Optimized YOLO model-based flow field key structure detection and feature point extraction method

Abstract

The invention discloses a flow field key structure detection and feature point extraction method based on an optimized YOLO model, and the method comprises the steps: S1, obtaining flow field time sequence schlieren images of an air-breathing aircraft under different working conditions through a high-speed schlieren collection system in a wind tunnel test; s2, preprocessing the flow field time sequence schlieren image to obtain a data set corresponding to the strong shock wave of the isolation section of the air-breathing aircraft, and dividing the data set into a training set, a verification set and a test set in proportion; s3, constructing an optimized YOLO target detection model, and obtaining an optimal weight model after iterative training and verification; and S4, adopting the optimal weight model to complete target detection and feature point extraction of the test set or the schlieren image to be detected. According to the method, the target area is focused through area cutting, meanwhile, a targeted data enhancement strategy of Gaussian noise, brightness adjustment, contrast ratio adjustment and saturation is designed, a high-quality annotation data set is constructed, and the model detection precision is effectively improved.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of intelligent fluid dynamics and computer vision. More specifically, this invention relates to a method for detecting key structures and extracting feature points in a flow field based on an optimized YOLO model. Background Technology

[0002] The isolator section of an air-breathing aircraft serves as a crucial connecting component between the inlet and the combustion chamber. The dynamic evolution characteristics of the strong shock waves within it directly impact the inlet's flow capture efficiency, the combustion stability of the combustion chamber, and the overall aerodynamic stability of the aircraft. Accurately acquiring the location, morphological characteristics, and dynamic evolution parameters of the strong shock waves within the isolator section is a core prerequisite for optimizing aircraft aerodynamic design, suppressing flow instability, and ensuring flight safety. This has significant theoretical research value and engineering application implications.

[0003] Existing methods for detecting key structures and extracting features in flow fields suffer from numerous technical shortcomings, making it difficult to meet practical engineering needs: While CFD numerical simulation methods can achieve qualitative analysis of flow fields, their accuracy in reproducing the dynamic evolution characteristics of complex unsteady flow fields is insufficient, especially for structures with abrupt changes and dynamic deformation characteristics, such as strong shock waves, where prediction errors are large and they cannot provide accurate location and morphological data; Manual identification methods rely on the experience and judgment of experts in the field of fluid mechanics, which is not only inefficient but also highly subjective, resulting in poor consistency of identification results and making it difficult to meet the batch processing needs of massive transient schlieren images; Traditional image processing algorithms require manual setting of key parameters such as thresholds and filtering parameters, and parameter selection depends on expert experience, resulting in poor adaptability to schlieren images under different working conditions and imaging conditions, and weak generalization ability; Existing deep learning object detection models are mainly divided into single-stage and two-stage methods. Single-stage object detection algorithms usually have higher speed and lower memory consumption, but their accuracy is usually slightly lower than that of two-stage algorithms. Meanwhile, mainstream single-stage target detection algorithms (such as the YOLO series) are mainly designed for natural images. In the scenario of shock wave recognition in the isolation section of air-breathing aircraft, their prior box size and network structure are not adapted to the geometry, specific size distribution characteristics and dynamic characteristics of key structures in the aircraft flow field (such as strong shock waves), resulting in low positioning accuracy and high false detection and missed detection rates for strong shock waves. At the same time, existing technologies lack a systematic extraction scheme for the core feature points of key structures in the flow field, which cannot effectively capture the dynamic evolution law of strong shock wave structures and is difficult to support subsequent in-depth analysis such as flow field oscillation characteristics and pulsating pressure correlation.

[0004] With the development of high-speed photography technology, transient schlieren images can capture the dynamic evolution information of strong shock waves within isolated sections non-contactly, with a wide field of view and high frame rate, providing rich raw data for flow field analysis. However, how to optimize deep learning target detection models for the unique characteristics of key flow field structures, achieve rapid and accurate localization of strong shock waves, and extract core feature points that stably reflect their dynamic evolution has become a pressing technical challenge. Therefore, it is urgent to propose an optimized single-stage method and a corresponding feature point extraction method to overcome existing technical bottlenecks and meet the high-precision and high-efficiency requirements for the detection and analysis of key structures in aircraft flow fields. Summary of the Invention

[0005] One object of the present invention is to solve at least the above-mentioned problems and / or defects, and to provide at least the advantages described below.

