Extraterrestrial object optical feature extraction method and system based on multi-stage preprocessing

Abstract

The application discloses an extraterrestrial body optical feature extraction method and system based on multi-stage preprocessing, and belongs to the technical field of extraterrestrial body optical image processing. The application extracts the contour of the celestial body, adjusts the image size to a specified value without changing the original contour proportion, introduces a light robust feature enhancement network for preprocessing of the normalized image in view of the non-uniform light field of the celestial body, and generates uniform and effective features which are robust to rotation and scale change based on an improved ORB feature detection method. The application realizes effective feature extraction for extraterrestrial body optical navigation, and avoids the influence of complex light conditions and deep space background interference. The extracted stable optical features can be used as the input of subsequent processing modules such as attitude solution, relative navigation and state estimation in the extraterrestrial body optical navigation system, and provide reliable visual information support for the probe in the approaching, flying-around and landing stages.

Description

Technical Field

[0001] This invention belongs to the field of extraterrestrial optical image processing technology, and particularly relates to a method and system for extracting extraterrestrial optical features based on multi-stage preprocessing. Background Technology

[0002] In extraterrestrial object exploration, star images acquired by navigation cameras face the dual challenges of complex lighting conditions and deep-space background interference: the response values ​​of feature points in traditional feature extraction algorithms decrease significantly in this region; the low-contrast background in the deep-space environment reduces the distinguishability between feature points and noise, making it difficult for existing methods to achieve effective feature selection; the relative motion between the detector and the target celestial body causes changes in image scale, and the fixed-scale space construction method of the traditional ORB (Oriented FAST and Rotated BRIEF) algorithm cannot meet the needs of multi-scale feature extraction.

[0003] Current technologies mainly rely on a combination of a single preprocessing technique (such as histogram equalization) and the original ORB algorithm. However, under complex lighting conditions, these technologies suffer from uneven feature point distribution and insufficient scale invariance, failing to meet the engineering requirements for high-precision navigation of extraterrestrial objects. For example, the patent "Self-calibration method for prior brightness equalization of small celestial body image illumination" (application number: CN202311725912.0) only considers brightness enhancement of celestial body images under low-light conditions, neglecting complex lighting situations with light-dark boundaries. The patent "Autonomous navigation and guidance control planning and scheduling method for small celestial body impact detection" (application number: CN201010109463.3) uses only traditional ORB features when conducting navigation research using optical characteristics, lacking robustness to rotation and scale changes. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method and system for extracting optical features of extraterrestrial objects based on multi-stage preprocessing. For target celestial navigation star maps captured by a detector at long distances, multi-stage preprocessing techniques combined with an improved feature extraction algorithm effectively avoid interference from deep-space background noise and complex lighting conditions, ultimately obtaining key features with multi-scale feature enhancement and affine invariance. The extracted stable optical features can serve as input to subsequent processing modules in the extraterrestrial object optical navigation system, such as attitude calculation, relative navigation, and state estimation, providing reliable visual information support for the detector during approach, fly-around, and landing phases.

[0005] The first aspect discloses a method for extracting optical features of extraterrestrial objects based on multi-stage preprocessing, the method comprising:

[0006] The input celestial image is preprocessed to obtain a standard-sized celestial image that retains the outline of the celestial body;

[0007] The standard-size celestial image is input into an illumination robust feature enhancement network. The illumination robust feature enhancement network divides the standard-size celestial image into multiple non-overlapping image regions, establishes illumination correlation relationships between each image region through a self-attention mechanism, generates an illumination response weight map based on the established illumination correlation relationships, performs frequency domain enhancement and spatial enhancement based on the illumination response weight map, and fuses and reconstructs the frequency domain enhancement results with the spatial enhancement results to obtain the enhanced preprocessed image.

[0008] Multi-scale feature construction is performed on the enhanced preprocessed image to obtain multiple feature maps at different scales;

[0009] Feature point detection is performed on the multiple feature maps at different scales respectively;

[0010] A corresponding affine matrix is ​​generated based on the feature direction and scale factor of each feature point, and the coordinates of pixels in the neighborhood of the feature point are normalized using the affine matrix.

[0011] High-dimensional vector representation is performed on the normalized image to obtain an optical feature description of the celestial image.

[0012] The second aspect discloses a system for extracting optical features of extraterrestrial objects based on multi-stage preprocessing, the system comprising:

[0013] The standardization processing module is used to preprocess the input celestial images to obtain standard-sized celestial images that retain the outlines of the celestial bodies.

[0014] The image enhancement preprocessing module is used to input the standard-size celestial image into an illumination robust feature enhancement network. The illumination robust feature enhancement network divides the standard-size celestial image into multiple non-overlapping image regions, establishes illumination correlation relationships between each image region through a self-attention mechanism, generates an illumination response weight map based on the established illumination correlation relationships, performs frequency domain enhancement and spatial enhancement based on the illumination response weight map, and fuses and reconstructs the frequency domain enhancement results with the spatial enhancement results to obtain the enhanced preprocessed image.

[0015] A feature map construction module at different scales is used to construct multi-scale features from the enhanced preprocessed image to obtain multiple feature maps at different scales.

