A dimm system star double imaging accurate identification method based on YOLO

Abstract

This invention discloses a precise stellar dual-image recognition method based on the YOLO-based DIMM system, belonging to the field of atmospheric measurement technology. The method includes: acquiring stellar samples; controlling the start of a rotating telescope system to stably track a designated star; forming a dual-image of the same star on a CCD detector using the first and second wedge mirrors of the rotating telescope system; acquiring several sample sets; grouping and labeling the samples; training, validating, and improving the model; importing the improved lightweight YOLO neural network model T into the DIMM system's image processing program; receiving the acquired dual-image stellar images in real time; completing stellar target recognition and dual-image frame localization; and outputting frame coordinates for star image centroid extraction and the variance σ of the dual-image centroid displacement. 2 The atmospheric coherence length is calculated and then quantitatively determined. This invention solves the problem that existing star identification methods cannot be adapted to the dual imaging mechanism of the DIMM system.

Description

Technical Field

[0001] This invention belongs to the field of atmospheric measurement technology, specifically relating to a precise method for dual-image recognition of stars based on the YOLO-based DIMM system. Background Technology

[0002] Atmospheric coherence length It is a core parameter characterizing the intensity of atmospheric turbulence, directly determining the performance of systems such as astronomical observation, laser communication, and adaptive optics. The Differential Image Motion Monitor (DIMM) is currently the primary instrument for measuring atmospheric coherence length. The core principle of this mainstream equipment is as follows: using double wedge mirrors to form a pair of images of the same star, acquiring binary star images through a CCD detector, calculating the variance of the centroid displacement of the binary star images, and then calculating the atmospheric coherence length. .

[0003] Stellar target identification is a prerequisite for accurate measurements by differential imaging motion monitoring instruments; the accuracy and stability of identification directly affect... The calculation results. Traditional differential image motion monitors mostly employ image processing methods such as adaptive threshold segmentation, morphological processing, and template matching, which have significant limitations in complex atmospheric environments.

[0004] In scenarios with strong turbulence and thick clouds, star images are prone to distortion, jitter, and blurring. Traditional methods have poor anti-interference capabilities and low recognition accuracy.

[0005] Failure to effectively utilize the homology of double wedge mirror imaging for paired correlation detection can easily lead to binary star matching errors, missed detections, or false alarms, directly resulting in distortion of the subsequent calculation of the centroid displacement variance of the binary star image.

[0006] It has poor adaptability to changes in observation conditions (such as exposure time, target magnitude, and turbulence intensity), making it difficult to meet the real-time and high-precision measurement requirements of the DIMM system.

[0007] Current neural network-based star recognition methods are mostly geared towards star sensor attitude determination tasks, primarily targeting single-star image detection. They fail to consider the geometric and brightness correlations between two star images in a DIMM system, nor do they model star image deformation and dynamic characteristics caused by atmospheric turbulence. Furthermore, they lack effective integration with atmospheric coherence length inversion processes. Therefore, their training samples and network structures are unsuitable for actual DIMM measurement scenarios and cannot be directly transferred and used.

[0008] In summary, the existing technology still lacks a highly robust identification scheme specifically designed for DIMM systems that can achieve accurate synchronous identification of two star images in complex atmospheric backgrounds and is deeply integrated with atmospheric coherence length inversion algorithms. This has become a key bottleneck restricting measurement accuracy and system stability. Summary of the Invention

[0009] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0010] A precise method for dual-stellar imaging identification in a YOLO-based DIMM system includes:

[0011] S1, collect stellar samples; control the turntable telescope system to start, stably track the designated star, and form a dual imaging image of the same star on the CCD detector based on the first and second wedge mirrors of the turntable telescope system, and collect several sample sets;

[0012] S2, Sample Grouping: The samples are randomly grouped into training and validation sets according to a preset ratio to ensure that the scene distribution of the training and validation sets is consistent.

