Space target image intelligent identification method

Abstract

The invention relates to the related field of image processing and computer vision, in particular to a space target image intelligent identification method, which optimizes a series of training strategies, including learning rate dynamic adjustment (such as attenuation and preheating), adaptive optimization algorithm and regularization technology (such as Dropout and data enhancement). And the accuracy, generalization ability and convergence speed of the model are obviously improved. In addition, through a model compression and reasoning optimization technology, efficient reasoning in a resource-constrained environment is realized, and the real-time performance of space image recognition is ensured. The flexible deployment strategy of the model enables the model to adapt to various platforms, including a cloud, a local server and edge equipment, and has good expansibility and practicability. The visualized recognition result is presented, so that the intuition of the recognition result is further improved, and powerful support is provided for subsequent analysis and decision making. The method has wide application prospect and practical value in space image recognition and other related fields.

Description

Technical Field

[0001] The present invention relates to the fields of image processing and computer vision, and specifically to an intelligent recognition method for spatial target images. Background Art

[0002] In the process of space exploration and earth observation, it is necessary to identify and classify a large number of spatial targets. Most traditional methods rely on manual analysis or feature-based machine learning methods, which are slow in processing speed and vulnerable to environmental changes. With the rise of deep learning technology, image recognition methods based on deep neural networks have gradually shown advantages. Summary of the Invention

[0003] The purpose of the present invention is to provide an intelligent recognition method for spatial target images to solve the problems raised in the above background art.

[0004] To achieve the above purpose, the present invention provides the following technical solution: an intelligent recognition method for spatial target images, including the following steps:

[0005] Step 1, data preprocessing: Collect multi-source spatial target image data, perform annotation, and at the same time perform normalization processing and data augmentation on the annotated image data to improve the generalization ability of the model;

[0006] Step 2, model design: Construct a deep convolutional neural network, the network includes multiple convolutional layers, pooling layers and fully connected layers, and use the cross-entropy loss function and the stochastic gradient descent optimization algorithm to train the model;

[0007] Step 3, training process: Use the annotated spatial target image data set in Step 1 to train the model in Step 2, and adjust the parameters to improve the recognition accuracy of the model;

[0008] Step 4, model evaluation and optimization: Evaluate the model performance on the validation data set, analyze the recognition accuracy through the confusion matrix, and then optimize and adjust the network structure according to the evaluation results;

[0009] Step 5, real-time recognition application: Deploy the trained model on the processing platform, input the real-time received spatial images for recognition, output the recognition results, and cooperate with the image processing algorithm to present the results visually.

[0010] Preferably, the specific steps of data preprocessing in Step 1 are as follows:

[0011] Step 11, collect multi-source spatial target image data:

[0012] Data source: Utilize satellite data, images taken by aircraft, and data from ground observation stations, and combine with image data from commercial satellite companies, government agencies, and public databases; among which the image data covers various environmental conditions, angles, sizes, and postures of different spatial targets;

[0013] Data format: The image data includes multiple data formats, specifically JPEG, PNG, TIFF, and RAW format, and the image data category contains different types of images such as multi-spectral, panchromatic, and infrared;

[0014] Step 12: Data annotation:

[0015] Annotation tools and methods: Use professional image annotation tools to annotate the image data, and the annotation categories include target types, specifically stars, planets, and spacecraft debris;

[0016] Annotation strategy: Adopt multi-level annotation to annotate the image data, specifically the first-level annotation of basic categories and the second-level detailed categories, and use expert review of the annotation results to ensure the annotation quality;

[0017] Dataset division: Divide the processed image data into a training set, a validation set, and a test set, with specific proportions of 70% / 15% / 15%;

[0018] Step 13: Image normalization processing: Normalize the pixel values of the image data to the 0-1 interval to reduce the influence of illumination changes; use global contrast normalization technology to enhance the contrast of image features; adopt bilinear interpolation or nearest neighbor interpolation methods for image data scaling to adjust the image size to a unified size to meet the input requirements of the neural network;

[0019] Step 14: Data augmentation: Randomly rotate the image by 0-360 degrees, flip it horizontally or vertically to simulate different observation angles; randomly scale the image size and translate the position to simulate different target distances and positions.