[0016] The feature point detection module is used to detect feature points in the multiple feature maps of different scales respectively;

[0017] The normalization processing module is used to generate a corresponding affine matrix based on the feature direction and scale factor of each feature point, and to perform coordinate normalization processing on the pixels in the neighborhood of the feature point through the affine matrix.

[0018] The celestial image optical feature description acquisition module is used to perform high-dimensional vector representation on the normalized image to obtain the celestial image optical feature description.

[0019] As can be seen from the above technical solutions, the present invention has the following beneficial effects:

[0020] Through the above technical solutions, this invention extracts the outline of celestial bodies and adjusts the image size to a specified value while maintaining the original outline proportions. Furthermore, for non-uniform illumination fields of celestial bodies, an illumination-robust feature enhancement network is introduced to preprocess the standardized image. Further, this invention generates uniform and effective feature descriptions robust to rotation and scale changes based on an improved ORB feature detection method. This invention achieves effective feature extraction for extraterrestrial optical navigation, avoiding the influence of complex illumination conditions and deep-space background interference. The extracted stable optical features can serve as input to subsequent processing modules such as attitude calculation, relative navigation, and state estimation in extraterrestrial optical navigation systems, providing reliable visual information support for the detector during approach, fly-around, and landing phases. Attached Figure Description

[0021] Figure 1 The present invention provides an overall flowchart of a method for extracting optical features of extraterrestrial objects based on multi-stage preprocessing.

[0022] Figure 2 This is a schematic diagram of the illumination robust feature enhancement network structure provided by the present invention.

[0023] Figure 3 The above are comparison images of celestial image preprocessing results provided in the embodiments of the present invention, wherein (a) is the original input celestial image, (b) is the standard size celestial image, and (c) is the enhanced celestial image.

[0024] Figure 4 The present invention provides two near-infrared imaging images of Titan, wherein (a) is a first image of Titan based on differential processing of near-infrared and visible light bands and orthographic projection, and (b) is a second image of Titan enhanced in high latitude regions based on different near-infrared band ratios.

[0025] Figure 5 This is a schematic diagram of the feature matching results using the original method.

[0026] Figure 6 This is a schematic diagram of the feature matching results using the present invention.

[0027] Figure 7 This invention provides a schematic diagram of an extraterrestrial optical feature extraction system architecture based on multi-stage preprocessing. Detailed Implementation

[0028] To make the objectives, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Several embodiments of the present invention are shown in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that the disclosure of the present invention will be thorough and complete.

[0029] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a 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 processes, methods, products, or apparatus.

[0030] It should be noted that when an element is referred to as being "fixed to" another element, it can be directly on the other element or there may be an intervening element. When an element is considered to be "connected to" another element, it can be directly connected to the other element or there may be an intervening element. The terms "vertical," "horizontal," "left," "right," "up," "down," and similar expressions used herein are for illustrative purposes only and are not intended to indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention.

[0031] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," "fixing," etc., 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 communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances. The term "and / or" as used herein includes any and all combinations of one or more of the related listed items.

[0032] In one embodiment, this invention discloses a method for extracting optical features of extraterrestrial objects based on multi-stage preprocessing. The method extracts the celestial object contours from the celestial image, adjusting the image size to a specified value while maintaining the original contour proportions. Then, for non-uniform illumination fields of the celestial object, a lightweight illumination-robust feature enhancement network is introduced to preprocess the standardized image. Furthermore, an inter-frame image visual navigation feature matching method based on an improved ORB is used to generate uniform and effective feature descriptions robust to rotation and scale changes. This invention is primarily used for effective feature extraction in extraterrestrial object optical navigation, avoiding the influence of complex illumination conditions and deep-space background interference.

[0033] like Figure 1 As shown, the specific steps include:

[0034] S101. Preprocess the input celestial image to obtain a standard-sized celestial image that retains the outline of the celestial body.

[0035] In one embodiment, the above steps specifically include:

[0036] The input celestial image is binarized to obtain a binarized celestial image;

[0037] A contour detection algorithm based on boundary tracking is used to extract the contours of celestial bodies from binarized celestial images.

[0038] A scaling strategy that preserves the aspect ratio of celestial bodies is used to scale the binarized celestial body image and center it on a black background to obtain a standard-sized celestial body image that retains the celestial body outline.

[0039] Specifically, the outline of the celestial body is extracted, and the image size is adjusted to a fixed size while keeping the original outline proportions unchanged. This image size is then used as the input for the illumination robust feature enhancement network in the subsequent steps.

[0040] Because the captured celestial images contain low-contrast backgrounds from a large area of ​​deep space, the distinguishability of feature points from noise is reduced. First, the input celestial image is Gaussian blurred, utilizing its linear smoothing filtering properties to suppress high-frequency noise while preserving image edge information. Next, threshold segmentation is performed, setting a threshold to divide the celestial image pixels into two classes, achieving initial differentiation between celestial bodies and the background and enhancing image contrast. In this example, a 5×5 Gaussian kernel with a standard deviation of 0 is used for Gaussian blurring. The threshold is set to 30 in the threshold segmentation process.

[0041] Contour detection is performed on binarized celestial images. A contour detection algorithm based on boundary tracking is used to extract contours from the binarized celestial images, with the Suzuki-Abe contour tracking algorithm being the preferred choice. An outer contour retrieval mode (RETR_EXTERNAL) is used to obtain the outermost closed contour of the target. The contour with the largest area is selected as the celestial target contour, and the position of the celestial body is determined by calculating its bounding box. Sub-regions are extracted by expanding pixels outward from the bounding box.

