A method and system for detecting and identifying craters in a star catalog

Abstract

The present application designs a star catalog impact crater detection and identification method and system, which comprises the following steps: S1: obtaining the pseudo-height spectral image of the star catalog impact crater; S2: extracting multiple features of the image as multiple pseudo-spectrum images respectively; S3: fusing and enhancing the multiple pseudo-spectrum to generate a PSEF image; S4: using the PSEF image to train a deep learning-based target detection model to obtain a trained impact crater detection model; S5: inputting the star image data to be detected into the impact crater detection model after being made into a PSEF image to obtain a detection result. Through the extraction and fusion of multi-modal features, the present application can effectively extract multiple features of small-scale impact craters under complex terrain and lighting conditions, the generated PSEF image can better adapt to the training process of the deep learning detection model, and the accuracy and robustness of the detection model for detecting small-scale lunar impact craters are improved.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing technology, and in particular relates to a method and system for detecting and identifying impact craters on star surfaces. Background Technology

[0002] Impact craters are among the most important geological features on a planet's surface. By studying impact craters, we can obtain important information about a planet's resources, geological age, and evolutionary history.

[0003] Small impact craters occur far more frequently than large and medium-sized impact craters on the surfaces of the Moon and Mars. Their widespread distribution and large number provide richer impact information, making them important subjects for studying surface processes and resource distribution on the Moon and other planets. Identifying and analyzing the spatial distribution characteristics of small lunar impact craters can reveal impact flux, morphological characteristics, crater age, and degradation processes. Furthermore, with the continuous enrichment of high-precision lunar imagery data, fully identifying small lunar impact craters and analyzing their spatial distribution characteristics has become an urgent need and fundamental guarantee for the study of lunar geological evolution.

[0004] Traditional impact crater detection methods are mostly focused on identifying large-scale impact craters, and their performance is poor when identifying small impact craters on the lunar and Martian surfaces. While some designs exist for identifying small-scale impact craters, current methods still suffer from the following problems:

[0005] (1) Existing methods that rely on pure digital orthophotos are easily affected by changes in illumination angle and terrain. For example, when the illumination angle is unknown and there is a lack of reference objects, it is difficult to directly distinguish between positive and negative terrain from image data. As a result, the deep learning models trained on these models depend on the illumination direction of the training data and have low universality.

[0006] (2) Existing methods that use fused data typically combine digital orthophotos and digital elevation data. However, digital elevation data has a low resolution and often cannot meet the requirements for automatic detection of impact craters measuring tens of meters or even several meters.

[0007] (3) Existing model evaluation methods usually only use data such as accuracy (P), recall (R), F1 score, mAP, etc. as indicators of whether the model predicts correctly. These indicators are all based on the proportion of correct predictions, but lack the representation of the size, shape and position error of the correct prediction box, and cannot further represent the accuracy of the correct prediction results.

[0008] Therefore, there is an urgent need to develop a detection method for small-scale impact craters to address the shortcomings of existing detection methods in terms of accuracy and robustness in detecting small-sized impact craters on the star surface. Summary of the Invention

[0009] The purpose of this invention is to solve the above-mentioned technical problems and provide a method and system for detecting and identifying impact craters on planetary surfaces. Through various feature extraction methods and image enhancement and fusion techniques, it achieves efficient and automatic detection of small-scale impact craters on the surfaces of planets such as the Moon and Mars.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0011] A method for detecting and identifying impact craters on a star surface, characterized by comprising the following steps:

[0012] S1: Acquire image data of impact craters on the planet's surface, divide them into image blocks of a predetermined size, and upsample each image block to obtain a pseudo-hyperspectral image of a predetermined size;

[0013] S2: Based on pseudo-hyperspectral images, multiple pseudo-spectral maps are generated. The specific method for generating multiple pseudo-spectral maps is as follows: extract multiple feature data of the target impact crater, and use them as multiple different pseudo-spectral maps; wherein, the multiple feature data include at least isotropic features, light incident direction features, and intrinsic frequency features; the multiple different pseudo-spectral maps include at least isotropic pseudo-spectrum, light incident direction pseudo-spectrum, and intrinsic frequency pseudo-spectrum;

[0014] S3: Fuse multiple pseudospectral images to generate a pseudomultispectral image, and enhance and fuse the pseudomultispectral image with the pseudohyperspectral image to obtain an enhanced fusion (PSEF) image of the impact crater;

[0015] S4: Construct a deep learning-based target detection model, use PSEF images to train the impact crater detection model, and adjust the model parameters to obtain the trained impact crater detection model.

[0016] S5: After converting the star catalog impact crater image data to be detected into a PSEF image, input it into the trained impact crater detection model to obtain the model detection results.

