A star map recognition method, device, and related medium based on neural networks.

Abstract

This invention discloses a star map recognition method, apparatus, and related medium based on a neural network. The method includes: selecting a primary star in a star map training image; constructing a coordinate feature vector using a Log-Polar algorithm to obtain a first brightness value of the primary star and second brightness values ​​of other neighboring stars; setting a relative brightness factor based on the first and second brightness values, mapping the relative brightness factor according to a preset brightness threshold to obtain a brightness feature vector; constructing a stellar feature vector set based on the coordinate feature vector and the brightness feature vector, and inputting it into a neural network for training and learning; the neural network outputs the primary star number to construct a star map recognition model; and using the star map recognition model to predict the number of target stars. This invention constructs a coordinate feature vector using the Log-Polar algorithm, then uses the relative brightness factor to map and obtain a brightness feature vector, and uses this to train and learn a neural network to construct a star map recognition model to predict the number of target stars, thus improving the recognition accuracy and speed.

Description

Technical Field

[0001] This invention relates to the field of star map recognition technology, and in particular to a star map recognition method, device and related medium based on neural networks. Background Technology

[0002] A star sensor is an autonomous, high-precision astronomical navigation device widely used in aerospace and maritime fields to provide attitude or position information for a vehicle. The operation of a star sensor presupposes the prior star image it captures. Star image recognition identifies the stars in the sky corresponding to the stars in the image based on their relative positions, thus enabling subsequent attitude or position calculations. However, existing star image recognition methods generally have some drawbacks. For example, when the number of stars in the image is small, the accuracy of existing methods is low. Furthermore, existing methods typically require building a large navigation star identification database and matching the identified data with this database, a cumbersome process. The large memory requirements of both the star image recognition method and the navigation star identification database also result in slow recognition speeds, failing to meet real-time requirements. Therefore, improving the accuracy and speed of star image recognition methods is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0003] This invention provides a star map recognition method, apparatus, computer device, and storage medium based on neural networks, aiming to improve the accuracy and recognition speed of star map recognition methods.

[0004] In a first aspect, embodiments of the present invention provide a star map recognition method based on a neural network, comprising:

[0005] A star map training image is acquired, and an observation star is arbitrarily selected from the star map training image as the primary star; wherein, the star map training image includes the primary star and at least one other neighboring star, and the primary star and the neighboring star are both stars;

[0006] The coordinate feature vector of the primary star is constructed using the Log-Polar algorithm, and the first brightness value of the primary star and the second brightness value of each other neighboring star in the star map training image are obtained based on the star map training image.

[0007] The relative brightness factor of the primary star is set according to the first brightness value and the second brightness value, and the relative brightness factor is mapped according to a preset brightness threshold to obtain a brightness feature vector.

[0008] A stellar feature vector set is constructed based on the coordinate feature vector and the brightness feature vector, and the stellar feature vector set is input into a neural network for training and learning. Then, the neural network outputs the number of the main star, thereby constructing a star map recognition model.

[0009] The star map recognition model is used to predict the number of a specified target star.

[0010] Secondly, embodiments of the present invention provide a star map recognition device based on a neural network, comprising:

[0011] An image acquisition unit is used to acquire a star map training image and arbitrarily select a star to be observed as the primary star in the star map training image; wherein, the star map training image includes the primary star and at least one other neighboring star, and the primary star and the neighboring star are both stars;

[0012] The coordinate vector construction unit is used to construct a coordinate feature vector for the primary star using the Log-Polar algorithm, and to obtain the first brightness value of the primary star and the second brightness value of each other neighboring star in the star map training image based on the star map training image;

[0013] A brightness vector construction unit is used to set the relative brightness factor of the primary star according to the first brightness value and the second brightness value, and to map the relative brightness factor according to a preset brightness threshold to obtain a brightness feature vector.

[0014] The model building unit is used to construct a set of stellar feature vectors based on the coordinate feature vectors and brightness feature vectors, input the set of stellar feature vectors into a neural network for training and learning, and then output the number of the primary star by the neural network to construct a star map recognition model.

[0015] The numbering prediction unit is used to predict the number of a specified target star using the star map recognition model.

[0016] Thirdly, embodiments of the present invention provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the neural network-based star map recognition method as described in the first aspect.

[0017] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the neural network-based star map recognition method as described in the first aspect.

[0018] This invention discloses a star map recognition method, apparatus, and related medium based on a neural network. The method includes: acquiring a star map training image and arbitrarily selecting a star to be observed as the primary star in the star map training image; wherein the star map training image includes the primary star and at least one other neighboring star, and the primary star and neighboring stars are all stars; constructing a coordinate feature vector for the primary star using the Log-Polar algorithm, and obtaining a first brightness value of the primary star and a second brightness value of each other neighboring star in the star map training image based on the star map training image; setting a relative brightness factor for the primary star according to the first brightness value and the second brightness value, and mapping the relative brightness factor according to a preset brightness threshold to obtain a brightness feature vector; constructing a stellar feature vector set based on the coordinate feature vector and the brightness feature vector, and inputting the stellar feature vector set into a neural network for training and learning, and then outputting the number of the primary star by the neural network to construct a star map recognition model; and using the star map recognition model to predict the number of a specified target star. This invention first selects the primary star from the star map training image and constructs its coordinate feature vector using the Log-Polar algorithm. Then, it obtains the brightness values ​​of the primary and neighboring stars and sets a relative brightness factor based on these values. A brightness feature vector is then obtained by mapping using the relative brightness factor. Next, a stellar feature vector set is constructed based on the coordinate and brightness feature vectors, and the neural network is trained to learn the model, thereby building a star map recognition model. Finally, the star map recognition model is used to predict the number of a specified target star. This invention expands the feature dimension of star map recognition by extracting and combining the feature vectors corresponding to the coordinates and brightness of stars, thus improving the accuracy of star map recognition. Furthermore, the speed of star map recognition is improved by training and constructing a star map recognition model for number prediction. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A schematic flowchart of a star map recognition method based on a neural network provided in an embodiment of the present invention;