[0042] Furthermore, to remove background interference, a pure black background image of the same size as the sub-region is created. After adjusting the coordinates of the celestial body's outline, the image is filled to form a binary mask. The mask is then compared pixel-by-pixel with the sub-region to preserve the celestial body's features. Finally, the image is scaled using a scaling strategy that maintains the aspect ratio of the celestial body's outline and centered on the black background to output a standard-sized celestial body image. In this example, the standard size is 640×480.

[0043] S102. The standard-size celestial image is input into an illumination robust feature enhancement network. The illumination robust feature enhancement network divides the standard-size celestial image into multiple non-overlapping image regions, establishes illumination correlation relationships between each image region through a self-attention mechanism, generates an illumination response weight map based on the established illumination correlation relationships, performs frequency domain enhancement and spatial enhancement based on the illumination response weight map, and fuses and reconstructs the frequency domain enhancement results with the spatial enhancement results to obtain the enhanced preprocessed image.

[0044] In one embodiment, the above steps specifically include:

[0045] The obtained standard-sized celestial images are divided into multiple non-overlapping image regions and converted into feature vectors of uniform dimension through linear mapping to obtain serialized feature representations.

[0046] The serialized feature representation is input into the Transformer encoder, and the illumination correlation between different image regions is established using the self-attention mechanism.

[0047] The encoding features of the Transformer encoder are input into the illumination sensing module. Illumination response weights of image regions are generated based on the encoding features. The enhancement degree of different image regions is adaptively adjusted to obtain the illumination response weight map.

[0048] The illumination response weight map is input into the Fourier transform residual module and the neural network joint spatial attention module respectively to obtain the frequency domain enhancement features output by the Fourier transform residual module and the spatial enhancement features output by the neural network joint spatial attention module.

[0049] The frequency domain enhancement features and spatial enhancement features are fused together, and the fused features are input into the reconstruction network to generate an enhanced preprocessed image.

[0050] This invention addresses the problems of local grayscale compression and insufficient texture response in optical images of extraterrestrial objects caused by variations in the solar incident angle, differences in regional illumination, and shadow coverage. It constructs a light perception enhancement network for optical navigation feature extraction tasks.

[0051] It should be noted that existing shadow processing networks typically aim to improve the visual quality of images or restore the appearance of shadow-free images, focusing on eliminating the visual differences between shadowed and unshadowed areas. However, the illumination perception enhancement network in this invention does not aim to remove shadow information from celestial images. Instead, it addresses the characteristic of shadowed areas being prevalent in deep space exploration scenarios and containing potential terrain structure information. By learning the illumination variation patterns in the image, it adaptively enhances the shadow-affected areas.

[0052] Specifically, the illumination-aware enhancement network establishes illumination relationships between image regions, identifies low-response areas caused by insufficient illumination, and uses feature reconstruction to improve the local contrast and structural representation of shadow areas. This enhances edges, textures, and grayscale changes that were previously difficult to detect due to brightness compression. Simultaneously, the network preserves the original spatial structure of the celestial surface, avoiding the introduction of artificial textures or geometric alterations caused by over-enhancement. This improves the repeatability and stability of key point detection in subsequent feature detection, description, and matching processes.

[0053] Specifically, the overall structure of the illumination robustness enhancement network is as follows: Figure 2 As shown, it mainly includes a Transformer encoder, a light sensing module, a Fourier transform residual module, and a neural network joint spatial attention module, with the Fourier transform residual module and the neural network joint spatial attention module connected in parallel.

[0054] First, the standard-sized celestial image obtained in step S101 is divided into 32 non-overlapping image regions, each containing local texture, edge, and illumination information. Then, each image region is flattened and converted into a feature vector of uniform dimension through linear mapping, thereby constructing a serialized feature representation suitable for Transformer encoder processing.

[0055] Then, the resulting serialized feature representation is input into the Transformer encoder. The encoder uses a self-attention mechanism to establish illumination relationships between different image regions, and jointly models the illumination distribution features, shadow region features, and surface texture features in the image.

[0056] Since deep-sky object images typically contain large areas of weak texture and low-light shadow areas, traditional local enhancement methods struggle to distinguish between changes in illumination and changes in the actual structure. Therefore, this invention introduces a lighting sensing module after the Transformer encoder, generating regional lighting response weights based on the encoded features to adaptively adjust the enhancement level for different regions.

[0057] In one embodiment, the steps of inputting the encoded features of the Transformer encoder into the illumination sensing module, generating illumination response weights for image regions based on the encoded features, and adaptively adjusting the enhancement level for different image regions to obtain an illumination response weight map specifically include:

[0058] Global average pooling is performed on the encoded features to extract the illumination response information corresponding to different channels;

[0059] The correspondence between illumination response information and regional illumination intensity is established through continuous convolutional mapping;

[0060] After normalization using the Sigmoid function, an illumination response weight map is generated, which is used to characterize the degree to which different image regions are affected by illumination changes.