[0017] In some embodiments of the present invention, the following steps are further included:

[0018] S6: Establish error analysis indicators to judge the accuracy of the detection results of the model. The error indicators should include at least the diameter relative error δ. D , Eccentricity deviation Δe and center point position deviation ΔL;

[0019] Diameter relative error δ D The calculation formula is:

[0020]

[0021] Where D, W, and H are the diameter, width, and height of the bounding box, respectively. The subscript TPi represents the i-th correctly predicted bounding box, and the subscript GTi represents the manually labeled ground truth box corresponding to TPi;

[0022] The formula for calculating the eccentricity deviation Δe is:

[0023]

[0024] Where e is the eccentricity of the bounding box as an ellipse, b is the length of the semi-minor axis of the ellipse, and a is the length of the semi-major axis;

[0025] The formula for calculating the center point position deviation ΔL is:

[0026]

[0027] Where x0 and y0 are the x and y coordinates of the center point, respectively;

[0028] After obtaining the model detection results, the diameter relative error, eccentricity deviation and center point position deviation are calculated based on the manually labeled ground truth boxes and the correctly predicted bounding boxes in the detection results, respectively. This allows for the statistical evaluation of the accuracy of the model prediction boxes, i.e. the edge detection accuracy of the impact crater automatic recognition algorithm.

[0029] In some embodiments of the present invention, the following steps are further included:

[0030] An optimized validation dataset was created based on the obtained PSEF images of the impact craters.

[0031] The data from the optimized validation dataset is input into the trained impact crater detection model. A discrete exhaustive search method is used to optimize the post-processing parameters in the optimized validation dataset to determine the optimal post-processing parameters for the impact crater detection model.

[0032] In some embodiments of the present invention, the post-processing parameters include the NMS confidence threshold and the NMS IoU threshold;

[0033] The specific method for optimizing post-processing parameters is as follows:

[0034] Set the value range for each of the two post-processing parameters;

[0035] Exhaustively search all combinations of post-processing parameters within the range of values ​​of the two post-processing parameters with their respective predetermined step sizes;

[0036] Each set of parameter combinations is assigned to the trained impact crater detection model, and the trained impact crater detection model is used to detect images in the optimized validation dataset.

[0037] Based on the detection results of each parameter combination, the optimal post-processing parameter combination is selected according to the F1 score and used as the post-processing parameter of the trained impact crater detection model.

[0038] In some embodiments of the present invention, step S2 specifically includes the following steps:

[0039] Isotropic pseudospectroscopy was extracted using a Gaussian Laplace filter;

[0040] The Sobel operator was used to extract the pseudospectral of the incident direction of light.

[0041] The intrinsic spatial frequency pseudospectrum is extracted using a spatial frequency domain bandpass filter.

[0042] In some embodiments of the present invention, the specific method for extracting the intrinsic spatial frequency pseudospectrum using a spatial frequency domain bandpass filter is as follows:

[0043] The image is converted to the spatial frequency domain by Fourier transform to obtain a spatial frequency domain image.

[0044] The spatial frequency domain map is processed using a bandpass mask of a predetermined shape that contains only the intrinsic frequencies of the target-scale impact crater, in order to filter out irrelevant frequency information.

[0045] The spatial frequency domain map after bandpass masking is inversely transformed to the spatial domain to obtain the intrinsic spatial frequency features of impact craters of a predetermined scale in the image.

[0046] In some embodiments of the present invention, the pseudo-multispectral image includes three color channels, each of which represents a feature, in order to preserve the differences and complementarities between the features.

[0047] In some embodiments of the present invention, the specific method for enhancing and fusing the pseudo-multispectral image and the pseudo-hyperspectral image in step S3 is as follows:

[0048] The pseudo-multispectral image is converted to the hue-saturation-intensity (HSI) space, the intensity channel is replaced with the grayscale image of the pseudo-hyperspectral image, and then converted back to the RGB space to generate an enhanced fusion (PSEF) image.

[0049] In some embodiments of the present invention, after obtaining the PSEF image of the impact crater in step S3, the following steps are also included;

[0050] The obtained PSEF images were manually annotated and divided into three independent datasets: a training dataset, a training validation dataset, and a test dataset, which were used to train, validate, and test the impact crater detection model, respectively.

[0051] Some embodiments of the present invention further provide a system for detecting and identifying star surface impact craters, comprising:

[0052] At least one processor;

[0053] At least one memory for storing at least one program;

[0054] When at least one program is executed by at least one processor, the at least one processor implements a method for detecting and identifying star catalog impact craters.