[0021] Figure 2 An example diagram of the Log-Polar algorithm in a star map recognition method based on a neural network provided in an embodiment of the present invention;

[0022] Figure 3 This is a schematic diagram of a sub-process of a star map recognition method based on a neural network provided in an embodiment of the present invention;

[0023] Figure 4 An example diagram of relative brightness factor mapping in a star map recognition method based on a neural network provided in an embodiment of the present invention;

[0024] Figure 5 An example diagram of a fully connected neural network in a star map recognition method based on a neural network provided in an embodiment of the present invention;

[0025] Figure 6 An example diagram of the ReLU function in a star map recognition method based on a neural network provided in an embodiment of the present invention;

[0026] Figure 7 A schematic block diagram of a star map recognition device based on a neural network provided in an embodiment of the present invention;

[0027] Figure 8 This is a schematic block diagram of a star map recognition device based on a neural network, provided as an embodiment of the present invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0030] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0031] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0032] Please see below. Figure 1 , Figure 1The flowchart of a star map recognition method based on a neural network provided in an embodiment of the present invention specifically includes steps S101 to S105.

[0033] S101. Obtain a star map training image, and arbitrarily select one star to be observed as the primary star in the star map training image; wherein, the star map training image includes the primary star and at least one other neighboring star, and the primary star and the neighboring star are both stars;

[0034] S102. Construct a coordinate feature vector for the primary star using the Log-Polar algorithm, and obtain the first brightness value of the primary star and the second brightness value of each other neighboring star in the star map training image based on the star map training image;

[0035] S103. Set the relative brightness factor of the primary star according to the first brightness value and the second brightness value, and map the relative brightness factor according to a preset brightness threshold to obtain a brightness feature vector.

[0036] S104. A star feature vector set is constructed based on the coordinate feature vector and the brightness feature vector, and the star feature vector set is input into the neural network for training and learning. Then, the neural network outputs the number of the main star, thereby constructing a star map recognition model.

[0037] S105. Use the star map recognition model to predict the number of the specified target star.

[0038] In this embodiment, firstly, a star to be observed is selected from the star map training image as the primary star, and other stars in the star map training image are selected as neighboring stars. A coordinate feature vector is constructed for the primary star using the Log-Polar algorithm. Then, the brightness values ​​of the primary star and neighboring stars are obtained, and a relative brightness factor is set based on the brightness values. A brightness feature vector is obtained by mapping the relative brightness factor. Next, a stellar feature vector set is constructed based on the coordinate and brightness feature vectors. This stellar feature vector set is then used to train a neural network, thereby constructing a star map recognition model. Finally, the star map recognition model is used to predict the number of a specified target star. This embodiment expands the feature dimension of star map recognition by extracting and combining the feature vectors corresponding to the coordinates and brightness of stars, thereby improving the accuracy of star map recognition. Furthermore, by training and constructing a star map recognition model for number prediction, the speed of star map recognition is improved.

[0039] It should be noted that the star map training images in this embodiment are derived from the navigation star database, which serves as the basis for the star sensor to perform star map recognition and attitude calculation. The navigation star database is generally loaded into the memory of the star sensor and is used to store data that can be compared with observed stars, specifically including star numbers and other data.

[0040] Combination Figure 2 As shown, in one embodiment, step S102 includes:

[0041] Obtain the first Cartesian coordinates of the primary star and the second Cartesian coordinates of other neighboring stars in the star map training image;

[0042] The Log-Polar algorithm is used to map the first star map rectangular coordinates of the primary star to the origin coordinates of the star map polar coordinate system. Then, the second star map rectangular coordinates of the other neighboring stars are converted to star map polar coordinates according to the following formula:

[0043]

[0044]

[0045] Where, x i The x-coordinate of the rectangular coordinates of the star map representing the i-th neighboring star is y. i θ represents the ordinate of the rectangular coordinates of the star map of the i-th neighboring star. i The polar angle r represents the polar coordinates of the star map for the i-th neighboring star. i The polar radius represents the polar coordinates of the star map for the i-th neighboring star.

[0046] In this embodiment, the first star map rectangular coordinates corresponding to the primary star in the star map training image are first obtained, and the second star map rectangular coordinates corresponding to the neighboring star are obtained. Then, the Log-Polar algorithm is used to convert the star map rectangular coordinates corresponding to the primary star and the neighboring star into star map polar coordinates.

[0047] The Log-Polar algorithm is based on the LPT (Log-Polar Transform). The principle of the LPT is that between the human retinal surface and the visual cortex, there exists a Log-Polar mapping that plays a crucial role in the recognition of targets with size, displacement, and rotation invariance. Through the Log-Polar mapping, the size, displacement, and rotation attributes of a target can be transformed into a single displacement change that varies with the parameters, thus transforming from a Cartesian coordinate system to a logarithmic radius polar coordinate system. By introducing the Log-Polar algorithm into star map recognition for feature vector construction, the constructed feature vectors become insensitive to positional noise and magnitude noise. Furthermore, it enhances the robustness of star map recognition when faced with false stars and missing stars in the image, thereby improving the recognition accuracy of the star map recognition algorithm.