[0020] Preferably, the specific steps of the model design in step 2 are as follows:

[0021] Step 21: Construct a deep convolutional neural network:

[0022] Input layer: Receive the image data after normalization processing in step 1;

[0023] Convolutional layer: Each convolutional layer extracts features through multiple convolutional kernels, that is, filters. The specific convolution operation formula is:

[0024]

[0025] Among them, input(i+m,j+n) represents the pixel value of the input image at position (i+m,j+n), and bias is the bias value;

[0026] Pooling layer: Max pooling is adopted to reduce the resolution of the feature map in the image data, reduce the number of parameters and the amount of computation. The max pooling formula:

[0027]

[0028] Among them, k is the pooling window size;

[0029] Fully connected layer: After being processed by multiple convolutional layers and pooling layers, the extracted feature information is input into the fully connected layer for further feature extraction and classification. The fully connected layer integrates and further processes the features extracted by the previous convolutional layer and pooling layer to obtain the final feature representation;

[0030] Step 22: Use the cross-entropy loss function: Train the CNN with a large amount of training data so that it can automatically learn the feature information of the spatial target. The backpropagation algorithm and optimization algorithm are used in the training process;

[0031] Define the loss function: The loss function is used to measure the difference between the network output and the true label. For the spatial target image classification task, the cross-entropy loss function is used as the loss function. The cross-entropy loss function formula:

[0032]

[0033] where y i represents the true label, represents the probability that the predicted label is class i;

[0034] Then use the backpropagation algorithm to calculate the gradient of the loss function with respect to the network parameters. Through the chain rule, the gradient of the loss function is propagated backward from the output layer to the input layer to update the weight parameters and bias terms of the network;

[0035] Then use the optimization algorithm to update the network parameters according to the gradient of the loss function to minimize the loss function.

[0036] Preferably, the training process of step 3 is specifically as follows:

[0037] Step 31: Use the labeled spatial target image dataset for training

[0038] Data preparation: Ensure that the dataset of image data has been preprocessed and divided into a training set, a validation set, and a test set. The training set is used to update the model parameters, the validation set is used to evaluate the model performance and tune the parameters, and the test set is used for the final performance evaluation;

[0039] Initialize model parameters: Use a suitable weight initialization method, specifically Xavier initialization or He initialization, to prevent gradient vanishing or explosion;

[0040] Training loop: The iterative training process is usually divided into several epochs. Each epoch represents a complete traversal of the entire training set. During training, the training data is randomly shuffled to ensure the independent and identically distributed nature of the data;

[0041] Step 32. Adjust the parameters to improve the recognition accuracy of the model

[0042] Select an optimization algorithm: Initially use the Stochastic Gradient Descent algorithm;

[0043] Learning rate adjustment: Use a learning rate decay strategy to gradually reduce the learning rate to improve the stability of training; Dynamically adjust the learning rate, specifically ReduceLROnPlateau: If the validation set performance does not improve after a certain number of epochs, then reduce the learning rate;

[0044] Step 33. Strategies to prevent overfitting

[0045] Early stopping: Monitor the loss or accuracy of the validation set during training. If the validation set performance does not improve within a certain number of epochs, then stop training early to prevent overfitting;

[0046] Data augmentation: Implement real-time data augmentation techniques during the training phase to expand the training data space and improve the generalization ability of the model;

[0047] Validation set evaluation: After each epoch, evaluate the model performance on the validation set, record the training and validation losses and accuracies of each epoch to intuitively monitor the training process of the model;

[0048] Step 34. Model evaluation and saving

[0049] During the validation phase, use metrics such as accuracy, precision, recall, and F1-score for performance evaluation, and save the optimal model, that is, the model weights that achieve the lowest loss or the highest accuracy on the validation set.