[0006] To achieve these objectives and other advantages of the present invention, a method for detecting key structures and extracting feature points in a flow field based on an optimized YOLO model is provided, comprising: S1. In wind tunnel testing, a high-speed schlieren acquisition system is used to acquire time-series schlieren images of the flow field of an air-breathing aircraft under different operating conditions. S2. By preprocessing the flow field time-series schlieren image, a data set corresponding to the strong shock wave of the isolation section of the air-breathing aircraft is obtained, and the data set is divided into training set, validation set and test set according to the proportion. S3. Construct an optimized YOLO object detection model, and iteratively train and validate the YOLO object detection model using the training set and validation set. After training, the model parameters with the highest average accuracy (AP) in the validation set are taken as the optimal weight model. S4. Use the optimal weight model to complete the target detection and feature point extraction of the test set or the schlieren image to be detected; The optimization of the YOLO object detection model includes: prior box clustering optimization based on the k-means algorithm, lightweight adaptation adjustment of the original YOLO model network structure, and optimization selection of the loss function of the YOLO object detection model.

[0007] Preferably, in S1, the flow field temporal schlieren image is acquired by acquiring 3,000 original images for each working condition that can completely capture the dynamic evolution of the strong shock wave, with the acquisition frame rate set to 9,000 FPS.

[0008] Preferably, in S2, the preprocessing includes: S20. Crop out isolated segments from the original image, and the resolution of the cropped isolated segments is uniformly 202×59 pixels; S21. Use the labeling tool LabelImg to label 300 randomly selected isolation segment images as labeling samples. The labeling content includes: the coordinates of the minimum bounding rectangle detection box of the strong shock wave target and the category label. The minimum bounding rectangle detection box should completely wrap around the core area of ​​the strong shock wave and fit the edge of the strong shock wave. The category label is strong shock wave in the isolation segment. S22. An image enhancement strategy is used to enhance the labeled samples nine times to obtain a set of 3,000 enhanced labeled images. The image enhancement strategy includes the following enhancement methods: adding Gaussian noise with a standard deviation of 0.02 to 0.08, and adjusting brightness, contrast, and saturation with adjustment factors of 0.7 to 1.3. The image enhancement strategy refers to applying each enhancement method independently or in combination.

[0009] Preferably, the prior box clustering optimization based on the kmeans algorithm refers to: based on the narrow shape and size distribution of the strong shock wave in the isolation segment, after extracting the width and height data corresponding to all labeled boxes in the training set, the kmeans clustering algorithm is used to perform clustering analysis on the width and height data; In the clustering process, the intersection-union ratio (IoU) between the bounding boxes and the cluster centers was used as the distance metric. The elbow rule was used to verify the clustering error under different numbers of clusters k. It was found that the matching degree between the prior boxes and the size of the strong shock wave target was the highest when k=6. The corresponding six sets of highest priority prior box sizes were obtained as follows: (54,136), (62,144), (60,151), (65,148), (65,151), (70,159). Replace the default prior boxes of the YOLO model with the six sets of highest priority bounding box sizes to obtain the initial anchor points for the model fitting target.

[0010] Preferably, lightweight adaptation and adjustment of the original YOLO model network structure refers to: The backbone network in the YOLO network is configured to include: 2 CBL layers + 3 CSP layers + 1 CBL layer; The feature fusion module for the neck structure in the YOLO network is set to be derived from the last layer CBL and the third-to-last layer CSP of the backbone network. The YOLO network head structure retains two scales of detection heads, with three prior boxes set for each scale.

[0011] Preferably, the loss function of the YOLO object detection model is the CIoU loss function.

[0012] Preferably, in S4, for the detection box coordinates (x1, y1, x2, y2) of the strong shock wave target in each image, the feature point is selected as a coordinate point (x1, y2) that can reflect the radial position change of the strong shock wave within the isolation section.