[0048] In one specific embodiment, when using a star map recognition model to predict the number of a specified target star in a captured star map image, the captured target image can first undergo image preprocessing. Image preprocessing may specifically include steps such as image filtering (neighborhood filtering) and star point extraction (extracting star pixels using the centroid method). After image preprocessing, stellar attribute data (coordinate data or brightness data) is obtained from the image, and a stellar feature vector set is constructed based on this stellar attribute data. However, the star map training images used to train the star map recognition model do not require preprocessing. This is because the star map training images are obtained from a navigation star database, which stores known stellar attribute data. Therefore, the star map training images already contain idealized stellar attribute data, thus eliminating the need for preprocessing.

[0049] In another specific embodiment, when there are many stars, a domain radius R can be set for the main star, and the domain of the main star can be divided according to the domain radius R. Then, other stars in the domain of the main star can be set as neighboring stars, thereby reducing the workload of star map recognition.

[0050] Combination Figure 3 As shown, in one embodiment, step S102 further includes steps S301 to S305:

[0051] S301. Obtain the pixel resolution m*n of the star map training image, and based on the pixel resolution, divide the polar angle coordinate axis direction corresponding to the polar angle θ into m polar angle intervals, and divide the polar radius coordinate axis direction corresponding to the polar radius r into n polar radius intervals, and label the m polar angle intervals as 1 to m, and label the n polar radius intervals as 1 to n.

[0052] S302. Project the other neighboring stars onto the polar coordinate axis to obtain the distribution of neighboring stars on the polar coordinate axis;

[0053] S303. Based on the distribution of neighboring stars, determine whether there are neighboring stars in each of the m polar angle intervals of the polar angle coordinate axis;

[0054] S304. For each polar angle interval, if it is determined that there is no neighboring star in the polar angle interval, the coordinate feature vector value of the primary star is set to 0; if it is determined that there is a neighboring star in the polar angle interval, the polar radius interval label of the existing neighboring star on the polar radius coordinate axis is obtained, and the polar radius interval label is used as the coordinate feature vector value of the primary star.

[0055] S305. Set the coordinate feature vector values ​​corresponding to all polar angle intervals into the coordinate feature vector of the primary star.

[0056] In this embodiment, based on the pixel resolution m*n of the star map training image, the polar angle θ and polar radius r are divided into the polar angle coordinate axis and polar radius coordinate axis, respectively, resulting in m polar angle intervals and n polar radius intervals. Then, neighboring stars are projected onto the polar angle coordinate axis to obtain the distribution of neighboring stars of the primary star on the polar angle coordinate axis. Next, it is determined whether each of the m polar angle intervals on the polar angle coordinate axis contains a neighboring star, thereby obtaining the coordinate feature vector value of the primary star in each polar angle interval based on the distribution of neighboring stars. Finally, the coordinate feature vector values ​​corresponding to all polar angle intervals are collected to obtain the coordinate feature vector of the primary star. Specifically, the coordinate feature vector of the primary star is represented as Pat1, Pat1 = [a1, a2, ..., a...]. i ,...,a m ], where a i This represents the coordinate eigenvector value of the primary star in the i-th polar angle interval. Here, when there is no neighboring star in the i-th polar angle interval, a i The value of a is 0; when there is a neighboring star in the i-th polar angle interval, a i The value is the polar radius interval number of the neighboring star on the polar radius coordinate axis. For example, if there is no neighboring star in the second polar angle interval, and there is a neighboring star in the third polar angle interval, and the polar radius interval number of the neighboring star in the third polar angle interval is 5 on the polar radius coordinate axis, then the coordinate eigenvector value a2 of the primary star in the second polar angle interval is 0, and the coordinate eigenvector value a3 of the primary star in the third polar angle interval is 5.

[0057] Furthermore, in one embodiment, step S304 includes:

[0058] When there are multiple neighboring stars in the polar angle interval, the polar radius interval labels corresponding to each of the multiple neighboring stars are obtained, and the smallest polar radius interval label is selected as the coordinate feature vector value of the primary star.

[0059] In this embodiment, for cases where multiple neighboring stars may exist within a polar angle interval, the polar radius interval labels corresponding to each neighboring star are compared, and the smallest polar radius interval label is selected as the coordinate feature vector value of the primary star. For example, if there are three neighboring stars in the fourth polar angle interval, and their polar radius interval labels on the polar radius coordinate axis are 1, 3, and 5 respectively, then the coordinate feature vector value a4 of the primary star in the fourth polar angle interval is 1. Furthermore, if more than one neighboring star in a polar angle interval has the smallest polar radius interval label, then any one of these neighboring stars is randomly selected as the neighboring star with the smallest polar radius interval label in that polar angle interval, and its polar radius interval label is used as the coordinate feature vector value of the primary star. For example, if there are three neighboring stars in the fifth polar angle interval, and their polar radius interval labels on the polar radius coordinate axis are 2, 2, and 4 respectively, then the coordinate feature vector value a5 of the primary star in the fifth polar angle interval is 2.