[0050] Preferably, the specific steps of the real-time recognition application in step 5 are as follows:

[0051] Step 51. Model deployment: Convert the model trained in step 4 into a format suitable for the target platform, and use tools such as TensorRT to optimize the model to improve the inference speed and efficiency; Configure the operating environment of the target platform, including installing necessary deep learning frameworks and dependency libraries; Deploy the model on the platform for real-time program calls;

[0052] Step 52, Receive and process real-time images: Receive spatial image data in real time and perform image preprocessing before inputting it into the model;

[0053] Step 53, Image recognition and result output: Use the deployed model to perform inference on the received real-time images and output classification labels and related probability values or confidence levels;

[0054] Step 54, Visual presentation of results: Use image processing algorithms to overlay the recognition results on the original image data, specifically including drawing bounding boxes, label texts, or heatmaps on the image data; Implement a user interface for real-time data display and update; The display methods include desktop application terminals, web pages, or direct output on the console.

[0055] Compared with the prior art, the beneficial effects of the present invention are as follows: Through optimized training strategies (such as learning rate adjustment, regularization, data augmentation, etc.), the accuracy of the model is effectively improved. At the same time, the introduction of regularization techniques (such as Dropout, weight decay, etc.) significantly enhances the generalization ability of the model, enabling it to maintain high-precision recognition results when processing unseen spatial images; Through the adaptive learning rate optimization algorithm and strategies such as learning rate warm-up and decay, the model can converge quickly in the initial stage of training and perform fine-tuning in the later stage of training, thereby greatly shortening the training time and improving the training efficiency; The present invention introduces a variety of regularization methods and data augmentation techniques, effectively preventing the problem of model overfitting and enhancing the robustness of the model to noise and deformation; After format conversion and optimization, the model in the present invention can be flexibly deployed on various platforms, including cloud servers, local workstations, and edge devices. This makes the present invention have wide applicability and scalability in practical applications and can meet the requirements in different scenarios, such as satellite ground stations, UAV monitoring, etc.; The present invention visualizes the recognition results through image processing algorithms, such as overlaying bounding boxes, label texts, etc. on the original image. This not only improves the readability and intuitiveness of the results but also provides strong support for subsequent analysis and decision-making. For example, in spatial target monitoring, targets can be quickly located and marked for real-time monitoring and emergency response. Brief Description of the Drawings

[0056] Figure 1 It is a schematic structural diagram of the method flow of the present invention. Detailed Embodiments

[0057] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0058] Please refer to Figure 1 , the present invention provides a technical solution: an intelligent recognition method for space target images, including the following steps:

[0059] Step 1, data preprocessing: Collect multi-source space target image data, perform annotation, and at the same time perform normalization processing and data augmentation on the annotated image data to improve the generalization ability of the model;

[0060] Step 2, model design: Construct a deep convolutional neural network, the network includes multiple convolutional layers, pooling layers and fully connected layers, and use the cross-entropy loss function and the stochastic gradient descent optimization algorithm to train the model;

[0061] Step 3, training process: Use the annotated space target image data set in Step 1 to train the model in Step 2, and adjust the parameters to improve the recognition accuracy of the model;

[0062] Step 4, model evaluation and optimization: Evaluate the model performance on the validation data set, analyze the recognition accuracy through the confusion matrix, and then optimize and adjust the network structure according to the evaluation results;

[0063] Step 5, real-time recognition application: Deploy the trained model on the processing platform, input the real-time received space images for recognition, output the recognition results, and cooperate with the image processing algorithm for visual presentation of the results.