[0013] S3, Sample labeling; Labeling the training set data to obtain the labeled data SA6;

[0014] S4, Model Training; For data SA6, an improved lightweight YOLO neural network model is selected for training, outputting accurate frame localization for dual imaging;

[0015] S5, Model Validation and Improvement: Input the validation set samples into the trained improved lightweight YOLO neural network model T to verify the performance of the improved lightweight YOLO neural network model;

[0016] S6, Model Application: The improved lightweight YOLO neural network model T, after meeting the standards, is imported into the image processing program of the DIMM system. It receives dual-image stellar images acquired by the CCD detector in real time, performs stellar target recognition and dual-image frame localization, and outputs frame coordinates for star image centroid extraction and dual-image centroid displacement variance σ. 2 Calculations were performed, and finally, the atmospheric coherence length was quantitatively calculated. .

[0017] The present invention has the following beneficial effects:

[0018] (1) Highly targeted, designed specifically for dual-imaging measurement of DIMM system; Based on the dual-wedge mirror homogeneous dual-star image structure of DIMM system, this invention constructs a multi-scene stellar dual-imaging sample set and adopts paired association annotation, enabling the improved lightweight YOLO neural network model to learn the spatial coupling relationship of dual-star images, ensuring a high degree of matching between the identification object and the atmospheric coherence length measurement requirements from the data level, and solving the problem that existing star identification methods cannot adapt to the dual-imaging mechanism of DIMM system.

[0019] (2) Significantly improved recognition accuracy and robustness under complex atmospheric conditions; By introducing a turbulence distortion feature enhancement module into the improved lightweight YOLO neural network model, and combining sample training under strong turbulence, thick clouds, different star magnitudes and different exposure conditions, this method can effectively suppress problems such as star image jitter, dispersion and noise interference. It can still achieve stable recognition of dual star imaging under harsh observation conditions. The recognition accuracy and positioning accuracy are far superior to traditional threshold segmentation, morphological processing and general target detection methods.

[0020] (3) Avoid mismatch of binary star images from the algorithm level and improve measurement reliability. This invention adds a binary star image pairing detection branch to the network structure and adds pairing constraints and spacing constraints to the loss function, forcing the model to output one-to-one corresponding binocular image frames, eliminating problems such as mismatch of different stars, missed detection of single stars, and misdetection of binary stars, and fundamentally avoiding the distortion of the calculation of the centroid displacement variance of binary star images caused by imaging pairing errors, thus greatly improving the reliability and consistency of atmospheric coherence length measurement results.

[0021] (4) Excellent real-time performance, meeting the online measurement requirements of DIMM system; The present invention adopts an improved lightweight YOLO structure, which achieves fast inference while ensuring high accuracy. The recognition time of a single frame image is ≤30ms, which can meet the high frame rate and real-time measurement requirements of DIMM system, avoid measurement errors caused by recognition delay, and improve the overall response speed of the system.

[0022] (5) The identification results are deeply coupled with the measurement process, resulting in higher measurement accuracy. The improved lightweight YOLO neural network model of this invention directly outputs the pixel-level bounding box coordinates and pairing labels of stellar dual imaging, which can be seamlessly connected to the subsequent centroid extraction, displacement variance calculation and atmospheric coherence length solution process, reducing the error accumulation caused by intermediate processing steps, realizing the integration of identification, positioning and measurement, and significantly improving the measurement accuracy and stability of the DIMM system.

[0023] (6) It has wide environmental adaptability and can work stably under various observation conditions. By setting the sample coverage to different CCD detector target positions, exposure times, star magnitudes, weather conditions and atmospheric turbulence intensity, the improved lightweight YOLO neural network model has extremely strong generalization ability and can maintain stable performance in the ever-changing outdoor test environment, thus expanding the applicable scenarios and working range of the DIMM system. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the structure of the turntable telescope system, where 1-two-dimensional turntable, 2-first wedge mirror, 3-second wedge mirror, 4-telescope, and 5-CCD detector;

[0025] Figure 2The diagram shows sample annotations using LabelImg (an open-source graphical image annotation tool); where (a) is a sample image of stars annotated at night using LabelImg software, and (b) is a sample image of stars annotated during the day using LabelImg software. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0027] The purpose of this invention is to overcome the shortcomings of existing technologies in DIMM systems, such as low star recognition accuracy, easy confusion between dual images, poor adaptability, and the inability of existing neural network star recognition methods to meet the requirements of DIMM systems. This invention provides a precise star dual-image recognition method for DIMM systems based on YOLO (a real-time target detection algorithm), achieving synchronous and accurate star dual-image recognition under complex atmospheric backgrounds and different scenarios (different turbulence, different exposure times, different weather conditions), and improving the atmospheric coherence length of DIMM systems. The precise calculations provide reliable star positioning data, improving the measurement accuracy and real-time performance of the DIMM system.