[0060] In one embodiment, step S103 includes:

[0061] The relative brightness factor of the i-th neighboring star relative to the primary star is calculated based on the first brightness value and the second brightness value according to the following formula:

[0062]

[0063] Where, λ i E represents the relative brightness factor of the i-th neighboring star relative to the primary star, E0 represents the first brightness value of the primary star, and E i This represents the second brightness value of the i-th neighboring star.

[0064] In this embodiment, the logarithm of the brightness ratio between the neighboring star and the primary star is set as the relative brightness factor of the primary star, which is also the second brightness value E of the neighboring star. i The relative brightness factor λ of the primary star is obtained by taking the logarithm of the ratio to the primary star's first brightness value E0. i Here, the relative brightness factor λ described in this embodiment... i The larger the value of λ, the brighter the i-th neighboring star is relative to the primary star; conversely, the smaller the value of λ, the brighter the relative brightness factor λ. i The smaller the value, the dimmer the i-th neighboring star is relative to the primary star. It should be noted that since there are generally large differences in brightness between stars, that is, large differences in brightness between the primary star and neighboring stars, this embodiment can narrow the range of feature factor values ​​by taking the logarithm of the brightness ratio between stars, strengthen the correlation between the relative brightness factor and the primary star, and reduce the sensitivity to brightness noise, thereby improving the robustness of the subsequently constructed brightness feature vector.

[0065] In one specific embodiment, the brightness of stars in a star map can be obtained by converting it into a grayscale image containing brightness information. In specific application scenarios, the brightness of stars requires a certain level of clarity in the star map. When the clarity of the captured star map does not meet the brightness acquisition requirements, image restoration or other technical means are needed to improve the clarity of the star map. If the improved clarity still does not meet the brightness acquisition requirements, the star map needs to be re-captured.

[0066] Combination Figure 4 As shown, in one embodiment, step S103 further includes:

[0067] Set the maximum relative brightness factor threshold and the minimum relative brightness factor threshold;

[0068] According to the following formula, all the relative luminance factors are mapped based on the maximum and minimum relative luminance factor thresholds to obtain multiple luminance feature values:

[0069]

[0070] Among them, b i a represents the brightness characteristic value of the primary star relative to the neighboring star with the smallest polar radius interval in the i-th polar angle interval. i λ represents the coordinate eigenvector value corresponding to the i-th polar angle interval. max λ represents the maximum relative brightness factor threshold. min This represents the minimum relative brightness factor threshold;

[0071] All brightness feature values ​​are set into the brightness feature vector of the primary star.

[0072] In this embodiment, the maximum threshold λ of the relative brightness factor is first set. max and minimum threshold λ min Then based on the maximum threshold λ max and minimum threshold λ min By mapping the relative brightness factor, we obtain the brightness feature values ​​of the primary star. Finally, we collect all the brightness feature values ​​to obtain the brightness feature vector of the primary star. By mapping the relative brightness factor to the brightness feature vector through a threshold range, we can effectively improve the robustness of the brightness feature vector to magnitude noise and increase the success rate of star map recognition. At the same time, by mapping the relative brightness factor to an integer brightness feature vector, we can also unify the feature vector format of the input neural network to integers, thereby improving the training convergence speed and recognition efficiency of the neural network.

[0073] In one specific embodiment, the brightness feature vector of the primary star is represented as Pat2, and Pat2 = [b1, b2, ..., b i ,...,b m ], where b i This represents the brightness characteristic value of the primary star relative to its neighboring star with the smallest polar radius interval in the i-th polar angle interval. Specifically, since the number of neighboring stars in the i-th polar angle interval may be 0, and the corresponding coordinate vector characteristic value 'a'... i It is also 0, therefore the brightness characteristic value b i Also set to 0. Furthermore, when λ i ≤λ min When the brightness of the neighboring star is less than that of the primary star and exceeds a certain range, then b will be... i Set to 1; when λ min <λ i ≤λ max When the brightness of the neighboring star is close to that of the primary star, then b is... i Set to 2; when λ i >λ max When the brightness of the neighboring star exceeds that of the primary star by a certain range, then b is... iSet it to 3.

[0074] like Figure 4 In this model, the primary star is located at the origin O. A radius R is set for the primary star's domain, and the corresponding domain is obtained. Then, other stars within this domain are considered neighboring stars. Figure 4 The primary stars are represented as Proxima 1 to Proxima 7. For Proxima 1, the angle between Proxima 1 and the x-axis in the rectangular coordinate system is its polar angle, and the logarithm of the straight line segment from the origin to Proxima 1 is its polar radius. Mapping the polar angle and polar radius onto the θ-axis (polar angle coordinate axis) and r-axis (polar radius coordinate axis) respectively converts Proxima 1 from its rectangular coordinate representation to its polar coordinate representation. Furthermore, since the relative brightness factor λ1 of Proxima 1 relative to the primary star is greater than λ... max Therefore, the brightness characteristic value b1 of the primary star relative to neighboring star 1 is 3. Figure 4 The length on the b-axis is 3. Similarly, the representation of Neighbor Star 2 is converted from rectangular coordinates to polar coordinates, and based on the fact that the relative brightness factor of Neighbor Star 2 relative to the primary star is λ2≤λ... min The brightness characteristic value b2 of the primary star relative to neighboring star 2 is obtained as 1. Figure 4 The length on the b-axis is 1, and so on, until all the remaining neighboring stars are transformed, yielding the brightness characteristic values ​​b3 to b7 of the primary star relative to neighboring stars 3 to 7. Here, b3 is 3 and b4 is 2. b5 b6 is 3, b7 is 1, and b8 is 2.