[0064] Further, the specific steps of data preprocessing in Step 1 are as follows:

[0065] Step 11, collect multi-source space target image data:

[0066] Data source: Utilize satellite data, images taken by aircraft, and data from ground observation stations, and combine the image data of commercial satellite companies, government agencies, and public databases; among them, the image data covers various environmental conditions, angles, sizes, and postures of different space targets;

[0067] Data format: The image data includes multiple data formats, specifically JPEG, PNG, TIFF, and RAW format, and the image data category includes different types of images such as multi-spectral, panchromatic, and infrared;

[0068] Step 12, data annotation:

[0069] Annotation tools and methods: Use professional image annotation tools to annotate the image data, and the annotation categories include target types, specifically stars, planets, and spacecraft debris;

[0070] Annotation strategy: Multilevel annotation is adopted to annotate the image data. Specifically, the basic categories are annotated at the first level, and the sub - categories are annotated at the second level. The expert review is used to ensure the quality of the annotation results;

[0071] Dataset division: The processed image data is divided into a training set, a validation set, and a test set, with specific proportions of 70% / 15% / 15%;

[0072] Step 13. Image normalization processing: Normalize the pixel values of the image data to the range of 0 - 1 to reduce the influence of illumination changes; Use global contrast normalization technology to enhance the contrast of image features; Adopt bilinear interpolation or nearest - neighbor interpolation methods to scale the image data and adjust the image size to a unified size to meet the input requirements of the neural network;

[0073] Step 14. Data augmentation: Randomly rotate the image by 0 - 360 degrees, flip it horizontally or vertically to simulate different observation angles; Randomly scale the image size and translate the position to simulate different target distances and positions.

[0074] Furthermore, the specific steps of model design in Step 2 are as follows:

[0075] Step 21. Build a deep convolutional neural network:

[0076] Input layer: Accept the image data after normalization processing in Step 1;

[0077] Convolutional layer: Each convolutional layer extracts features through multiple convolutional kernels, that is, filters. The specific convolutional operation formula is:

[0078]

[0079] Among them, input(i + m, j + n) represents the pixel value of the input image at the position (i + m, j + n), and bias is the bias value;

[0080] Pooling layer: Adopt max - pooling to reduce the resolution of the feature map in the image data, reduce the number of parameters and the amount of calculation. The max - pooling formula:

[0081]

[0082] Among them, k is the pooling window size;

[0083] Fully - connected layer: After being processed by multiple convolutional layers and pooling layers, the extracted feature information is input into the fully - connected layer for further feature extraction and classification. The fully - connected layer integrates and further processes the features extracted by the previous convolutional layers and pooling layers to obtain the final feature representation;

[0084] Step 22: Use the cross-entropy loss function: Train the CNN with a large amount of training data so that it can automatically learn the feature information of spatial targets. The backpropagation algorithm and optimization algorithm are used in the training process;

[0085] Define the loss function: The loss function is used to measure the difference between the network output and the true label. For the spatial target image classification task, the cross-entropy loss function is used as the loss function. The formula for the cross-entropy loss function is:

[0086]

[0087] where y i represents the true label, represents the probability that the predicted label is class i;

[0088] Then use the backpropagation algorithm to calculate the gradient of the loss function with respect to the network parameters. Through the chain rule, the gradient of the loss function is propagated backward from the output layer to the input layer to update the weight parameters and bias terms of the network;

[0089] Then use the optimization algorithm to update the network parameters according to the gradient of the loss function to minimize the loss function.

[0090] Further, the specific process of step 3 training is as follows:

[0091] Step 31: Use the labeled spatial target image dataset for training

[0092] Data preparation: Ensure that the dataset of image data has undergone preprocessing steps and is divided into a training set, a validation set, and a test set. The training set is used to update the model parameters, the validation set is used to evaluate the model performance and adjust the parameters, and the test set is used for the final performance evaluation;

[0093] Initialize the model parameters: Adopt a suitable weight initialization method, specifically use Xavier initialization or He initialization to prevent gradient vanishing or explosion;

[0094] Training loop: The iterative training process is usually divided into several epochs. Each epoch represents a complete traversal of the entire training set. During the training process, the training data is randomly shuffled to ensure the independent and identically distributed data;

[0095] Step 32: Adjust the parameters to improve the recognition accuracy of the model

[0096] Select the optimization algorithm: Initially use the stochastic gradient descent algorithm;