[0028] This invention presents a precise stellar dual-imaging recognition method for YOLO-based DIMM systems, used for the identification and extraction of stellar targets in complex atmospheric backgrounds and different scenes within DIMM (Differential Imaging Motion Measurement) systems, with atmospheric coherence length as the determining factor. The calculation provides accurate star image data support, including the following steps: star sample acquisition, sample annotation, model training and model prediction; star sample acquisition and sample annotation are designed for the same star dual imaging generated by the DIMM system dual wedge mirrors, and the output of model prediction is the accurate frame positioning of the dual imaging, which is used for subsequent star image barycenter extraction and star image barycenter displacement variance calculation.

[0029] The YOLO-based DIMM system-based method for accurate stellar dual-imaging identification is based on a rotating telescope system. This system comprises a two-dimensional rotating platform 1, a first wedge mirror 2, a second wedge mirror 3, a telescope 4, and a CCD (charge-coupled device) detector 5. The rotating telescope system can stably track a specified star, and through the refraction of the two wedge mirrors, forms a dual-imaging image of the same star on the CCD detector 5 (adapting to the dual-pupil measurement principle of the DIMM system). The method includes:

[0030] S1, collecting stellar samples; including:

[0031] S1.1 controls the turntable telescope system to make a small offset so that the dual imaging images of the same star appear simultaneously at any different position on the CCD detector 5 target surface. The dual imaging images of the star are acquired and saved to obtain a star sample set SA1 covering different positions on the CCD detector 5 target surface. This is used to improve the adaptability of the lightweight YOLO neural network model to the star image position offset and to adapt to the small attitude changes during the DIMM system tracking process.

[0032] S1.2, change the exposure time of CCD detector 5 (the exposure time range is controlled within 1~9ms, not exceeding the atmospheric coherence time τ, to avoid recognition errors caused by star image blurring), collect and save star dual imaging images under different exposure times, and obtain sample set SA2, which is used to adapt to the strict requirements of atmospheric coherence length measurement on exposure time in the DIMM system, and ensure that the recognition accuracy matches the measurement accuracy.

[0033] S1.3, control the turntable telescope system to track stars of different brightness (covering the range of commonly detected magnitudes in the DIMM system), acquire dual-image images of stars of different brightness and save them to obtain sample set SA3, which is used to improve the robustness of the improved lightweight YOLO neural network model in recognizing stars of different magnitudes.

[0034] S1.4 tracks the same star under different weather conditions (clear sky, thin clouds, thick clouds), collects and saves dual-image images of the star under different weather conditions, and obtains sample set SA4, which is used to adapt to the DIMM measurement needs in complex outdoor observation environments.

[0035] S1.5, stellar dual-image acquisition and storage were performed under different atmospheric turbulence conditions (strong turbulence, moderate turbulence, and weak turbulence) to obtain sample set SA5. This specifically addresses the star image distortion and jitter issues caused by atmospheric turbulence, ensuring accurate dual-image identification under various turbulence conditions, and providing atmospheric coherence length. This provides the foundation for accurate calculations.

[0036] S2, Sample grouping; including: dividing all collected samples into training set and validation set according to a preset ratio (e.g., 7:3); the samples are all dual-image (paired) samples of the same star, and the sample scenes in the training set and validation set are not repeated, so as to avoid overfitting of the improved lightweight YOLO neural network model.

[0037] S3, Sample Labeling; includes: labeling the training set data using the professional labeling tool LabelImg v1.8.1 to obtain the labeled data SA6; the labeling method is: drawing bounding boxes for the double-image of the same star in each sample image and labeling them with paired labels, with labeling accuracy down to the pixel level, providing accurate labels for training the improved lightweight YOLO neural network model and avoiding subsequent measurement errors caused by confusion of double-image.