[0075] Combination Figure 5 As shown, in one embodiment, step S104 includes:

[0076] All star feature vectors in the star feature vector set are input into the input layer of a fully connected neural network, and the star feature vectors are batch standardized using the batch standardization layer in the input layer.

[0077] The first and second hidden layers of a fully connected neural network are used sequentially to train and predict the stellar feature vectors after batch normalization, and the training prediction results are used as the stellar number prediction results. The first and second hidden layers have the same network structure. The first hidden layer includes a first fully connected layer, a first activation layer, a first batch normalization layer, and a first dropout layer.

[0078] The star number prediction result is output using the output layer of a fully connected neural network, and the star number prediction result is used as the star map recognition result.

[0079] In this embodiment, stellar feature vectors are first input into the input layer of a fully connected neural network and batch-normalized. Then, the batch-normalized stellar feature vectors are sequentially input into the first and second hidden layers of the fully connected neural network. The stellar feature vectors are trained and predicted using the first and second hidden layers, and the training prediction results are used as the stellar number prediction results. Finally, the output layer of the fully connected neural network outputs the stellar number prediction results, which are ultimately used as the star map recognition results. It can be understood that the training prediction in this embodiment specifically involves: based on the weights and biases of the neural network, the batch-normalized stellar feature vectors are weighted and calculated through multiple hidden layers. Then, the weighted stellar feature vectors are nonlinearly transformed using activation functions in the multiple hidden layers to obtain the results used for the output layer. Through training prediction, this embodiment obtains the pattern relationship between the input stellar feature vectors and stars, thereby mapping the stellar feature vectors input to the neural network to the output stellar numbers. It should be noted that the weights and biases of the neural network, in addition to being pre-built, can also be adjusted based on the training and learning results of the neural network.

[0080] A neural network mainly consists of an input layer, hidden layers, and an output layer. The input layer receives the raw input data, the output layer outputs the prediction result, and the hidden layers connect the input and output through the weights and biases of the neural network, undertaking the task of information transformation and processing. During the training process of the neural network, the final prediction accuracy is improved by continuously adjusting the weights and biases between the layers. In specific application scenarios, the training process of a neural network can be carried out using supervised learning or unsupervised learning. Supervised learning adjusts the weights and biases using a training dataset to minimize the error between the predicted result and the true value; unsupervised learning automatically discovers the feature representation of the hidden layers based on the features of the input data, thereby achieving tasks such as data dimensionality reduction or clustering.

[0081] This embodiment uses a neural network approach for star map recognition, which maintains a high recognition rate while ensuring robustness against interference. Furthermore, it can be trained with large amounts of data, storing the association information between input features and star numbers in neurons, thereby reducing memory usage and improving recognition speed and accuracy. In addition, each neuron in a fully connected neural network is connected to the neuron in the previous layer, resulting in a simple structure that is easy to train and implement. Deep learning can also be used to train the weight parameters between neurons, further enhancing the recognition effect.

[0082] In specific application scenarios, the star feature vector Pat(s) is constructed by combining the coordinate feature vector Pat1 and the brightness feature vector Pat2, where Pat(s) = [Pat1Pat2] = [a1,a2,...,a2]i ,...,a m ,b1,b2,...,b i ,...,b m For each polar angle interval of the primary star, each coordinate eigenvector value a i Each has a corresponding brightness feature vector value b i By combining coordinate feature vectors and brightness feature vectors, the correlation between them can be strengthened, which can effectively increase the robustness of the identified feature information and features, thus enabling better identification even when the number of stars is small. It also improves the success rate of the final star map identification. Furthermore, by combining operations to unify the data format of coordinate feature vectors and brightness feature vectors, the convergence speed of training and learning of fully connected layers can be accelerated.

[0083] Furthermore, this embodiment accelerates the training of the fully connected neural network and maintains its stability by adding a batch normalization (BatchNorm) layer to the stellar feature vectors input to the fully connected neural network. Specifically, the main principle of BatchNorm is to normalize a batch of input features so that the mean of the input feature data distribution is 0 and the variance is 1, thereby solving problems such as gradient vanishing and gradient exploding in fully connected neural networks. It also accelerates the convergence speed of the neural network and improves the generalization performance of the model by reducing the variation in the input distribution of each layer during training.

[0084] Combination Figure 6 As shown, in one specific embodiment, a rectified unit (ReLU function) is added after the fully connected layer as the activation function of the neural network. The ReLU function is a piecewise function and is essentially a nonlinear function. Using the ReLU function as the activation function of the fully connected neural network can effectively improve the nonlinear expressive power of the fully connected neural network. The ReLU function is specifically expressed as follows:

[0085]

[0086] When the input z of the ReLU function is greater than or equal to 0, the output of the ReLU function is z; when the input z is less than 0, the output of the ReLU function is zero. Furthermore, compared to other activation functions, the ReLU function has the advantages of simple computation and low computational cost, enabling faster training speeds during backpropagation training of neural networks and effectively reducing the probability of gradient vanishing in fully connected neural networks. Simultaneously, when the input is less than 0, the ReLU function outputs 0. This characteristic allows most neurons in the fully connected neural network to remain inactive, while activating only a few neurons for processing. This reduces redundant computation in the fully connected neural network, lowers redundancy, improves the model's computational efficiency and real-time performance, thereby increasing the recognition speed of star map recognition.