[0061] Specifically, the illumination sensing module consists of global average pooling, two layers of convolutional mapping, and sigmoid activation. First, global average pooling is applied to the encoded features output by the Transformer to extract the overall illumination response information corresponding to different channels. Then, two consecutive layers of convolutional mapping are used to establish the correspondence between the illumination response information and the regional illumination intensity. Finally, after normalization using the sigmoid function, an illumination response weight map is generated. The illumination response weight map characterizes the degree to which different spatial regions are affected by illumination changes. Regions with larger weights indicate significant shadow suppression or insufficient brightness, requiring focused enhancement to restore the local grayscale dynamic range; regions with smaller weights retain their original structural information, avoiding the introduction of unrealistic information.

[0062] Furthermore, after the Transformer-encoded features are input into the illumination sensing module, they are divided into two parallel processing paths. The first path is the frequency domain enhancement path, which is input into the Fourier Transform Residual (FTR) module. The FTR module transforms the spatial domain features into the frequency domain, and enhances high-frequency texture and edge information that are affected by illumination-induced reduction in response by analyzing the contribution of different frequency components to the image structural information, and uses residual connections to maintain the consistency of the original structure.

[0063] The second path is the spatial structure enhancement path. This path inputs a two-round recurrent neural network with a joint spatial attention module (TWRNN). The TWRNN module dynamically adjusts the importance of different regions through a multi-stage spatial feature update process, enabling the network to focus on effective structural information within and around shadow areas.

[0064] In the TWRNN module, the input features are iteratively updated through four stages: F1, F2, F3 and F4. After completing the initial spatial attention modeling, the output features of the F1 stage continue to be passed to the F2 stage, and are also fed back to the Fourier Transform (FFT) block of the FTR module through cross-module connections to guide feature reweighting and structural adjustment in the frequency domain enhancement process.

[0065] The FFT block performs Fourier transform on the input backbone features, mapping the features from the spatial domain to the frequency domain, and then enhances and reconstructs different frequency components to output enhanced features. This output feature is also divided into two parts: one part serves as the backbone output and enters the subsequent reconstruction network to generate the preprocessed image; the other part serves as auxiliary features fed back into the later stages of the TWRNN (F2 and subsequent modules) to supplement the structural information during the spatial domain attention modeling process.

[0066] Through the aforementioned dual-path input, cross-module feedback, and interactive update mechanism between the frequency domain and spatial domain, joint optimization among Transformer features, frequency domain enhancement features, and spatial attention features is achieved.

[0067] Subsequently, the frequency domain enhancement features output by the FTR module are fused with the spatial enhancement features output by the TWRNN. The fused features are then input into the reconstruction network, which uses a step-by-step coded feature recovery method to restore the fused deep features to the input image size and generate an enhanced preprocessed image.

[0068] Figure 3 The figures show a comparison of the preprocessed celestial images. (a) is the original input celestial image, (b) is the standard-sized celestial image, and (c) is the enhanced celestial image. As can be seen from the figures, the enhanced preprocessed celestial image, while maintaining the original spatial structure, improves the grayscale levels and local contrast in shadow areas, enhances the structural response in weakly textured areas, and improves the feature representation ability in areas affected by non-uniform illumination. This provides a more stable input image for subsequent feature detection, description, and matching.

[0069] S103. Perform multi-scale feature construction on the enhanced preprocessed image to obtain multiple feature maps of different scales.

[0070] To address the image scale variation caused by the relative motion between the detector and the target celestial body, and to improve the stability and consistency of feature point detection under imaging conditions at different distances, a multi-layer image pyramid structure is constructed for multi-scale feature extraction.

[0071] In one embodiment, the step of constructing multi-scale features from the enhanced preprocessed image to obtain multiple feature images of different scales specifically includes:

[0072] Multi-scale features are constructed on the enhanced preprocessed image using an N-layer image pyramid structure. The N-layer image pyramid structure includes N sampling layers from top to bottom. The first layer upsamples the enhanced preprocessed image by a first preset factor. The second to Nth layers downsample the features of the layer above them by a second preset factor. The first preset factor and the second preset factor are reciprocals of each other. N is a positive integer greater than 2.

[0073] Specifically, this example uses a five-layer image pyramid structure, with each layer numbered sequentially from bottom to top. A first preset multiplier and a second preset multiplier are set; the first preset multiplier is used for upsampling, and the second preset multiplier is used for downsampling. In this embodiment, the first preset multiplier is 5 / 3, and the second preset multiplier is 3 / 5. The original input image is first magnified by 5 / 3 to form the first layer, denoted as [layer name missing]. The second layer is a downsampled version of the first layer, denoted as 3 / 5. Similarly, the 5th layer is 3 / 5 downsampled from the 4th layer, denoted as... This multi-layered pyramid structure enables the same spatial structure to participate in the feature detection process at different scales, thereby enhancing the response capability of ORB feature points under scale-varying conditions and avoiding the problems of a small number of feature points and uneven cluster distribution caused by single-scale detection.

[0074] It should be noted that the specific number of layers in the image pyramid structure can be set according to actual needs.

[0075] S104. Perform feature point detection on the multiple feature maps of different scales respectively.