[0038] S4, Model Training; including: training an improved lightweight YOLO neural network model on data SA6. This improved lightweight YOLO neural network model is obtained by adapting to the DIMM scene based on the standard lightweight YOLO structure. This includes: adding a turbulence distortion feature enhancement module to the feature fusion layer to improve the feature representation ability of jittery and diffuse star images; adding a dual-star image pairing association branch to the detection head to force the output of paired, homogeneous star targets; and introducing a comprehensive loss function including spacing constraints to ensure that the dual-image spacing conforms to the prior knowledge of the dual-wedge mirror optical structure. During training, the learning rate, batch size, and number of iterations are selected as core debugging parameters. Parameter optimization is performed by minimizing the loss function to obtain the trained improved lightweight YOLO neural network model T. The trained improved lightweight YOLO neural network model T balances recognition accuracy and inference speed, adapts to the real-time measurement requirements of the DIMM system, and can quickly achieve synchronous detection and localization of dual-star images.

[0039] S5, Model Validation and Improvement; including: using the validation set to validate the inference accuracy and inference speed of the trained improved lightweight YOLO neural network model T, with validation metrics including: dual-image recognition accuracy, single-image inference time, and star frame positioning error; adjusting model parameters (such as learning rate decay coefficient and number of iterations) based on the validation results to improve the trained improved lightweight YOLO neural network model T until it meets the actual requirements of atmospheric coherence length measurement in the DIMM system.

[0040] S6, Model Application; including: importing the improved lightweight YOLO neural network model T (after meeting the standards) into the image processing program of the DIMM system, receiving stellar dual-image images acquired by the CCD detector 5 in real time, completing stellar target recognition and dual-image frame localization, and outputting frame coordinate data for star image centroid extraction and dual-image centroid displacement variance σ. 2 The calculation ultimately yielded the atmospheric coherence length. The precise solution.

[0041] The present invention will be described in detail below with reference to the embodiments.

[0042] This invention relates to a YOLO-based DIMM system stellar dual-imaging precise identification method, which utilizes a turntable telescope system. The turntable telescope system includes: a customized two-dimensional turntable 1 (positioning accuracy: ±30″), a telescope 4 (Midea: LX200, focal length: 2438mm, aperture: 305mm), a first wedge mirror 2, a second wedge mirror 3 (wedge angle: 2′, material: K9), and a CCD detector 5 (pixels: 2592×2048, pixel size: 4.8μm, frame rate: not less than 100fps). Along the optical path, the first wedge mirror 2 and the second wedge mirror 3 are positioned at the front end of the telescope 4, and the CCD detector 5 is connected to the rear end of the telescope 4. The first wedge mirror 2, the second wedge mirror 3, and the telescope 4 are mounted as a whole on the upper part of the two-dimensional turntable 1. This system can stably track a specified star, forming two identical images of the same star through refraction by the first wedge mirror 2 and the second wedge mirror 3, adapting to the dual-pupil measurement principle of the DIMM system.

[0043] Software environment: LabelImg v1.8.1 is used for annotation; PyTorch (deep learning framework) 1.12.0 and Python (programming language) 3.9 are used for model training; NVIDIA professional graphics card is used for hardware acceleration; the image processing program is written based on QT (QT framework, cross-platform application development framework) and is adapted to the real-time data transmission of the DIMM system.

[0044] The present invention provides a precise method for dual-image stellar identification based on YOLO in a DIMM system, comprising the following steps:

[0045] S1, collect stellar samples;

[0046] The control turntable telescope system is activated and stably tracks a designated star (selected magnitude range: -1 to 3, adapted to the DIMM system's detection capabilities). Based on the first wedge mirror 2 and the second wedge mirror 3, a dual-image of the same star is formed on the CCD detector 5. Several (e.g., 5) sample sets are collected in the following manner, with multiple (e.g., 1000) images collected in each sample set, for a total of 5000 paired samples:

[0047] S1.1, Sample Set SA1 Acquisition: Control the two-dimensional turntable 1 to make a slight offset so that the dual imaging images of the same star appear simultaneously at different positions on the CCD detector target surface (covering the center, edge, and other areas of the target surface). Acquire the images and save them in the naming format "SA1_positionX_serial number.jpg" to ensure that the samples cover different target surface positions, which is used to improve the adaptability of the improved lightweight YOLO neural network model to position offset.