[0087] In another specific embodiment, Dropout regularization is introduced into the hidden layers of the fully connected neural network. Dropout regularization effectively avoids overfitting in fully connected neural networks. Overfitting refers to a neural network performing well on the training set but poorly on the test set, resulting in a model lacking good generalization ability. Dropout regularization reduces overfitting by randomly dropping some neurons in the neural network. Specifically, during the training process, each neuron has a dropout probability, representing the probability that the neuron will be dropped during training. In each iteration of the neural network's training, Dropout regularization randomly sets the output of some neurons to 0 based on the dropout probability, thereby reducing the dependencies between neurons in the neural network. This prevents the neural network from becoming overly reliant on any particular neuron, allowing it to better adapt to various combinations of neurons, learn more robust features, better handle noise and uncertainty, and ultimately make the neural network more stable, reducing overfitting.

[0088] Figure 7 This is a schematic block diagram of a neural network-based star map recognition device 700 provided in this embodiment. The device 700 includes:

[0089] Image acquisition unit 701 is used to acquire a star map training image and arbitrarily select a star to be observed as the main star in the star map training image; wherein, the star map training image includes the main star and at least one other neighboring star, and the main star and the neighboring star are both stars;

[0090] The coordinate vector construction unit 702 is used to construct a coordinate feature vector for the primary star using the Log-Polar algorithm, and to obtain the first brightness value of the primary star and the second brightness value of each other neighboring star in the star map training image based on the star map training image;

[0091] The brightness vector construction unit 703 is used to set the relative brightness factor of the primary star according to the first brightness value and the second brightness value, and to map the relative brightness factor according to a preset brightness threshold to obtain a brightness feature vector.

[0092] The model building unit 704 is used to construct a star feature vector set based on the coordinate feature vector and the brightness feature vector, input the star feature vector set into the neural network for training and learning, and then output the number of the main star by the neural network to construct a star map recognition model.

[0093] The numbering prediction unit 705 is used to predict the number of a specified target star using the star map recognition model.

[0094] In one embodiment, the coordinate vector construction unit 702 includes:

[0095] A rectangular coordinate acquisition unit is used to acquire the first rectangular coordinates of the primary star and the second rectangular coordinates of other neighboring stars in the star map training image.

[0096] The coordinate transformation unit is used to map the first star map rectangular coordinates of the primary star to the origin coordinates of the star map polar coordinate system using the Log-Polar algorithm, and to convert the second star map rectangular coordinates of other neighboring stars to star map polar coordinates according to the following formula:

[0097]

[0098]

[0099] Where, x i The x-coordinate of the rectangular coordinates of the star map representing the i-th neighboring star is y. i θ represents the ordinate of the rectangular coordinates of the star map of the i-th neighboring star. i The polar angle r represents the polar coordinates of the star map for the i-th neighboring star. i The polar radius represents the polar coordinates of the star map for the i-th neighboring star.

[0100] Combination Figure 8 As shown, in one embodiment, the coordinate vector construction unit 702 further includes:

[0101] The interval division unit 801 is used to obtain the pixel resolution m*n of the star map training image, and divide the polar angle coordinate axis direction corresponding to the polar angle θ into m polar angle intervals based on the pixel resolution, and divide the polar radius coordinate axis direction corresponding to the polar radius r into n polar radius intervals, and number the m polar angle intervals sequentially from 1 to m, and number the n polar radius intervals sequentially from 1 to n.

[0102] The neighboring star projection unit 802 is used to project the other neighboring stars onto the polar coordinate axis to obtain the distribution of neighboring stars on the polar coordinate axis.

[0103] The neighboring star distribution judgment unit 803 is used to determine whether there are neighboring stars in each of the m polar angle intervals of the polar angle coordinate axis based on the neighboring star distribution.

[0104] The coordinate vector value setting unit 804 is used to, for each polar angle interval, if it is determined that there is no neighboring star in the polar angle interval, set the coordinate feature vector value of the main star to 0; if it is determined that there is a neighboring star in the polar angle interval, obtain the polar radius interval label of the existing neighboring star on the polar radius coordinate axis, and use the polar radius interval label as the coordinate feature vector value of the main star.

[0105] The coordinate vector value set unit 805 is used to set the coordinate feature vector values ​​corresponding to all polar angle intervals into the coordinate feature vector of the main star.

[0106] In one embodiment, the coordinate vector value setting unit 804 includes:

[0107] The minimum interval selection unit is used to obtain the polar radius interval labels corresponding to each of the multiple neighboring stars when there are multiple neighboring stars in the polar angle interval, and select the smallest polar radius interval label as the coordinate feature vector value of the main star.

[0108] In one embodiment, the luminance vector construction unit 703 includes:

[0109] A brightness factor calculation unit is used to calculate the relative brightness factor of the i-th neighboring star relative to the primary star according to the first brightness value and the second brightness value, using the following formula:

[0110]

[0111] Where, λ i E represents the relative brightness factor of the i-th neighboring star relative to the primary star, E0 represents the first brightness value of the primary star, and E i This represents the second brightness value of the i-th neighboring star.

[0112] In one embodiment, the luminance vector construction unit 703 further includes:

[0113] The threshold setting unit is used to set the maximum relative brightness factor threshold and the minimum relative brightness factor threshold.

