicates upsampling to a quarter scale. , This represents a 1×1 convolution projection. These are shallow detail features. It represents deep semantic features.

5. The lunar rock extraction method based on feature enhancement and illumination dark polarity constraint according to claim 1, characterized in that, The first enhanced feature, local contrast, and edge intensity are fused to obtain a weight map, and the corresponding fusion expression is: in, For weighted graphs, This represents the Sigmoid activation function. This represents a 3×3 convolution operation. As the first enhancement feature, These are the local contrast features after feature mapping. The edge intensity features are obtained after feature mapping. , Feature mapping functions representing local contrast and edge intensity, respectively. For local contrast, This represents the edge strength.

6. The lunar rock extraction method based on feature enhancement and illumination dark polarity constraint according to claim 1, characterized in that, The first enhanced feature is recalibrated using the weighted graph, and the corresponding residual recalibration expression is: in, As the second enhancement feature, As the first enhancement feature, This is a weighted graph.

7. The lunar rock extraction method based on feature enhancement and illumination dark polarity constraint according to claim 1, characterized in that, The expression for obtaining the local contrast is: in, For local contrast, For pixels in a normalized grayscale image grayscale value at that location It is a local mean. It is the numerical stability constant. Standard deviation, , Represents pixel coordinates.

8. The method for extracting lunar rocks based on feature enhancement and illumination dark polarity constraint according to claim 1, characterized in that, The process of obtaining the edge strength includes: Using the Sobel operator, the horizontal and vertical gradients of the overlapping image blocks are calculated; further, based on the horizontal and vertical gradients, the edge intensity is calculated, and the corresponding calculation formula is as follows: in, For edge strength, For horizontal gradient, For vertical gradient, The horizontal Sobel convolution kernel is used. The Sobel convolution kernel is in the vertical direction. Represents pixel coordinates, The input image.

9. The method for extracting lunar rocks based on feature enhancement and illumination dark polarity constraint according to claim 1, characterized in that, The preprocessing process includes sequentially dividing the lunar orbit image into segments, scaling the grayscale image, and normalizing it.

10. A lunar rock extraction system based on feature enhancement and illumination dark polarity constraint, characterized in that, It includes a memory and a processor, the memory storing a computer program, the processor invoking the computer program to perform the steps of the method as described in any one of claims 1 to 9.

Images