[0055] The beneficial effects of this invention are as follows:

[0056] 1. This invention employs pseudospectral spatial feature extraction and enhancement fusion technology to extract multimodal features of star surface impact craters under complex terrain and lighting conditions from multiple perspectives. The PSEF image generated by enhancing and fusing the extracted multimodal features contains rich detail information, which can better adapt to the training process of deep learning target detection models, improve the model's correct prediction ability, edge detection accuracy, and the accuracy and robustness of the model's detection results for small-scale star surface impact craters.

[0057] 2. This invention establishes an error index and introduces three error analysis indices for correctly predicted bounding boxes, thereby enabling the statistical analysis of the recognition accuracy of the impact crater detection model. This overcomes the problem that conventional indices such as precision and recall can only reflect the proportion of correct predictions, but cannot characterize whether the size, shape, and position of the correctly predicted box are accurate.

[0058] 3. This invention optimizes post-processing parameters after model training, which can determine the range of post-processing parameters that perform well, thereby providing guidance for the actual deployment of the model and enabling the model to perform better in the testing phase and in actual deployment. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 A flowchart of a method for detecting and identifying impact craters on a star catalog;

[0061] Figure 2 This is a schematic diagram of the pseudohyperspectral data provided in this embodiment;

[0062] Figure 3 This is a schematic diagram of the isotropic pseudospectral provided in this embodiment;

[0063] Figure 4 This is a schematic diagram of the pseudospectral pattern of the incident light direction provided in this embodiment;

[0064] Figure 5A schematic diagram of the spatial frequency domain diagram containing the bandpass mask and its inverse transform result;

[0065] In the figure, (a), (c), and (e) are the spatial frequency domain diagrams of pseudo-hyperspectral images using different bandpass masks, and (b), (d), and (f) are schematic diagrams of the inverse transform results of (a), (c), and (e), respectively.

[0066] Figure 6 This is a schematic diagram of the intrinsic spatial frequency pseudospectral provided in this embodiment;

[0067] Figure 7 This is a schematic diagram of the pseudomultispectral data provided in this embodiment;

[0068] Figure 8 This is a schematic diagram of the enhanced fusion image provided in this embodiment;

[0069] Figure 9 A heatmap showing the F1 score comparison results of the impact crater detection model and the control group model provided in this embodiment for parameter optimization.

[0070] Wherein, (a) is a schematic diagram of the F1 score results of the impact crater detection model provided in this embodiment for parameter optimization, and (b) is a schematic diagram of the F1 score results of the control group model for parameter optimization. Detailed Implementation

[0071] To make the objectives, technical solutions, and advantages of this application clearer, the application is described and illustrated 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 scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application.

[0072] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations according to this application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. Furthermore, it should be understood that the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or apparatus.

[0073] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0074] The technical solution of the present invention will be described in detail below with reference to specific embodiments and accompanying drawings.

[0075] As attached Figure 1 - Appendix Figure 9 As shown in the illustrative embodiment of a method for detecting and identifying star surface impact craters according to the present invention, the method includes the following steps.

[0076] Step S1: Based on high-resolution remote sensing data from the lunar surface, such as that from the Lunar Reconnaissance Orbiter (LRO), acquire a digital orthophoto image (DOM) of the lunar surface. The acquired DOM typically has high spatial resolution and can clearly display detailed features of the lunar surface. In this embodiment, the DOM image acquired using LRO has a resolution of 1.4 m / pixel.

[0077] The acquired raw digital orthophoto image is sliced ​​into several image blocks of size W×H for subsequent feature extraction and image processing. The size of the slices can be adjusted according to specific needs, and is usually selected to contain multiple impact craters on the order of meters.

[0078] Each image patch after slicing was upsampled using bilinear and bicubic interpolation methods to obtain a single-channel pseudohyperspectral (PHS) image with a size of 4W×4H. The upsampled image is shown in the attached figure. Figure 2 As shown, the purpose of upsampling is to improve the resolution of the image, making the details of the image clearer and facilitating subsequent feature extraction and enhancement. Other applicable upsampling methods can be used for images of star surface impact craters of different scales and under different environments, and this invention does not limit them.

[0079] The obtained PHS image can be denoted as M(x,y), where x and y are the horizontal and vertical coordinates of the pixels in the PHS image, respectively.

[0080] It should be noted that, in order to distinguish it from conventional hyperspectral images directly acquired by optical equipment, the image obtained by upsampling image blocks of digital orthophotos in this invention is defined as a pseudo-hyperspectral image, i.e., a PHS image. The pseudo-spectral images and pseudo-multispectral images provided in the following embodiments are definitions made by this invention to distinguish them from conventional spectral images and multispectral images, and they are all derived from pseudo-hyperspectral images.

[0081] S2: Create various pseudo-spectral maps based on pseudo-hyperspectral images.