[0013] The present invention has at least the following beneficial effects: Firstly, this invention combines the structural characteristics of the strong shock wave in the isolated segment to be detected, focuses the target area through region clipping, and designs Gaussian noise and adjusts the brightness, contrast, and saturation of the targeted data augmentation strategy to construct a high-quality labeled dataset, effectively improving the model detection accuracy.

[0014] Secondly, this invention addresses the specific size distribution and elongated shape of strong shock waves in the isolation section by using the k-means clustering algorithm to cluster the width and height data of the labeled boxes and setting 6 sets of optimal fitting prior boxes. This solves the problems of low positioning accuracy and slow model training convergence caused by the mismatch between the prior boxes of the original YOLO model and the target size of the flow field, and significantly improves the initial matching accuracy of strong shock wave detection.

[0015] Third, this invention performs lightweight network pruning on the original YOLO network, achieving synergistic optimization of detection accuracy (AP=98.3%) and detection speed (64FPS), adapting to the rapid processing needs of massive transient schlieren images.

[0016] Fourth, based on the dynamic evolution characteristics of the strong shock wave in the isolation section and the engineering analysis requirements, this invention specifically selects the lower left corner point (x1, y2) of the detection frame as the core feature point. This feature point can stably reflect the positional changes and dynamic characteristics of the strong shock wave, providing high-quality and high-reliability shock wave structural feature data for subsequent flow field oscillation analysis, aerodynamic performance correlation and other studies, thus solving the problem of the lack of a systematic extraction scheme for key structural feature points of the flow field in the existing technology.

[0017] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 The raw schlieren image of the strong shock transient in the isolation section before preprocessing Figure 1 ; Figure 3 To the original Figure 1 A schematic diagram of the core area after the red distinction has been cropped; Figure 4 To Figure 2 Example image of a strong shock wave target labeled in the image; Figure 5 For the original data before data augmentation Figure 2 ; Figure 6 To augment the original data Figure 2 The image showing the effect after contrast adjustment; Figure 7 To augment the original data Figure 2 The effect after adjusting the brightness; Figure 8 To augment the original data Figure 2 The effect after adding Gaussian noise; Figure 9 This is a schematic diagram showing the distribution of the k-means clustering of the dataset when k=6, based on the YOLO model optimized by this invention. Figure 10 This is a schematic diagram of the network structure; Figure 11 The original embodiment Figure 3 ; Figure 12 To use the optimized model of this invention to analyze the original Figure 3 A diagram illustrating the results after prediction; Figure 13 To use the original model for the original Figure 3 A diagram illustrating the results after prediction; Figure 14 The original embodiment Figure 4 ; Figure 15 To use the optimized model of this invention to analyze the original Figure 4 A diagram illustrating the results after prediction; Figure 16 To use the original model for the original Figure 4 A diagram illustrating the results after prediction. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0020] This invention provides a method for detecting key structures and extracting feature points in the flow field of aircraft based on an optimized YOLO model. It is particularly suitable for the rapid and accurate localization of key flow field structures such as strong shock waves and the extraction of core feature points in transient schlieren images of the isolation section of air-breathing aircraft. It can directly provide high-quality data support for subsequent research such as aircraft aerodynamic design optimization and flow field dynamic evolution analysis.

[0021] Furthermore, this invention achieves precise localization of strong shock waves in the isolation section of an air-breathing aircraft and extracts core feature points through data augmentation processing targeting data features, prior box clustering optimization, network structure adjustment, and loss function improvement. Figure 1 The diagram shown is a schematic representation of the process of this invention. The specific steps are as follows: Step 1: Construct a dataset of transient schlieren images of strong shock waves in the isolation section of an air-breathing aircraft. 1. Schreiber Image Acquisition and Preprocessing: A high-speed schreiber acquisition system was used to acquire flow field schreiber images of the air-breathing aircraft under different operating conditions (such as varying Mach number and angle of attack) in wind tunnel tests. The acquisition frame rate was set to 9000 FPS to ensure complete capture of the dynamic evolution process of strong shock waves. 3000 raw images were acquired for each operating condition.

[0022] To eliminate Figure 2 Irrelevant information such as the background and aircraft fuselage is removed from the original image, and regions are cropped to retain only those irrelevant details. Figure 3 The core region of the isolated segment (isolation segment) shown is cropped to a uniform resolution of 202×59 (width×height) pixels. This processing method will provide a basic condition for the model to effectively focus on the detection target region and reduce the computational complexity of subsequent training.