[0076] In one embodiment, the above steps specifically include:

[0077] The feature maps at different scales are divided into multiple image blocks;

[0078] For each candidate pixel in the feature map at different scales, it is determined whether the preset conditions of the feature points are met. The preset conditions of the feature points are that the pixel values ​​of at least 3 points at a preset position in the preset neighborhood are simultaneously greater than the sum of the pixel values ​​of the candidate pixel and the adaptive threshold or less than the difference between the pixel values ​​of the candidate pixel and the adaptive threshold.

[0079] Pixels that meet the preset conditions for feature points are used as feature points.

[0080] Specifically, the FAST algorithm is first used to quickly locate feature points in the image. By comparing the brightness of a pixel with its surrounding neighboring pixels, it is determined whether the point is a corner point, i.e., a feature point. However, the traditional ORB algorithm detects a small number of feature points and is prone to clustering. This invention uses an inter-frame image visual navigation feature matching method based on an improved ORB. The specific process is as follows:

[0081] The feature maps at different scales were divided into 4×4 blocks, resulting in 16 image blocks. Then, a mask was constructed to perform ORB algorithm detection on the image blocks with a step size of one grid, ensuring that the feature points were evenly distributed.

[0082] For candidate pixels Select with A circle with a radius of 3 pixels is drawn around the center, and this circle is used as a preset neighborhood. Only when at least three of the four preset positions (1, 5, 9, and 13) on the circle have pixel values ​​greater than [a certain value] will a search be conducted. or less That is, this point is considered a feature point. Among them, The pixel value of the candidate pixel. This is the pixel detection threshold. It should be noted that setting... To achieve adaptive thresholding, instead of manually setting a threshold, the average gray value of the neighborhood of a candidate pixel in the image is used as the threshold for that point. An adaptive selection method is used to make it applicable to images with different contrast features, as shown in the following formula (1):

[0083] (1)

[0084] in, This indicates the number of pixels within the preset neighborhood of the candidate pixel. The pixel values ​​of each point within a preset neighborhood of the candidate pixel. This represents the average pixel value of all points within the preset neighborhood of the candidate pixel.

[0085] In one embodiment, to avoid feature points clustering at the edges of cave corners, ground depressions, etc., within each image patch, non-maximum suppression is used to remove overly clustered feature points. Specifically, the present invention also includes:

[0086] Traverse all feature points in a feature map of the same size, sort the feature points in the same image block by pixel value, and retain the feature point with the largest pixel value.

[0087] Specifically, the pixels of the entire feature map of the same size are traversed, sorted according to the pixel values ​​of the calculated feature points, and finally only the feature points with the largest pixel values ​​in the image block are retained.

[0088] S105. Generate a corresponding affine matrix based on the feature direction and scale factor of each feature point, and perform coordinate normalization processing on the pixels in the neighborhood of the feature point using the affine matrix.

[0089] Since the keypoints detected by the FAST algorithm only have location information and not directionality, the direction is further determined by calculating the image gray-level centroid of the image block where the feature point is located. First, the image block is calculated... moment The specific formula is shown in formula (2) below:

[0090] (2)

[0091] in, Represents image blocks Mid-pixel coordinates The parameters representing coordinate values ​​in different directions, i.e. hour, , hour, , Represents pixels The pixel values. Image patches can be found using moments. The centroid of the image patch is obtained by connecting the geometric center O with the centroid C, resulting in a direction vector. Therefore, the feature direction of the feature point can be defined by formula (3):

[0092] (3)

[0093] in, and express , Moments in different directions. The scale factor of a feature point is uniquely determined by its pyramid level. Let the pyramid scaling factor, also known as the second preset multiplier, be 3 / 5, then the scale factor corresponding to the k-th level... The definition is shown in the following formula (4):

[0094] (4)

[0095] Among them, the basic scale Defined as 1 (unit scale), used for normalizing scale space representation.

[0096] Furthermore, based on the characteristic direction and scale factor Generate affine matrix The specific formula is shown in formula (5) below:

[0097] (5)

[0098] The affine matrix is ​​used to normalize the coordinates of the image patch in the neighborhood of the feature point, mapping it from the original image coordinate system to the standard reference coordinate system. Specifically, firstly, any pixel in the neighborhood of the feature point... Transform to local coordinates with the feature point as the origin. The specific formula is shown in formula (6) below:

[0099] (6)

[0100] in, This represents the location of the feature point in the image. The pixel is obtained in the local coordinate system through an affine transformation matrix. Normalized coordinates The specific formula is shown in formula (7) below:

[0101] (7)

[0102] The adjusted feature points have the same orientation and other attributes as the original feature points, but their scale is adjusted.

[0103] In one embodiment, to enhance the robustness of the scale factor, the present invention further includes:

[0104] The effective scale factor is obtained by adjusting the scale factor using a scale adjustment coefficient.

[0105] Generate the corresponding affine matrix based on the feature direction and effective scale factor of each feature point.

[0106] Specifically, a scale adjustment coefficient is introduced based on the pyramid scale. The final effective scaling factor is obtained. The specific formula is shown in formula (8) below:

[0107] (8)

[0108] The scale adjustment coefficient in this embodiment Then, based on the feature direction and effective scale factor of each feature point, the corresponding affine matrix is ​​generated.

[0109] S106. Perform high-dimensional vector representation on the normalized image to obtain the optical feature description of the celestial image.