[0082] The specific method for creating multiple pseudo-spectral maps is as follows: extract multiple feature data of the target impact crater and use them as multiple different pseudo-spectral maps; wherein, the multiple feature data include at least isotropic features, light incident direction features and intrinsic feature frequency features, and correspondingly, the multiple different pseudo-spectral maps include at least isotropic pseudo-spectrum, light incident direction pseudo-spectrum and intrinsic feature frequency pseudo-spectrum.

[0083] In some embodiments of the present invention, step S2 specifically includes the following steps:

[0084] Isotropic features, or isotropic pseudospectral features, of the image are extracted using a Laplace-Gaussian (LoG) filter. The LoG filter detects zero-crossing points at edge locations by calculating the second derivative of the image, thereby extracting edge features in different directions. Specifically, the LoG filter is applied to the obtained pseudohyperspectral PHS image to extract isotropic features, resulting in a single-channel isotropic pseudospectral image of size 4W×4H, as shown in the attached figure. Figure 3 As shown.

[0085] An isotropic feature image can be denoted as G0(x,y), and its definition is as follows:

[0086]

[0087] Where M is the input pseudo-hyperspectral image, i and j are the indices of the convolution kernel traversal, the values ​​of i and j range from -k to k, k is the radius of the LoG convolution kernel, and the value of k depends on the parameter σ of the Gaussian function.

[0088] The LoG operator is defined as follows:

[0089]

[0090] Where σ is a parameter of the Gaussian function.

[0091] The Sobel operator is used to extract the light incident direction features of an image, i.e., the light incident direction pseudospectral. The Sobel operator is a directional operator that can extract edge features of an image along a specific illumination direction. It should be noted that the illumination incident direction is the same for digital orthophotos of the same region and at the same time. Specifically, the Sobel operator is applied to the PHS image to extract the edge features along the light incident direction, obtaining a feature image of size 4W×4H, as shown in the attached image. Figure 4 As shown.

[0092] The pseudospectral diagram of the incident light direction is denoted as B0(x,y), and is defined as follows:

[0093]

[0094] Where S is the Sobel operator operation, g represents the directional gradient, θ is the angle between the incident light direction and the X direction in the image, and x and y are the X and Y directions, respectively.

[0095] The intrinsic spatial frequency features, or intrinsic spatial frequency pseudospectral, of impact craters at the target scale in an image are extracted using a spatial frequency domain bandpass filter. A specific type of impact crater at a particular scale has its own specific intrinsic spatial frequencies, which are extracted using the following method:

[0096] The image is transformed to the spatial frequency domain using Fourier transform, resulting in a spatial frequency domain map. The center of the spatial frequency domain map corresponds to the low-frequency information of the spatial domain image, while the periphery corresponds to the high-frequency information. The larger the scale of the impact crater, the closer its main information, i.e., its intrinsic spatial frequency, is to the center of the spatial frequency map, and vice versa.

[0097] A bandpass mask of a predetermined shape, containing only the intrinsic frequencies of the target-scale impact crater, is applied to process the spatial frequency domain map. A bandpass filter is then used to retain the intrinsic frequency information of the target impact crater while filtering out irrelevant frequency information. The shape of the bandpass mask can be specifically designed according to the size of the impact crater to be extracted. In this embodiment, as shown in the attached figure… Figure 5 As shown in (e), the bandpass mask is in the shape of concentric double rings, which can be used to extract impact craters on the scale of several meters. The mask near the center can extract information about larger impact craters. Figure 5 This is a schematic diagram showing the effects of different bandpass masks on the spatial frequency domain map of a PHS image and the corresponding inverse transform results. Sub-images (a), (c), and (e) are obtained by inverse Fourier transform to sub-images (b), (d), and (f), respectively.

[0098] Finally, the spatial frequency domain map after bandpass masking is inversely transformed to the spatial domain to obtain a single-channel feature image with a size of 4W×4H, as shown in the attached figure. Figure 6 As shown.

[0099] The spatial frequency domain map transformed from the pseudohyperspectral map by the Discrete Fourier Transform (DFT) is denoted as F(u,v), and is defined by the following formula:

[0100]

[0101] Where (u,v) are the coordinates of a point on the spatial frequency domain plot, W img and H img These represent the width and height of the pseudo-hyperspectral image, respectively.

[0102] The bandpass mask U(u,v) is defined as follows:

[0103]

[0104] Where A is the region on the spatial domain map where the intrinsic frequency information of the target impact crater is located, for example, the concentric double rings in this embodiment.