[0023] 2. Target annotation and label definition: Randomly select 300 preprocessed valid images as annotation samples. For example... Figure 4 As shown, LabelImg, a dedicated labeling tool for target detection, was used for annotation. The annotation content includes the coordinates of the minimum bounding rectangle detection box of the strong shock wave target and its category label: the detection box coordinates are based on the upper left corner of the image as the origin, and the target's position information is recorded as (x1, y1, x2, y2), where (x1, y1) is the coordinate of the upper left corner of the detection box, and (x2, y2) is the coordinate of the lower right corner of the detection box, ensuring that the detection box completely encompasses the core area of ​​the strong shock wave and fits the edge of the strong shock wave; the target's category label is uniformly marked as "isolated segment strong shock wave".

[0024] 3. Data Augmentation and Dataset Partitioning: To address the overfitting problem caused by insufficient labeled samples and improve the model's generalization ability and robustness, data augmentation was performed on 300 images and their corresponding labels.

[0025] In this step, due to the strong directionality and physical constraints of the supersonic shock wave flow field structure, using conventional image enhancement strategies such as random horizontal / vertical / mirror flipping will lead to the model learning incorrect physical associations. This will prevent the model from accurately capturing key target detection features such as the critical shock wave structure and vortex core location under complex background interference, thus reducing the model's training accuracy. Considering the relatively limited range of motion and relatively stable morphology of the strong shock wave in the isolation section, a targeted image enhancement strategy is designed and implemented: such as... Figure 5 The original shown Figure 2It can be designed to include adding contrast adjustments (such as...) Figure 6 As shown, the contrast adjustment factor is 0.7~1.3), and the brightness adjustment (such as...) Figure 7 As shown, the brightness adjustment factor is 0.7~1.3), Gaussian noise (such as... Figure 8 As shown, the standard deviation in Gaussian noise was set to 0.02~0.08, and the saturation was adjusted (saturation adjustment factor 0.7~1.3). Each enhancement method was applied independently or in combination, and a total of 9 enhancements were completed, resulting in a dataset of 3000 enhanced labeled images.

[0026] Subsequently, the dataset was randomly divided into a training set (2400 images), a validation set (300 images), and a test set (300 images) in a ratio of 8:1:1. The training set was used for model parameter learning, the validation set was used for hyperparameter optimization and model performance evaluation during the training process, and the test set was used for final method effectiveness verification.

[0027] Step 2: Optimization Design of the YOLO Object Detection Model 1. Prior Box Clustering Optimization Based on k-means Algorithm: The size of the prior boxes in the YOLO model directly affects the initial matching accuracy of the detection boxes and the model convergence speed. Considering the elongated shape and specific size distribution of the strong shock wave in the isolation segment, the width (w) and height (h) data of all labeled boxes in the training set were extracted. The k-means clustering algorithm was used to perform clustering analysis on the width and height data of the labeled boxes. During the clustering process, the intersection-over-union ratio (IoU) between the labeled boxes and the cluster centers was used as the distance metric. The elbow rule was used to verify the clustering error under different cluster sizes k (k=3~8). Considering both the accuracy of the initial matching boxes during model training and the complexity of the network, it was finally determined that k=6 resulted in the optimal clustering effect, at which point the matching degree between the prior boxes and the strong shock wave target size was highest. The final six sets of highest priority bounding box sizes were obtained: (54,136), (62,144), (60,151), (65,148), (65,151), and (70,159). These replaced the default prior boxes of the YOLO model, providing more suitable initial anchor points for model training. Figure 9 The distribution of k-means clustering in the dataset when k=6 is given. The circles in the figure represent the size (width and height) of the shock structures in the samples. Shock structures with similar sizes are clustered into a central cluster. The six colors correspond to samples with six centers in the k-means clustering. The six small black crosses are the six cluster centers, which are the final determined prior box sizes.