[0097] Learning rate adjustment: Use a learning rate decay strategy to gradually reduce the learning rate to improve the stability of training; Dynamically adjust the learning rate, specifically ReduceLROnPlateau: If the validation set performance does not improve after several epochs, then reduce the learning rate;

[0098] Step 33. Overfitting prevention strategy

[0099] Early stopping: Monitor the loss or accuracy of the validation set during training. If the validation set performance does not improve within a certain number of epochs, then stop training early to prevent overfitting;

[0100] Data augmentation: Implement real-time data augmentation techniques during the training phase to expand the training data space and improve the generalization ability of the model;

[0101] Validation set evaluation: After each epoch, evaluate the model performance on the validation set, record the training and validation losses and accuracies of each epoch to intuitively monitor the training process of the model;

[0102] Step 34. Model evaluation and saving

[0103] During the validation phase, use metrics such as accuracy, precision, recall, and F1-score for performance evaluation, and save the optimal model, that is, the model weights that achieve the lowest loss or the highest accuracy on the validation set.

[0104] Furthermore, the specific steps of the real-time recognition application in Step 5 are as follows:

[0105] Step 51. Model deployment: Convert the model trained in Step 4 into a format suitable for the target platform, and use tools such as TensorRT to optimize the model to improve the inference speed and efficiency; Configure the running environment of the target platform, including installing necessary deep learning frameworks and dependency libraries; Deploy the model on the platform for real-time program calls;

[0106] Step 52. Receive and process real-time images: Receive spatial image data in real-time and perform image preprocessing before inputting it into the model;

[0107] Step 53. Image recognition and result output: Use the deployed model to perform inference on the received real-time images and output classification labels and related probability values or confidence levels;

[0108] Step 54. Visualization and presentation of results: Use image processing algorithms to overlay the recognition results on the original image data, specifically including drawing bounding boxes, label texts, or heatmaps on the image data; Implement a user interface for real-time data display and update; The display methods include desktop application clients, web pages, or direct output on the console.

[0109] The present invention provides a deep learning model training and deployment solution for real-time spatial image recognition applications. Through a series of optimized training strategies and algorithm improvements, the performance, generalization ability and reasoning efficiency of the model are significantly improved. Specifically, the present invention achieves technical improvements and optimizations in the following aspects:

[0110] By adjusting the learning rate (including attenuation, preheating, adaptive optimization algorithm, etc.), batch size adjustment, regularization methods (such as Dropout, weight decay, data enhancement), etc., the training efficiency and accuracy of the model are improved, and the generalization ability of the model is enhanced, making it more stable in dealing with different environments and tasks;

[0111] The converted and optimized models can be flexibly deployed on a variety of platforms, including the cloud, local servers, and edge devices, with good scalability and adaptability. This provides technical support for system integration and real-time response in practical applications;

[0112] The recognition results are visualized through image processing algorithms (such as bounding boxes, label text, etc.), which improves the intuitiveness and readability of the recognition results and provides strong support for subsequent analysis and decision-making.

[0113] In general, the present invention significantly improves the performance of deep learning models in spatial image recognition through a series of technical optimizations and innovations, and has multiple advantages such as real-time, high efficiency, stability and flexibility, and is suitable for spatial image recognition tasks in a variety of complex scenarios. This makes the present invention have broad application prospects and practical value in related fields.

[0114] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent recognition of space target images, characterized in that: The following steps are involved: Step 1: Data preprocessing: Collect multi-source spatial target image data, annotate them, and perform normalization and data enhancement on the annotated image data to improve the generalization ability of the model; Step 2: Model design: Build a deep convolutional neural network, which includes multiple convolutional layers, pooling layers, and fully connected layers. Use the cross entropy loss function and stochastic gradient descent optimization algorithm to train the model. Step 3, training process: Use the space target image dataset labeled in step 1 to train the model in step 2, and adjust the parameters to improve the recognition accuracy of the model; Step 4: Model evaluation and optimization: Evaluate the model performance of the model trained in step 3 on the validation data set, identify the accuracy through confusion matrix analysis, and then optimize and adjust the network structure based on the evaluation results; Step 5: Real-time recognition application: Deploy the trained model on the processing platform, input the real-time received spatial image for recognition, output the recognition results, and use the image processing algorithm to visualize the results.