[0114] A luminance factor mapping unit is used to map all the relative luminance factors according to the following formula, based on the maximum relative luminance factor threshold and the minimum relative luminance factor threshold, to obtain multiple luminance feature values:

[0115]

[0116] Among them, b i a represents the brightness characteristic value of the primary star relative to the neighboring star with the smallest polar radius interval in the i-th polar angle interval. i λ represents the coordinate eigenvector value corresponding to the i-th polar angle interval. max λ represents the maximum relative brightness factor threshold. min This represents the minimum relative brightness factor threshold;

[0117] A brightness feature set unit is used to set all brightness feature values ​​into a brightness feature vector of the primary star.

[0118] In one embodiment, the model building unit 704 includes:

[0119] The feature vector input unit is used to input all star feature vectors in the star feature vector set into the input layer of the fully connected neural network, and to perform batch standardization processing on the star feature vectors using the batch standardization layer in the input layer;

[0120] The training prediction unit is used to train and predict the star feature vectors after batch normalization by using the first hidden layer and the second hidden layer in the fully connected neural network in turn, and to use the training prediction result as the star number prediction result; wherein, the first hidden layer and the second hidden layer have the same network structure, and the first hidden layer includes a first fully connected layer, a first activation layer, a first batch normalization layer and a first dropout layer in turn.

[0121] The result output unit is used to output the star number prediction result using the output layer of the fully connected neural network, and to use the star number prediction result as the star map recognition result.

[0122] Since the embodiments of the apparatus and the embodiments of the method correspond to each other, please refer to the description of the embodiments of the method for the embodiments of the apparatus, which will not be repeated here.

[0123] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed, can perform the steps provided in the above embodiments. The storage medium may include various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0124] This invention also provides a computer device, which may include a memory and a processor. The memory stores a computer program, and when the processor calls the computer program in the memory, it can implement the steps provided in the above embodiments. Of course, the computer device may also include various network interfaces, power supplies, and other components.

[0125] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to in the method section. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of the claims of this application.

[0126] It should also be noted that, in this specification, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A neural network-based star pattern recognition method, characterized by, include: A star map training image is acquired, and an observation star is arbitrarily selected from the star map training image as the primary star; wherein, the star map training image includes the primary star and at least one other neighboring star, and the primary star and the neighboring star are both stars; The coordinate feature vector of the primary star is constructed using the Log-Polar algorithm, and the first brightness value of the primary star and the second brightness value of each other neighboring star in the star map training image are obtained based on the star map training image. The relative brightness factor of the primary star is set according to the first brightness value and the second brightness value, and the relative brightness factor is mapped according to a preset brightness threshold to obtain a brightness feature vector. A stellar feature vector set is constructed based on the coordinate feature vector and the brightness feature vector, and the stellar feature vector set is input into a neural network for training and learning. Then, the neural network outputs the number of the main star, thereby constructing a star map recognition model. The star map recognition model is used to predict the number of specified target stars; The step of constructing a coordinate feature vector for the primary star using the Log-Polar algorithm includes obtaining the first star map Cartesian coordinates of the primary star and the second star map Cartesian coordinates of other neighboring stars in the star map training image; using the Log-Polar algorithm to map the first star map Cartesian coordinates of the primary star to the origin coordinates of the star map polar coordinate system, and converting the second star map Cartesian coordinates of other neighboring stars to star map polar coordinates; wherein... The polar angle r represents the polar coordinates of the star map for the i-th neighboring star. i The polar radius represents the polar coordinates of the star map for the i-th neighboring star; The step of constructing coordinate feature vectors for the primary star using the Log-Polar algorithm also includes obtaining the pixel resolution m*n of the star map training image, and determining the polar angle based on the pixel resolution. The corresponding polar coordinate axis direction is divided into m polar angle intervals, and the polar radius coordinate axis direction corresponding to the polar radius r is divided into n polar radius intervals. The m polar angle intervals are sequentially labeled as 1~m, and the n polar radius intervals are sequentially labeled as 1~n. The other neighboring stars are projected onto the polar coordinate axis to obtain the distribution of neighboring stars on the polar coordinate axis. Based on the distribution of neighboring stars, it is determined whether each of the m polar angle intervals of the polar coordinate axis has a neighboring star. For each polar angle interval, if it is determined that there is no neighboring star in the polar angle interval, the coordinate feature vector value of the primary star is set to 0. If it is determined that there is a neighboring star in the polar angle interval, the polar radius interval label of the existing neighboring star on the polar radius coordinate axis is obtained, and the polar radius interval label is used as the coordinate feature vector value of the primary star. The coordinate feature vector values ​​corresponding to all polar angle intervals are collected into the coordinate feature vector of the primary star. The step of setting the relative brightness factor of the primary star based on the first brightness value and the second brightness value, and mapping the relative brightness factor according to a preset brightness threshold to obtain a brightness feature vector, includes: The relative brightness factor of the i-th neighboring star relative to the primary star is calculated based on the first brightness value and the second brightness value according to the following formula: ; in, E represents the relative brightness factor of the i-th neighboring star relative to the primary star, E0 represents the first brightness value of the primary star, and E i This represents the second brightness value of the i-th neighboring star.

2. The star map recognition method based on neural networks according to claim 1, characterized in that, The conversion of the Cartesian coordinates of the second star map of other neighboring stars to polar coordinates is calculated using the following formula: ； Where, x i The x-coordinate of the rectangular coordinates of the star map representing the i-th neighboring star is y. i The ordinate of the rectangular coordinates of the star map represents the coordinates of the i-th neighboring star. The polar angle r represents the polar coordinates of the star map for the i-th neighboring star. i The polar radius represents the polar coordinates of the star map for the i-th neighboring star.