[0105] Convolving the mask U with F and then performing an inverse transform yields the pseudospectral R0(x,y) of the intrinsic frequency of the impact crater at the predetermined target scale, which is defined as follows:

[0106]

[0107] This invention generates multiple pseudospectral representations of impact crater images using a multimodal feature extraction method. This effectively captures various features of small-scale impact craters, thus alleviating the problem of insufficient feature extraction capabilities in traditional methods. The Sobel filter focuses on edge information along the incident light direction; the LoG filter, as a second-order isotropic differential operator, extracts edges that can supplement directional edges; and the spatial frequency domain bandpass filter, due to the intrinsic spatial frequency characteristics of the impact crater, is more sensitive to the crater's size, emphasizing scale information. These three feature maps are considered as three pseudospectral representations, describing the characteristics of the target impact crater from multiple different dimensions. When they are combined into a three-channel pseudomultispectral (PMS) image, they complement each other, improving the richness and accuracy of feature extraction, and consequently improving the precision and recall of impact crater detection.

[0108] S3: The three feature data are superimposed and fused to generate a pseudomultispectral (PMS) image, denoted as [R0 G0 B0]. T As attached Figure 7 As shown, the generated pseudo-multispectral image contains various extracted feature information, with a size of 4W×4H, including three color channels. Each channel represents a different feature; for example, the red channel represents the isotropic pseudospectrum, the green channel represents the pseudospectrum along the incident light direction, and the blue channel represents the intrinsic frequency pseudospectrum of the target-scale impact crater. The generated pseudo-multispectral image also preserves the differences and complementarities between the features contained in each pseudospectrum.

[0109] Panchromatic sharpening is used to enhance and fuse pseudo-multispectral and pseudo-hyperspectral images to obtain a crater enhanced fusion (PSEF) image. Panchromatic sharpening is an image fusion method that can fuse low-resolution multispectral images with high-resolution panchromatic images to enhance the spatial resolution and detail of the image.

[0110] This invention generates PSEF images containing rich detail information through pseudospectral spatial feature analysis and enhanced fusion technology. The generated PSEF images can better adapt to and improve the detection performance of deep learning-based detection models, thereby helping to improve the mean accuracy (mAP) and F1 score of the subsequently applied deep learning-based detection models.

[0111] Furthermore, compared to the original grayscale image, the PSEF image, as a color image, has different modal information in its three channels. Most existing deep learning-based object detection networks are developed for color images. The PSEF image generated in this invention can better adapt to existing object detection network models and has good adaptability.

[0112] In some embodiments of the present invention, the specific implementation method for enhancing and fusing pseudo-multispectral images and pseudo-hyperspectral images using full-color sharpening technology is as follows:

[0113] The pseudomultispectral image was converted to the hue-saturation-intensity (HSI) space to obtain [I0t]. 10 t 20 ] T :

[0114]

[0115] Among them, t 10 t 20 The relationship between hue (H) and saturation (S) is as follows:

[0116]

[0117] After replacing the intensity (I0) channel with a grayscale image of the pseudo-hyperspectral image, it is converted back to RGB space to generate a three-channel enhanced fusion (PSEF) image of size 4W×4H based on pseudospectral spatial feature extraction and fusion, as shown in the attached image. Figure 8 As shown, the PSEF image is denoted as [R]. PSEF G PSEF B PSEF ] T It is expressed by the following formula:

[0118]

[0119] Pancolor sharpening is commonly used in remote sensing for multispectral and hyperspectral fusion, resulting in an image that possesses both high hyperspectral resolution and visualizes color information from multiple spectra. This invention introduces pancolor sharpening into pure image data derived from monochrome. The monochrome image data is upsampled to obtain pseudo-hyperspectral data, and then subjected to a series of feature extractions to obtain pseudo-multispectral data. Finally, pancolor sharpening is applied in the pseudospectral space to obtain an enhanced fused image. This image emphasizes the extracted details while preserving the global information of the original image data, improving the accuracy of impact crater detection.

[0120] In some embodiments of the present invention, after obtaining the enhanced fused image of the impact crater in step S3, the following steps are further included.

[0121] The obtained enhanced fused images were manually labeled and divided into four independent datasets: a training dataset, a training validation dataset, an optimization validation dataset, and a test dataset. The training and training validation datasets were used to learn the weights of the target detection framework during the training of the impact crater detection model. The optimization validation dataset was used to optimize the post-processing parameters of the trained impact crater detection model after training was completed. The test dataset was used to test and evaluate the impact crater detection model.

[0122] S4: Construct a deep learning-based target detection model. Use the training dataset, training validation dataset, and test dataset to train, validate, and test the model in sequence. Adjust the model parameters to obtain the trained impact crater detection model, abbreviated as PSEF model.

[0123] The following uses the Basal model as a control group model and compares it with the PSEF model provided in this invention to illustrate the beneficial effects of the PSEF model.