2. The method for intelligent recognition of space target images according to claim 1, characterized in that: The specific steps of data preprocessing in step 1 are as follows: Step 11: Collect multi-source spatial target image data: Data sources: satellite data, images taken by aircraft, and data from ground observation stations, combined with image data from commercial satellite companies, government agencies, and public databases; the image data covers a variety of environmental conditions, angles, sizes, and attitudes of different space targets; Data format: Image data includes multiple data formats, specifically JPEG, PNG, TIFF, and RAW formats, and the image data class includes different types of images such as multispectral, panchromatic, and infrared; Step 12: Data labeling: Annotation tools and methods: Use professional image annotation tools to annotate image data. Annotation categories include target types, specifically stars, planets, and spacecraft debris. Annotation strategy: Multi-level annotation is used to annotate image data, specifically the first-level annotation of basic categories and the second-level annotation of subcategories. Experts are used to review the annotation results to ensure the quality of annotation. Dataset division: The processed image data is divided into training set, validation set and test set, with a specific ratio of 70% / 15% / 15%; Step 13: Image normalization: normalize the pixel values ​​of the image data to the range of 0-1 to reduce the impact of illumination changes; Use global contrast normalization technology to enhance the contrast of image features; use bilinear interpolation or nearest neighbor interpolation method to scale image data and adjust the image size to a uniform size to meet the neural network input requirements; Step 14: Data augmentation: randomly rotate the image 0-360 degrees, flip it horizontally or vertically to simulate different observation angles; randomly scale the image size and translate the position to simulate different target distances and positions.

3. The method for intelligent recognition of space target images according to claim 1, characterized in that: The specific steps of model design in step 2 are as follows: Step 21: Build a deep convolutional neural network: Input layer: accepts the image data after normalization in step 1; Convolutional layer: Each convolutional layer extracts features through multiple convolution kernels, i.e. filters. The specific convolution operation formula is: Among them, input(i+m,j+n) represents the pixel value of the input image at position (i+m,j+n), and bias is the bias value; Pooling layer: Maximum pooling is used to reduce the resolution of feature maps in image data, reduce the number of parameters and calculations, and the maximum pooling formula is: Among them, k is the pooling window size; Fully connected layer: After being processed by multiple convolutional layers and pooling layers, the extracted feature information is input into the fully connected layer for further feature extraction and classification. The fully connected layer integrates and further processes the features extracted by the previous convolutional layer and pooling layer to obtain the final feature representation; Step 22: Use the cross entropy loss function: Train CNN with a large amount of training data so that it can automatically learn the feature information of the spatial target. The training process uses the back propagation algorithm and the optimization algorithm. Definition of loss function: The loss function is used to measure the difference between the network output and the true label. For the spatial target image classification task, the loss function uses the cross entropy loss function, where the cross entropy loss function formula is: Among them, y i represents the true label, Represents the probability that the predicted label is class i; Then, the back-propagation algorithm is used to calculate the gradient of the loss function with respect to the network parameters. Through the chain rule, the gradient of the loss function is back-propagated from the output layer to the input layer to update the weight parameters and bias terms of the network. Then use the optimization algorithm to update the network parameters according to the gradient of the loss function to minimize the loss function.