3. The star map recognition method based on neural networks according to claim 1, characterized in that, If it is determined that a neighboring star exists in the polar angle interval, then the polar radius interval label of the existing neighboring star on the polar radius coordinate axis is used as the coordinate feature vector value of the primary star, including: When there are multiple neighboring stars in the polar angle interval, the polar radius interval labels corresponding to each of the multiple neighboring stars are obtained, and the smallest polar radius interval label is selected as the coordinate feature vector value of the primary star.

4. The star map recognition method based on neural networks according to claim 1, characterized in that, The step of setting the relative brightness factor of the primary star based on the first brightness value and the second brightness value, and mapping the relative brightness factor according to a preset brightness threshold to obtain a brightness feature vector, further includes: Set the maximum relative brightness factor threshold and the minimum relative brightness factor threshold; According to the following formula, all the relative luminance factors are mapped based on the maximum and minimum relative luminance factor thresholds to obtain multiple luminance feature values: ； Among them, b i a represents the brightness characteristic value of the primary star relative to the neighboring star with the smallest polar radius interval in the i-th polar angle interval. i This represents the coordinate eigenvector value corresponding to the i-th polar angle interval. This represents the maximum relative brightness factor threshold. This represents the minimum relative brightness factor threshold; All brightness feature values ​​are set into the brightness feature vector of the primary star.

5. The star map recognition method based on neural networks according to claim 1, characterized in that, The process of constructing a stellar feature vector set based on the coordinate feature vector and the brightness feature vector, inputting the stellar feature vector set into a neural network for training, and then having the neural network output the number of the primary star includes: All star feature vectors in the star feature vector set are input into the input layer of a fully connected neural network, and the star feature vectors are batch standardized using the batch standardization layer in the input layer. The first and second hidden layers of a fully connected neural network are used sequentially to train and predict the stellar feature vectors after batch normalization, and the training prediction results are used as the stellar number prediction results. The first and second hidden layers have the same network structure. The first hidden layer includes a first fully connected layer, a first activation layer, a first batch normalization layer, and a first dropout layer. The star number prediction result is output using the output layer of a fully connected neural network, and the star number prediction result is used as the star map recognition result.

6. A star map recognition device based on a neural network, characterized in that, include: An image acquisition unit is used to acquire a star map training image and arbitrarily select a star to be observed as the primary star in the star map training image; wherein, the star map training image includes the primary star and at least one other neighboring star, and the primary star and the neighboring star are both stars; The coordinate vector construction unit is used to construct a coordinate feature vector for the primary star using the Log-Polar algorithm, and to obtain the first brightness value of the primary star and the second brightness value of each other neighboring star in the star map training image based on the star map training image; A brightness vector construction unit is used to set the relative brightness factor of the primary star according to the first brightness value and the second brightness value, and to map the relative brightness factor according to a preset brightness threshold to obtain a brightness feature vector. The model building unit is used to construct a set of stellar feature vectors based on the coordinate feature vectors and brightness feature vectors, input the set of stellar feature vectors into a neural network for training and learning, and then output the number of the primary star by the neural network to construct a star map recognition model. The numbering prediction unit is used to predict the number of a specified target star using the star map recognition model. The coordinate vector construction unit is specifically used to obtain the first star map rectangular coordinates of the primary star and the second star map rectangular coordinates of other neighboring stars in the star map training image; it uses the Log-Polar algorithm to map the first star map rectangular coordinates of the primary star to the origin coordinates of the star map polar coordinate system, and converts the second star map rectangular coordinates of other neighboring stars to star map polar coordinates; wherein... The polar angle r represents the polar coordinates of the star map for the i-th neighboring star. i The polar radius represents the polar coordinates of the star map for the i-th neighboring star; The coordinate vector construction unit is further specifically used to obtain the pixel resolution m*n of the star map training image, and to calculate the polar angle based on the pixel resolution. The corresponding polar coordinate axis direction is divided into m polar angle intervals, and the polar radius coordinate axis direction corresponding to the polar radius r is divided into n polar radius intervals. The m polar angle intervals are sequentially labeled as 1~m, and the n polar radius intervals are sequentially labeled as 1~n. The other neighboring stars are projected onto the polar coordinate axis to obtain the distribution of neighboring stars on the polar coordinate axis. Based on the distribution of neighboring stars, it is determined whether each of the m polar angle intervals of the polar coordinate axis has a neighboring star. For each polar angle interval, if it is determined that there is no neighboring star in the polar angle interval, the coordinate feature vector value of the primary star is set to 0. If it is determined that there is a neighboring star in the polar angle interval, the polar radius interval label of the existing neighboring star on the polar radius coordinate axis is obtained, and the polar radius interval label is used as the coordinate feature vector value of the primary star. The coordinate feature vector values ​​corresponding to all polar angle intervals are collected into the coordinate feature vector of the primary star. The brightness vector construction unit is specifically used to calculate the relative brightness factor of the i-th neighboring star relative to the primary star according to the first brightness value and the second brightness value, as follows: ； in, E represents the relative brightness factor of the i-th neighboring star relative to the primary star, E0 represents the first brightness value of the primary star, and E i This represents the second brightness value of the i-th neighboring star.

7. A computer device, characterized in that, The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the neural network-based star map recognition method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the neural network-based star map recognition method as described in any one of claims 1 to 5.

Images