[0124] The PSEF model is an impact crater detection model trained using enhanced fused images.

[0125] The Basal model is a target detection model that uses a 4W×4H three-channel pseudo-hyperspectral image for training, validation, tuning, and testing. The remaining steps are the same as those of the PSEF model.

[0126] Compared to the Basal model, the PSEF model shows improvements in both accuracy (P) and recall (R). In a specific embodiment, as shown in Table 1, during validation using data from the validation dataset, the impact crater detection model provided by this invention achieved an accuracy (P) of 0.943, a recall (R) of 0.891, mAP@50 of 0.937, mAP@50-95 of 0.658, and an F1 score of 0.916, all of which are improvements over the Basal model. During testing with the test dataset, under optimal post-processing parameter configuration, the impact crater detection model provided by this invention achieved an accuracy (P) of 0.969, an R of 0.932, and an F1 score of 0.950, all of which are improvements over the Basal model.

[0127] Table 1 PSEF Model and Basal Model

[0128]

[0129] This embodiment combines pseudospectral spatial feature extraction and feature enhancement fusion techniques to extract multimodal features of impact craters on the star surface under complex terrain and lighting conditions from multiple perspectives, improving the accuracy of edge detection. The PSEF image generated by fusing the extracted multimodal features contains rich detail information, which can better adapt to the training process of deep learning-based target detection models, improving the model's detection performance, accuracy, and robustness in detecting small-scale impact craters on the star surface; in particular, the accuracy of the detection results for impact craters smaller than 10 pixels (corresponding to a scale of less than 14 meters on the star surface in this embodiment) is significantly improved.

[0130] In some embodiments of the present invention, in order to enable the model to perform better during the testing phase and in actual deployment, the following steps are further included.

[0131] An independent, optimized validation dataset was created based on the enhanced fused images of the obtained impact craters.

[0132] The data from the optimized validation dataset is input into the trained impact crater detection model. A discrete exhaustive search method is used to optimize the post-processing parameters in the optimized validation dataset to determine the optimal post-processing parameters for the impact crater detection model.

[0133] In a specific embodiment of the present invention, the post-processing parameters are optimized using a YOLOv5-based impact crater detection model as an example. The post-processing mainly includes non-maximum suppression (NMS). The NMS process can reduce duplicate prediction boxes for the same target. The post-processing parameters include the NMS confidence threshold (σ) and the NMS IoU threshold (τ).

[0134] The specific method for optimizing post-processing parameters is as follows:

[0135] Set the value ranges for the two post-processing parameters respectively. The NMS confidence threshold (σ) ranges from 0 to 0.7 with a step size of 0.1; the NMS IoU threshold (τ) ranges from 0 to 0.4 with a step size of 0.1.

[0136] Within the value range of the two post-processing parameters, all combinations of post-processing parameters are exhaustively enumerated with predetermined step sizes for each parameter. In this embodiment, based on the above parameter range and step size, a total of 40 combinations of post-processing parameters are set.

[0137] Each set of parameter combinations is assigned to the trained impact crater detection model, and the trained impact crater detection model is used to detect images in the optimized validation dataset.

[0138] Based on the detection results of each parameter combination, the range of post-processing parameter combinations that can balance accuracy (P) and recall (R) is selected according to the F1 score. In subsequent processing, the optimal post-processing parameter combination can be selected within this range and used as the post-processing parameter of the trained impact crater detection model.

[0139] It should be noted that a larger range indicates better robustness of the model to post-processing parameters. Post-processing parameters are crucial in the practical deployment of the model. Due to the dense distribution and large number of small-scale impact craters, it is difficult to label a large number of targets to select the optimal post-processing parameters in practical applications. Traditional deep learning-based automatic impact crater detection models typically only focus on the design and training of the network framework, lacking attention to post-processing parameters. This invention incorporates a discrete enumeration search optimization process for post-processing parameters during model training, which can determine the range of post-processing parameters that perform well, thus providing guidance for practical model applications. Furthermore, the model trained on PSEF imagery exhibits good robustness to post-processing parameters, meaning it still performs well with a wider range of post-processing parameter selections, which is beneficial for the practical deployment of the model.

[0140] The specific process of using the trained impact crater detection model to detect images in the optimized validation dataset is as follows: For all predicted bounding boxes in the detection results with a confidence level greater than the predetermined NMS confidence threshold (σ), sort them from high to low confidence; then calculate the intersection-union ratio (IoU) between the current predicted bounding box and the subsequent predicted bounding boxes one by one. If the IoU is greater than the NMS IoU threshold (τ), then exclude the predicted bounding boxes with lower confidence.