[0110] Specifically, keypoint features are described by a high-dimensional vector consisting entirely of 0s and 1s. The ORB algorithm uses a lookup table when calculating the descriptor, which can retrieve 128 sets of points in the neighborhood of the feature point in the same pattern. Then, it compares the size relationship between two random pixels p and q in each set. If p > q, it takes 1; otherwise, it takes 0.

[0111] To further illustrate the beneficial effects achieved by the present invention, specific embodiments are described below.

[0112] To verify the applicability of this invention in weak-texture scenarios of deep-space planetary moons, real remote sensing image data of Saturn's moon Titan was selected as input data. All image data were derived from publicly available data. In a specific embodiment, two near-infrared imaging products of Titan acquired by the Cassini probe between March and April 2007 were selected as input images. The first image is an orthographic projection image of Titan based on differential processing of the near-infrared and visible light bands, with a center wavelength of approximately 939 nanometers and a spatial resolution of approximately 7 kilometers per pixel. The second image is an enhanced image of the high-latitude region of Titan based on processing with different near-infrared band ratios, with a spatial resolution of approximately 9 kilometers per pixel. Figure 4 As shown, (a) is the first image of Titan based on differential processing of the near-infrared and visible light bands, and (b) is the second image of Titan's high-latitude region enhanced based on different near-infrared band ratios.

[0113] Image preprocessing and feature point detection are performed on the first and second images using both the present invention and the original method. Then, bidirectional robust filtering and matching are performed on the extracted image features to eliminate false matches and ensure geometric consistency. Specifically, a BFMatcher (brute-force matcher) is used to match binary feature descriptors, which correspond to the local feature vector representation of each feature point. The matching criterion is set to Hamming distance, and cross-checking is disabled. The feature descriptors of the first and second images are matched, with each query descriptor returning a positive match result. Similarly, the feature descriptors of the second and first images are matched to obtain negative match results.

[0114] Furthermore, a ratio test and symmetry verification are performed. Ratio test: For each pair of matches in the forward matching results, if the distance of the first match is less than a certain multiple of the distance of the second match, then the match is considered a good match. Symmetry verification: For matches that pass the ratio test, for each forward match, check if a corresponding match exists in the reverse matching to verify the symmetry condition.

[0115] Finally, geometric verification is performed. Matching point pairs between two images are extracted from the high-quality matching list that has undergone bidirectional consistency verification, and corresponding two-dimensional point sets are constructed. The geometric transformation relationship between the two views is estimated based on the Random Sample Consensus (RANSAC) algorithm, which describes the spatial correspondence between the two images. Furthermore, during model estimation, a minimum number of sample points are randomly selected from the matching point pairs to calculate the initial geometric transformation model parameters, and this model is used to verify the consistency of all matching points.

[0116] Specifically, matching consistency is measured by calculating the spatial deviation of each matching point under the transformation model, and a preset reprojection threshold is used as the inlier determination criterion. When the deviation is less than the reprojection threshold, the matching point pair is considered to satisfy geometric consistency. Matches that satisfy geometric consistency are selected as the final valid matching results.

[0117] In this embodiment, the ratio test multiplier is set to 0.75. If the distance of the first match is less than 0.75 times the distance of the second match, it is considered a good match point pair. The reprojection threshold for geometric verification is 3-5 pixels, with a confidence level of not less than 0.99.

[0118] In this embodiment, the feature matching results of the first and second images using the original method show no valid matching point pairs that meet the requirements. By removing the matching verification step, the top twenty closest matching point pairs in the bidirectional matching results are displayed as follows: Figure 5 As shown, it is clear that the original image features resulted in mismatches. The feature matching results in the embodiments of this invention are as follows: Figure 6 As shown in the results, it can be seen that the present invention can achieve feature extraction of distant target celestial bodies and obtain uniform and effective feature matching results.

[0119] This invention also provides an extraterrestrial optical feature extraction system based on multi-stage preprocessing, corresponding to the above-described method embodiments. Since the system embodiments are basically similar to the method embodiments, the description is relatively simple. For details of the relevant technical features and the effects achieved, please refer to the corresponding descriptions of the above-described method embodiments. This invention discloses an extraterrestrial optical feature extraction system based on multi-stage preprocessing, such as... Figure 7 As shown, the system includes:

[0120] The standardization processing module is used to preprocess the input celestial images to obtain standard-sized celestial images that retain the outlines of the celestial bodies.

[0121] The image enhancement preprocessing module is used to input the standard-size celestial image into an illumination robust feature enhancement network. The illumination robust feature enhancement network divides the standard-size celestial image into multiple non-overlapping image regions, establishes illumination correlation relationships between each image region through a self-attention mechanism, generates an illumination response weight map based on the established illumination correlation relationships, performs frequency domain enhancement and spatial enhancement based on the illumination response weight map, and fuses and reconstructs the frequency domain enhancement results with the spatial enhancement results to obtain the enhanced preprocessed image.

[0122] A feature map construction module at different scales is used to construct multi-scale features from the enhanced preprocessed image to obtain multiple feature maps at different scales.

[0123] The feature point detection module is used to detect feature points in the multiple feature maps of different scales respectively;

[0124] The normalization processing module is used to generate a corresponding affine matrix based on the feature direction and scale factor of each feature point, and to perform coordinate normalization processing on the pixels in the neighborhood of the feature point through the affine matrix.