4. The method for intelligent recognition of space target images according to claim 1, characterized in that: The training process of step 3 is as follows: Step 31: Use the labeled space target image dataset for training Data preparation: Ensure that the image data set has undergone preprocessing steps and is divided into training set, validation set, and test set. The training set is used to update model parameters, the validation set is used to evaluate model performance and adjust parameters, and the test set is used for final performance evaluation. Initialize model parameters: Use appropriate weight initialization methods, specifically Xavier initialization or He initialization, to prevent gradient disappearance or explosion; Training cycle: The iterative training process is usually divided into several epochs, each epoch represents a complete traversal of the entire training set. During the training process, the training data is randomly shuffled to ensure that the data is independent and identically distributed; Step 32: Adjust parameters to improve the recognition accuracy of the model Select optimization algorithm: Initially use stochastic gradient descent algorithm; Learning rate adjustment: Use the learning rate decay strategy to gradually reduce the learning rate to improve the stability of training; dynamically adjust the learning rate, specifically ReduceLROnPlateau: if the performance of the validation set does not improve after several epochs, reduce the learning rate; Step 33: Strategy to prevent overfitting Early stopping: Monitor the loss or accuracy of the validation set during training. If the validation set performance does not improve within a certain number of epochs, stop training early to prevent overfitting. Data enhancement: Implement real-time data enhancement technology during the training phase to expand the training data space and improve the model's generalization ability; Validation set evaluation: After each epoch, the model performance is evaluated on the validation set, and the training and validation loss and accuracy of each epoch are recorded to intuitively monitor the model training process; Step 34: Model evaluation and saving In the validation phase, performance is evaluated using metrics such as accuracy, precision, recall, and F1 score, and the optimal model is saved, that is, the model weight that achieves the lowest loss or highest accuracy on the validation set.

5. The method for intelligent recognition of space target images according to claim 1, characterized in that: The specific contents of the model evaluation and optimization in step 4 are as follows: Step 41: Evaluate model performance on the validation data set. Use the validation set to test the model to obtain the performance of the model on unseen data. Step 42: Use confusion matrix to analyze recognition accuracy: Construct a confusion matrix, where the rows represent the actual categories and the columns represent the predicted categories. For a binary classification problem, it contains the following elements: TP: True positive class, correctly identified; TN: True negative class, correctly excluded; FP: False Positive, incorrectly identified as a target; FN: False negative class, missed detection as target; Analysis: Calculate performance metrics including precision, recall, and F1 score: Identify the strengths and weaknesses of the model on different categories; Step 43: Optimize the network structure according to the evaluation results: find out the deficiencies of the network in certain categories from the confusion matrix and performance index analysis, and improve the feature extraction capability by adjusting the depth of the network, and adjust the convolution kernel to better capture details or global information; Step 44: Improve hyperparameters: adjust the learning rate according to the actual structure. A learning rate that is too high may lead to unstable convergence, while a learning rate that is too low may lead to slow convergence. Adjust batch size: Small batches help more frequent weight updates, but increase training time; large batches are more stable, but require more memory. Step 45: Regularization method: Use Dropout to prevent the model from overfitting and increase the robustness of the model by randomly ignoring neurons and their connections; L2 regularization: Add a weight decay term to the loss function to limit the numerical range of the weight; Step 45, model training strategy optimization: after adjustment in steps 44 and 45, the model is retrained using the optimized training strategy; Step 46: Iterative evaluation and optimization: After multiple evaluation and adjustment iterations, the model performance can be continuously improved.

6. The method for intelligent recognition of space target images according to claim 1, characterized in that: The specific steps of the real-time recognition application of step 5 are as follows: Step 51, model deployment: Convert the model trained in step 4 into a format suitable for the target platform, and use tools such as TensorRT to optimize the model to improve reasoning speed and efficiency; configure the operating environment of the target platform, including installing the necessary deep learning framework and dependent libraries; deploy the model on the platform for real-time call by the program; Step 52, receiving and processing real-time images: receiving spatial image data in real time and performing image preprocessing before inputting into the model; Step 53: Image recognition and result output: Use the deployed model to infer the received real-time image and output the classification label and the related probability value or confidence level; Step 54: Visualize the results: Use an image processing algorithm to overlay the recognition results on the original image data, including drawing a bounding box, label text or heat map on the image data; implement a user interface for real-time data display and update; display methods include desktop application, web page or direct output on the console.

Images