[0048] S1.2, Sample Set SA2 Acquisition: Fix the position of the 2D turntable 1 and track the star, change the exposure time of the CCD detector 5, and set the exposure time to 1ms, 3ms, 5ms, 7ms, and 9ms (all not exceeding the atmospheric coherence time τ≈10ms to avoid star image blurring). Acquire 200 images for each exposure time and save them as "SA2_exposure timeX_serial number.jpg" to adapt to the exposure requirements of different measurement scenarios of the DIMM system.

[0049] S1.3, Sample set SA3 acquisition: Control the two-dimensional turntable 1 to track stars of different brightness (negative magnitude, first magnitude, second magnitude, and third magnitude), with a fixed exposure time of 5ms, and acquire 200 dual-image images of each star, saving them as "SA3_magnitude X_serial number.jpg" to improve the robustness of the improved lightweight YOLO neural network model in recognizing stars of different magnitudes.

[0050] S1.4, Sample Set SA4 Acquisition: Select the same star and acquire 300+ dual-image images under various weather conditions such as clear sky, thin clouds, and thick clouds, with a fixed exposure time of 5ms. Save the images as "SA4_weatherX_serial number.jpg" to adapt to complex outdoor observation environments.

[0051] S1.5, Sample Set SA5 Acquisition: Select the same star, under different atmospheric turbulence conditions, with a fixed exposure time of 5ms, and acquire 300+ dual-image images under each atmospheric turbulence condition, saving them as "SA5_turbulenceX_serial number.jpg" to specifically solve the problem of star image distortion recognition caused by atmospheric turbulence.

[0052] S2, sample grouping;

[0053] By using a random grouping method, 5000 samples are divided into training and validation sets according to a certain ratio to ensure that the scene distribution of the training and validation sets is consistent, avoid overfitting of the improved lightweight YOLO neural network model, and ensure that the validation results can truly reflect the performance of the model in real-world scenarios.

[0054] S3, Sample Labeling; includes:

[0055] Open the LabelImg annotation tool, import the training set samples, and use the rectangular bounding box annotation method to draw bounding boxes for the double images of the same star in each image; at the same time, label the two bounding boxes with paired labels to distinguish the double images of the same star from the images of different stars. After the annotation is completed, save it as a .txt file to obtain the annotated data SA6, which is used for model training.

[0056] S4, Model Training;

[0057] An improved lightweight YOLO neural network model was selected, comprising an input layer, a backbone feature extraction network, a neck feature fusion network, and a head detection network connected sequentially. A turbulence distortion enhancement module was added to the feature fusion network, and a binary star image pairing constraint branch was added to the head detection network. The hidden layers of the backbone feature extraction network and the neck feature fusion network uniformly adopted the ReLU (Rectified Linear Unit) activation function, while the output layer of the head detection network adopted a sigmoid function. The labeled data S6 was input into the improved lightweight YOLO neural network model, and the following debugging parameters were set: initial learning rate 0.001, adaptive moment estimation (Adam) optimizer, batch size = 16, number of iterations = 100 epochs, and the loss function was a comprehensive loss function combining CIoU (Complete Intersection over Union, a bounding box regression loss function) loss with binary star spacing constraints and pairing consistency constraints (adapting to target detection and localization requirements, superior to the cross-entropy loss function). During training, the training loss and validation loss were monitored in real time. Training was stopped when the validation loss did not decrease for 10 consecutive epochs, resulting in the trained improved lightweight YOLO neural network model T.

[0058] S5, Model Validation and Improvement;

[0059] The validation set samples are input into the trained improved lightweight YOLO neural network model T to verify its performance. The core validation metrics include: dual-image recognition accuracy, single-image inference time, and star frame positioning error. If the validation results do not meet the requirements, the parameters are adjusted and the model is retrained until the performance of the trained improved lightweight YOLO neural network model T meets the requirements for atmospheric coherence length measurement in the DIMM system.