[0125] The celestial image optical feature description acquisition module is used to perform high-dimensional vector representation on the normalized image to obtain the celestial image optical feature description.

[0126] This invention also provides an electronic device, which includes a processor and a memory. The memory stores at least one instruction or at least one program, which is loaded and executed by the processor. This is a method for extracting optical features of extraterrestrial objects based on multi-stage preprocessing, as provided in the above-described method embodiments.

[0127] Furthermore, the electronic device may participate in or include the apparatus or system provided in the embodiments of the present invention. The electronic device may include one or more processors (processors may include, but are not limited to, processing devices such as microprocessors (MCUs) or programmable logic devices (FPGAs), memory for storing data, and transmission devices for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the I / O interface), a network interface, a power supply, and / or a camera.

[0128] It should be noted that the aforementioned one or more processors and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuit may be a single, independent processing module, or may be integrated, in whole or in part, into any other element within the device (or mobile device). As involved in the embodiments of the present invention, the data processing circuit serves as a processor control mechanism (e.g., selection of a variable resistor termination path connected to an interface).

[0129] The memory can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the method described in the embodiments of the present invention. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, thereby realizing the above-mentioned data processing method. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to electronic devices via a network. Examples of the above-mentioned networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0130] The transmission device is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the device's communication provider. In one example, the transmission device includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0131] The display can be, for example, a touchscreen liquid crystal display (LCD), which allows users to interact with the user interface of an electronic device (or mobile device).

[0132] This invention also provides a computer storage medium storing at least one instruction or at least one program, which is loaded and executed by a processor to implement the extraterrestrial optical feature extraction method based on multi-stage preprocessing provided in the above-described method embodiments.

[0133] Optionally, in this embodiment, the aforementioned computer storage medium may be located at at least one of the multiple network servers in a computer network. Optionally, in this embodiment, the aforementioned storage medium may include, but is not limited to, various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0134] This invention also provides a computer program product or computer program, which includes computer instructions stored in a computer storage medium. The processor of an electronic device reads the computer instructions from the computer storage medium and executes the computer instructions, causing the electronic device to perform the extraterrestrial optical feature extraction method based on multi-stage preprocessing provided in the above-described method embodiments.

[0135] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, specific embodiments have been described above. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims can be performed in a different order than that shown in the embodiments and still achieve the desired result. Additionally, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0136] It should be understood that the above description of the preferred embodiments is quite detailed and should not be construed as a limitation on the scope of protection of the present invention. Those skilled in the art can make substitutions or modifications under the guidance of the present invention without departing from the scope of protection of the claims of the present invention, and all such substitutions or modifications fall within the scope of protection of the present invention. The scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for extracting optical features of extraterrestrial objects based on multi-stage preprocessing, characterized in that, The method includes: The input celestial image is preprocessed to obtain a standard-sized celestial image that retains the outline of the celestial body; The standard-size celestial image is input into an illumination robust feature enhancement network. The illumination robust feature enhancement network divides the standard-size celestial image into multiple non-overlapping image regions, establishes illumination correlation relationships between each image region through a self-attention mechanism, generates an illumination response weight map based on the established illumination correlation relationships, performs frequency domain enhancement and spatial enhancement based on the illumination response weight map, and fuses and reconstructs the frequency domain enhancement results with the spatial enhancement results to obtain the enhanced preprocessed image. Multi-scale feature construction is performed on the enhanced preprocessed image to obtain multiple feature maps at different scales; Feature point detection is performed on the multiple feature maps at different scales respectively; A corresponding affine matrix is ​​generated based on the feature direction and scale factor of each feature point, and the coordinates of pixels in the neighborhood of the feature point are normalized using the affine matrix. High-dimensional vector representation is performed on the normalized image to obtain an optical feature description of the celestial image.

2. The method for extracting optical features of extraterrestrial objects based on multi-stage preprocessing according to claim 1, characterized in that, The preprocessing of the input celestial image to obtain a standard-sized celestial image that retains the outline of the celestial body includes: The input celestial image is binarized to obtain a binarized celestial image; A contour detection algorithm based on boundary tracking is used to extract the contours of celestial bodies from binarized celestial images. A scaling strategy that preserves the aspect ratio of celestial bodies is used to scale the binarized celestial body image and center it on a black background to obtain a standard-sized celestial body image that retains the celestial body outline.

3. The method for extracting optical features of extraterrestrial objects based on multi-stage preprocessing according to claim 1, characterized in that, The illumination robust feature enhancement network includes a Transformer encoder, an illumination sensing module, a Fourier transform residual module, and a neural network joint spatial attention module. The Fourier transform residual module and the neural network joint spatial attention module are connected in parallel. The illumination robust feature enhancement network divides the standard-size celestial image into multiple non-overlapping image regions, establishes illumination correlation relationships between these regions through a self-attention mechanism, generates an illumination response weight map based on the established illumination correlation relationships, performs frequency domain enhancement and spatial enhancement based on the illumination response weight map, and fuses and reconstructs the frequency domain enhancement results with the spatial enhancement results to obtain an enhanced preprocessed image, including: The obtained standard-sized celestial images are divided into multiple non-overlapping image regions and converted into feature vectors of uniform dimension through linear mapping to obtain serialized feature representations. The serialized feature representation is input into the Transformer encoder, and the illumination correlation between different image regions is established using the self-attention mechanism. The encoding features of the Transformer encoder are input into the illumination sensing module. Illumination response weights of image regions are generated based on the encoding features. The enhancement degree of different image regions is adaptively adjusted to obtain the illumination response weight map. The illumination response weight map is input into the Fourier transform residual module and the neural network joint spatial attention module respectively to obtain the frequency domain enhancement features output by the Fourier transform residual module and the spatial enhancement features output by the neural network joint spatial attention module. The frequency domain enhancement features and spatial enhancement features are fused together, and the fused features are input into the reconstruction network to generate an enhanced preprocessed image.