[0141] The F1 scores of the impact crater detection model after parameter adjustment and the control group before parameter adjustment are compared in the appendix. Figure 9 As shown. (From the appendix) Figure 9It can be seen that, in a total of 40 sets of data with a step size of 0.1, an NMS confidence threshold (σ) ranging from 0 to 0.7, and an NMS IoU threshold (τ) ranging from 0 to 0.4, the impact crater detection model had an F1 score higher than 0.9 in 19 sets, while only 11 sets in the control group had a score higher than 0.9. This indicates that the impact crater detection model provided by this invention has better robustness to post-processing parameters. (See Appendix) Figure 9 In the image, the content selected by the rectangle is the F1 score result of the model using the optimal combination of post-processing parameters.

[0142] S5: The image data of the star catalog impact craters to be detected is processed into an enhanced fused image and then input into the trained impact crater detection model to obtain the model detection results.

[0143] Since small-scale impact craters are densely distributed and numerous, simply calculating the proportion of correctly predicted craters is insufficient to fully evaluate the detection performance of the model. Therefore, it is necessary to further evaluate the error between the correct prediction and the true value. In some embodiments of the present invention, the following steps are also included.

[0144] S6: Establish error analysis indicators to judge the accuracy of the detection results of the model, and realize error analysis of correct predictions.

[0145] In the impact crater detection problem, the target's bounding box is determined by four degrees of freedom: x-coordinate, y-coordinate, lateral width, and longitudinal width. Therefore, in this embodiment, the error index includes at least the diameter relative error δ. Di , Eccentricity deviation Δe and center point position deviation ΔL.

[0146] After obtaining the model detection results, the relative diameter error, eccentricity deviation, and center point position deviation are calculated based on the manually labeled ground truth boxes and the predicted bounding boxes in the detection results, respectively, to determine the accuracy of the model detection results. Specifically, the smaller the values ​​of the three error indicators, the more accurately the impact crater is identified.

[0147] Among them, the relative error of diameter δ Di The calculation formula is:

[0148]

[0149] Where D, W, and H are the diameter, width, and height of the bounding box, respectively. The subscript TPi represents the i-th correctly predicted bounding box, and the subscript GTi represents the manually labeled ground truth box corresponding to TPi.

[0150] The formula for calculating the eccentricity deviation Δe is:

[0151]

[0152] Where e is the eccentricity of the bounding box as an ellipse, b is the length of the semi-minor axis of the ellipse, and a is the length of the semi-major axis.

[0153] The formula for calculating the center point position deviation ΔL is:

[0154]

[0155] Where x0 and y0 are the x and y coordinates of the center point, respectively.

[0156] Traditional model performance evaluations often only include metrics such as P, R, mAP, and F1 scores, which reflect the proportion of correct predictions, but lack analysis of the accuracy of the predicted bounding box positions. For small-scale objects, even a few pixels of deviation can significantly reduce the Intersection over Union (IoU). Therefore, the accuracy of the bounding box is extremely important for the detection of small-scale impact craters. For example, two squares with sides of 2 pixels, if their widths and heights are offset by 1 pixel, will have an IoU of 25%; while an impact crater with sides of 100 pixels, if its widths and heights are offset by 1 pixel, will have an IoU of approximately 96%. This invention establishes error metrics and introduces three predicted bounding box error analysis metrics to statistically analyze the recognition accuracy of the impact crater detection model.

[0157] Some embodiments of the present invention further provide a system for detecting and identifying star surface impact craters, comprising:

[0158] At least one processor;

[0159] At least one memory for storing at least one program;

[0160] When at least one program is executed by at least one processor, the at least one processor implements a method for detecting and identifying star catalog impact craters.

[0161] Finally, it should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0162] The above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them; although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can still be made to the specific implementation of the present invention or equivalent substitutions can be made to some technical features without departing from the spirit of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the technical solutions claimed in the present invention.

Claims

1. A method for detecting and identifying impact craters on a star catalog, characterized in that, Includes the following steps: S1: Acquire image data of impact craters on the planet's surface, divide them into image blocks of a predetermined size, and upsample each image block to obtain a pseudo-hyperspectral image of a predetermined size; S2: Based on the pseudo-hyperspectral image, extract multiple feature data of impact craters of a predetermined scale in the image, and use them as multiple different pseudo-spectral images; the multiple feature data include at least isotropic features, light incident direction features, and intrinsic feature frequency features; The various pseudospectral patterns include at least isotropic pseudospectrals, light incident direction pseudospectrals, and intrinsic characteristic frequency pseudospectrals; S3: Superimpose and fuse multiple different pseudospectral images to generate a pseudomultispectral image, and enhance and fuse the pseudomultispectral image with the pseudohyperspectral image to obtain an enhanced fusion (PSEF) image of the impact crater. S4: Construct a deep learning-based target detection model, use the PSEF image to train the impact crater detection model, and adjust the model parameters to obtain the trained impact crater detection model. S5: After converting the star catalog impact crater image data to be detected into a PSEF image, input it into the trained impact crater detection model to obtain the model detection results.