[0060] S6, Model Application;

[0061] The improved lightweight YOLO neural network model T, after meeting the standards, is imported into the image processing program of the DIMM system. It receives stellar dual-images acquired by the CCD detector 5 in real time, completes stellar target recognition and dual-image frame localization, and outputs the coordinates of the two image frames (two sets of target boxes) P1 and P2. The coordinates of the upper left of P1 are... The coordinates of the lower right of P1 are The coordinates of the upper left of P2 are P2 lower right coordinates Based on the coordinates of the two sets of target boxes P1 and P2, the sub-pixel centroid coordinates of the corresponding star images are calculated. Multiple frames of images are continuously acquired, and the relative offset of the two sets of sub-pixel centroid coordinates is statistically analyzed. The variance of the centroid displacement of the two star images is then calculated using a known algorithm for atmospheric turbulence measurement in the DIMM system. Then Substitute the following atmospheric coherence length The solution formula was used to finally determine the atmospheric coherence length. Quantitative calculations enable accurate, real-time measurement of the DIMM system.

[0062] ;

[0063] ;

[0064] ;

[0065] ;

[0066] in, The zenith angle of the observed star. The path atmospheric coherence length The atmospheric coherence length is the length along the line connecting the sub-apertures. The length of atmospheric coherence in the direction perpendicular to the line connecting the two points. Let be the variance of the angular undulation along the direction of the line connecting the child pupils. Let V be the variance of the angle of arrival along the direction perpendicular to the line connecting the sub-pupils. For the telescope's focal length, The wavelength corresponding to the coherence length. The diameter of the telescope's sub-pupil. This is the distance between the centers of the telescope's sub-pupils.

[0067] Figure 2 This is a schematic diagram illustrating sample annotations using the LabelImg software (an open-source graphical image annotation tool). Figure 2 Image (a) is a sample image of stars labeled at night using LabelImg software. Figure 2 (b) is a sample image of daytime stars labeled using LabelImg software.

[0068] The above description is merely an embodiment of the present invention and does not limit the scope of the invention. Any equivalent structural or procedural transformations made based on the description and drawings of this invention, or direct or indirect applications in other related system fields, are similarly included within the protection scope of this invention. Contents not described in detail in this specification are prior art known to those skilled in the art.

Claims

1. A precise method for dual-stellar imaging identification based on YOLO-based DIMM systems, characterized in that, include: S1, collect stellar samples; control the turntable telescope system to start, stably track the designated star, and form a dual imaging image of the same star on the CCD detector based on the first and second wedge mirrors of the turntable telescope system, and collect several sample sets; S2, Sample Grouping: The samples are randomly grouped into training and validation sets according to a preset ratio to ensure that the scene distribution of the training and validation sets is consistent. S3, Sample labeling; Labeling the training set data to obtain the labeled data SA6; S4, Model Training; For data SA6, an improved lightweight YOLO neural network model is selected for training, outputting accurate frame localization for dual imaging; S5, Model Validation and Improvement: Input the validation set samples into the trained improved lightweight YOLO neural network model T to verify the performance of the improved lightweight YOLO neural network model; S6, Model Application: The improved lightweight YOLO neural network model T, after meeting the standards, is imported into the image processing program of the DIMM system. It receives dual-image stellar images acquired by the CCD detector in real time, performs stellar target recognition and dual-image frame localization, and outputs frame coordinates for star image centroid extraction and dual-image centroid displacement variance σ. 2 Calculations were performed, and finally, the atmospheric coherence length was quantitatively calculated. .

2. The method for accurate stellar dual-image identification based on YOLO in a DIMM system according to claim 1, characterized in that, The rotating telescope system includes a two-dimensional rotating platform, a first wedge mirror, a second wedge mirror, a telescope, and a CCD detector. Along the optical path, the first and second wedge mirrors are set at the front end of the telescope, and the CCD detector is connected to the rear end. The first wedge mirror, the second wedge mirror, and the telescope are mounted on the upper part of the two-dimensional rotating platform. The rotating telescope system stably tracks a designated star and, through the refraction of the first and second wedge mirrors, forms a dual-image of the same star on the CCD detector.