4. The method for extracting optical features of extraterrestrial objects based on multi-stage preprocessing according to claim 3, characterized in that, The step of inputting the encoded features of the Transformer encoder into the illumination sensing module, generating illumination response weights for image regions based on the encoded features, adaptively adjusting the enhancement level for different image regions, and obtaining an illumination response weight map includes: Global average pooling is performed on the encoded features to extract the illumination response information corresponding to different channels; The correspondence between illumination response information and regional illumination intensity is established through continuous convolutional mapping; After normalization using the Sigmoid function, an illumination response weight map is generated, which is used to characterize the degree to which different image regions are affected by illumination changes.

5. The method for extracting optical features of extraterrestrial objects based on multi-stage preprocessing according to claim 1, characterized in that, Multi-scale feature construction is performed on the enhanced preprocessed image to obtain multiple feature images of different scales, including: Multi-scale features are constructed on the enhanced preprocessed image using an N-layer image pyramid structure. The N-layer image pyramid structure includes N sampling layers from top to bottom. The first layer upsamples the enhanced preprocessed image by a first preset factor. The second to Nth layers downsample the features of the layer above them by a second preset factor. The first preset factor and the second preset factor are reciprocals of each other. N is a positive integer greater than 2.

6. The method for extracting optical features of extraterrestrial objects based on multi-stage preprocessing according to claim 1, characterized in that, The step of detecting feature points on the multiple feature maps at different scales includes: The feature maps at different scales are divided into multiple image blocks; For each candidate pixel in the feature map at different scales, it is determined whether the preset conditions of the feature points are met. The preset conditions of the feature points are that the pixel values ​​of at least 3 points at a preset position in the preset neighborhood are simultaneously greater than the sum of the pixel values ​​of the candidate pixel and the adaptive threshold or less than the difference between the pixel values ​​of the candidate pixel and the adaptive threshold. Pixels that meet the preset conditions for feature points are used as feature points.

7. The method for extracting optical features of extraterrestrial objects based on multi-stage preprocessing according to claim 6, characterized in that, The method further includes: Traverse all feature points in a feature map of the same size, sort the feature points in the same image block by pixel value, and retain the feature point with the largest pixel value.

8. The method for extracting optical features of extraterrestrial objects based on multi-stage preprocessing according to claim 6, characterized in that, The step of generating a corresponding affine matrix based on the feature direction and scale factor of each feature point, and then normalizing the coordinates of pixels in the neighborhood of the feature point using the affine matrix, includes: The feature orientation is determined by calculating the image gray-level centroid of the image block where the feature point is located. A corresponding affine matrix is ​​generated based on the feature direction and scale factor of each feature point. The scale factor is determined based on the second preset multiple of the multi-layer image pyramid structure and the corresponding level, and the scale factor is different for feature maps of different scales. Transform the pixels within the image patch containing the feature point to a local coordinate system with the feature point as the origin; The coordinates of pixels within the image block containing feature points in the local coordinate system are normalized using an affine matrix.

9. The method for extracting optical features of extraterrestrial objects based on multi-stage preprocessing according to claim 8, characterized in that, The method further includes: The effective scale factor is obtained by adjusting the scale factor using a scale adjustment coefficient. Generate the corresponding affine matrix based on the feature direction and effective scale factor of each feature point.

10. A system for extracting optical features of extraterrestrial objects based on multi-stage preprocessing, characterized in that, The system includes: The standardization processing module is used to preprocess the input celestial images to obtain standard-sized celestial images that retain the outlines of the celestial bodies. The image enhancement preprocessing module is used to input the standard-size celestial image into an illumination robust feature enhancement network. The illumination robust feature enhancement network divides the standard-size celestial image into multiple non-overlapping image regions, establishes illumination correlation relationships between each image region through a self-attention mechanism, generates an illumination response weight map based on the established illumination correlation relationships, performs frequency domain enhancement and spatial enhancement based on the illumination response weight map, and fuses and reconstructs the frequency domain enhancement results with the spatial enhancement results to obtain the enhanced preprocessed image. A feature map construction module at different scales is used to construct multi-scale features from the enhanced preprocessed image to obtain multiple feature maps at different scales. The feature point detection module is used to detect feature points in the multiple feature maps of different scales respectively; The normalization processing module is used to generate a corresponding affine matrix based on the feature direction and scale factor of each feature point, and to perform coordinate normalization processing on the pixels in the neighborhood of the feature point through the affine matrix. The celestial image optical feature description acquisition module is used to perform high-dimensional vector representation on the normalized image to obtain the celestial image optical feature description.

Images