2. The method for detecting and identifying impact craters on a star surface according to claim 1, characterized in that, Further steps include: S6: Establish an error analysis index to judge the accuracy of the detection results of the model, wherein the error index includes at least the diameter relative error δ. D , Eccentricity deviation Δe and center point position deviation ΔL; The diameter relative error δ D The calculation formula is: Where D, W, and H are the diameter, width, and height of the bounding box, respectively; the subscript TPi is the i-th correctly predicted bounding box, and the subscript GTi is the manually labeled ground truth box corresponding to TPi; The formula for calculating the eccentricity deviation Δe is: Where e is the eccentricity of treating the bounding box as an ellipse; b is the length of the semi-minor axis of the ellipse; and a is the length of the semi-major axis. The formula for calculating the center point position deviation ΔL is: Where x0 and y0 are the x and y coordinates of the center point, respectively; After obtaining the model detection results, the diameter relative error, eccentricity deviation and center point position deviation are calculated based on the manually annotated ground truth boxes and the correctly predicted bounding boxes in the detection results. The accuracy of the model's detection results is analyzed based on the calculation results.

3. The method for detecting and identifying impact craters on a star surface according to claim 1, characterized in that, Further steps include: An optimized validation dataset was created based on the obtained PSEF images of the impact craters. The data from the optimized validation dataset is input into the trained impact crater detection model, and the post-processing parameters are optimized in the optimized validation dataset using a discrete exhaustive search method to determine the optimal post-processing parameters for the impact crater detection model.

4. The method for detecting and identifying impact craters on a star surface according to claim 3, characterized in that, The post-processing parameters include the NMS confidence threshold and the NMS IoU threshold; The specific method for optimizing post-processing parameters is as follows: Set the value range for each of the two post-processing parameters; Exhaustively search all combinations of post-processing parameters within the range of values ​​of the two post-processing parameters with their respective predetermined step sizes; Each set of parameter combinations is assigned to the trained impact crater detection model, and the trained impact crater detection model is used to detect images in the optimized verification dataset. Based on the detection results of each parameter combination, the optimal post-processing parameter combination is selected according to the F1 score and used as the post-processing parameter of the trained impact crater detection model.

5. The method for detecting and identifying impact craters on a star surface according to claim 1, characterized in that, Step S2 specifically includes the following steps: Isotropic pseudospectroscopy was extracted using a Gaussian Laplace filter; The pseudospectral of the incident light direction was extracted using a Sobel filter; The intrinsic spatial frequency pseudospectrum is extracted using a spatial frequency domain bandpass filter.

6. The method for detecting and identifying impact craters on a star surface according to claim 5, characterized in that, The specific method for extracting the intrinsic spatial frequency pseudospectrum of an image using a spatial frequency domain bandpass filter is as follows: The image is converted to the spatial frequency domain by Fourier transform to obtain a spatial frequency domain image. The spatial frequency domain map is processed using a bandpass mask of a predetermined shape that contains only the intrinsic frequencies of the target-scale impact crater to filter out irrelevant frequency information. The spatial frequency domain map after bandpass masking is inversely transformed to the spatial domain to obtain the intrinsic spatial frequency pseudospectrum of the impact crater of a predetermined scale.

7. The method for detecting and identifying impact craters on a star surface according to any one of claims 1, characterized in that, The pseudo-multispectral image includes three color channels, each representing a pseudospectral representation of a feature, to preserve the differences and complementarity between the features.

8. The method for detecting and identifying impact craters on a star surface according to claim 1 or 7, characterized in that, The specific method for enhancing and fusing the pseudo-multispectral image and the pseudo-hyperspectral image in step S3 is as follows: The pseudo-multispectral image is converted to hue-saturation-intensity space, and the intensity channel is replaced with the grayscale image of the pseudo-hyperspectral image. Then, it is converted back to RGB space to obtain the enhanced fusion (PSEF) image.

9. The method for detecting and identifying impact craters on a star surface according to claim 1, characterized in that, After obtaining the PSEF image of the impact crater in step S3, the following steps are also included; The obtained PSEF images are manually annotated and divided into independent training datasets, training validation datasets, and test datasets, which are used to train, validate, and test the impact crater detection model, respectively.

10. A system for detecting and identifying impact craters on a star surface, characterized in that, At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the star catalog impact crater detection and identification method as described in any one of claims 1-9.

Images