3. The method for accurate stellar dual-image identification based on YOLO in a DIMM system according to claim 2, characterized in that, The method for collecting the sample set in S1 is as follows: S1.1, control the turntable telescope system to make a slight offset so that the dual imaging images of the same star appear simultaneously at any different position on the CCD detector target surface, acquire and save the dual imaging images of the star, and obtain the star sample set SA1 covering different positions on the CCD detector target surface. S1.2, change the exposure time of the CCD detector, acquire and save star dual imaging images under different exposure times, and obtain sample set SA2; S1.3, control the turntable telescope system to track stars of different brightness, acquire dual imaging images of stars of different brightness and save them to obtain sample set SA3; S1.4: Track the same star under different weather conditions, collect and save dual-image images of the star under different weather conditions, and obtain sample set SA4; S1.5: Two-image stellar images were acquired and saved under different atmospheric turbulence conditions to obtain sample set SA5.

4. The method for accurate stellar dual-image identification based on YOLO in a DIMM system according to claim 1, characterized in that, In S2, the samples are dual-image samples of the same star, and the sample scenes in the training set and the validation set are not repeated.

5. The method for accurate stellar dual-image identification based on YOLO in a DIMM system according to claim 3, characterized in that, In S3, the annotation method is as follows: for each sample image, the double-image of the same star is labeled separately, and the two frames are labeled with paired labels to distinguish the double-image of the same star from the images of different stars; after the annotation is completed, it is saved to obtain the annotated data SA6.

6. The method for accurate stellar dual-image identification based on YOLO in a DIMM system according to claim 5, characterized in that, In S4, the improved lightweight YOLO neural network model is obtained by making DIMM scene adaptation improvements on the basis of the standard lightweight YOLO structure. These improvements include: adding a turbulence distortion feature enhancement module to the feature fusion layer to improve the feature representation ability of jittery and diffuse star images; adding a dual star image pairing association branch to the detection head to force the output of paired stars of the same origin; and introducing a comprehensive loss function that includes spacing constraints to make the dual imaging spacing conform to the prior of the dual wedge mirror optical structure.

7. The method for accurate stellar dual-image identification based on YOLO in a DIMM system according to claim 1, characterized in that, In S4, during the training process, the learning rate, batch size, and number of iterations are selected as core debugging parameters. The parameters are optimized by minimizing the loss function to obtain the trained improved lightweight YOLO neural network model T.

8. The method for accurate stellar dual-image identification of a YOLO-based DIMM system according to claim 1, characterized in that, In S4, the improved lightweight YOLO neural network model includes an input layer, a backbone feature extraction network, a neck feature fusion network, and a head detection network connected in sequence. The hidden layers of the backbone feature extraction network and the neck feature fusion network uniformly adopt the linear rectified unit activation function, while the output layer of the head detection network adopts the sigmoid function.

9. The method for accurate identification of dual stars in a YOLO-based DIMM system according to claim 6, characterized in that, In S5, the verification metrics include: dual-image recognition accuracy, single-image inference time, and star frame positioning error.

10. The method for accurate identification of dual stars in a YOLO-based DIMM system according to claim 9, characterized in that, S6 include: The improved lightweight YOLO neural network model T, after meeting the standards, is imported into the image processing program of the DIMM system. It receives dual-image stellar images acquired by the CCD detector in real time, performs stellar target recognition and dual-image frame localization, and outputs the coordinates of two sets of target frames P1 and P2. Based on the coordinates of the two sets of target frames P1 and P2, the sub-pixel centroid coordinates of the corresponding star images are calculated. Multiple frames are continuously acquired, and the relative offset of the two sets of sub-pixel centroid coordinates is statistically analyzed to calculate the variance of the dual-image centroid displacement. Then Substitute into the following formula to obtain the atmospheric coherence length. Quantitative calculation: ； ； ； ； in, The zenith angle of the observed star. The path atmospheric coherence length The atmospheric coherence length is the length along the line connecting the sub-apertures. The length of atmospheric coherence in the direction perpendicular to the line connecting the two points. Let be the variance of the angular undulation along the direction of the line connecting the child pupils. Let V be the variance of the angle of arrival along the direction perpendicular to the line connecting the sub-pupils. For the telescope's focal length, The wavelength corresponding to the coherence length. The diameter of the telescope's sub-pupil. This is the distance between the centers of the telescope's sub-pupils.

Images