[0028] 2. Lightweight Network Structure Adaptation: Addressing the low resolution (202×59) and strong shock wave target characteristics of the isolated segment schlieren image, the YOLO model network structure was lightweighted and optimized to improve computational speed while maintaining detection accuracy. For example... Figure 10As shown, the optimized YOLO network's original backbone feature extraction network (also known as the backbone network, which is mainly responsible for feature extraction) consists of 2 CBL layers, 3 CSP layers, and 1 CBL layer. This structure can effectively capture shallow features such as the edge contours of strong shock waves, meeting the feature extraction requirements of strong shock wave targets. The feature fusion module of the neck network (also known as the neck structure, which is mainly responsible for feature fusion) is derived from the last CBL layer and the third-to-last CSP layer of the original backbone feature extraction network. These two paths are then subjected to feature fusion processing through upsampling. This feature pyramid structure enhances the extraction of multi-scale features of the target. The prediction head network (also known as the head structure, which is mainly responsible for outputting detection information through the detection head) retains prediction modules for two scales, with three prior boxes set for each scale.

[0029] 3. Loss Function Optimization: To improve the regression accuracy of strong shock wave detection boxes, the CIoU (Complete Intersection over Union) loss function is adopted as the regression loss for the detection boxes. This loss function comprehensively considers the overlapping area of ​​the detection boxes, the Euclidean distance between the center points, and the aspect ratio. Compared with traditional IoU loss and GIoU loss, it can more comprehensively and accurately evaluate the matching degree between the predicted box and the ground truth box, effectively solving the problem of insufficient regression accuracy for narrow shock wave structure targets.

[0030] Step 3: Optimize the training of the YOLO model and select the optimal weights. 1. The training environment is configured as follows: CPU Intel(R) Xero(R) E-2276M, GPU NVIDIA Quadro RTX5000, deep learning framework PyTorch1.2.

[0031] 2. Model Training and Optimal Weight Saving: Input the training set partitioned in Step 1 into the optimized YOLO model, initiate iterative training, and perform hyperparameter tuning and performance comparison. During training, the validation set is used to evaluate model performance, with evaluation metrics including precision, recall, and average precision (AP). Precision reflects the accuracy of the model's detection results (no false alarms), recall reflects the model's ability to capture real targets (no missed detections), and average precision (AP) comprehensively measures the model's overall detection performance. After training, the model with the highest AP on the validation set is selected as the optimal weight model for subsequent strong shock wave detection tasks. The hyperparameters of the optimal weight model are specifically set as follows: batch size = 32; initial learning rate = 1e-3, using a cosine annealing learning rate scheduling strategy, with the learning rate decaying periodically with each training round to balance the model's rapid convergence in the early stages with accurate optimization in the later stages; total number of training rounds = 100 rounds; the optimizer is Adam, with parameters β1 = 0.9, β2 = 0.999, and a weight decay coefficient = 5e-4, which suppresses model overfitting through weight decay.

[0032] Step 4: Strong shock wave target detection and key feature point extraction Temporal image strong shock wave detection: Load the optimal weight model obtained in step three, and input the test set images or newly acquired transient schlieren images of strong shock waves in the isolated segment into the model. The model outputs the bounding box coordinates (x1, y1, x2, y2) and classification confidence score of the strong shock wave target in each image. Set the confidence score threshold to 0.5, filter out low-confidence detection results with confidence scores below the threshold, eliminate false detection targets, and ensure the accuracy of strong shock wave detection. Figure 11 and Figure 14 For the image to be predicted, Figure 12 and Figure 15 The corresponding prediction results are obtained using the optimized model of this invention. Figure 13 and Figure 16 As can be clearly seen from the prediction results using the original model, the optimized model of this invention can achieve more accurate shock wave structure target identification and detection.

[0033] Key feature point extraction of shock wave structure: To accurately capture the dynamic evolution and oscillation characteristics of strong shock wave structures, core feature points reflecting key locations within the strong shock wave's action area are extracted from the detection frame. Considering the dynamic characteristics of the strong shock wave in the isolation section and the requirements of aerodynamic analysis, the key feature points for subsequent analysis should reflect the radial positional changes of the strong shock wave within the isolation section, providing core data support for subsequent strong shock wave oscillation analysis, pulsating pressure correlation, and other studies. In this example, the corresponding coordinates are "x1, y2" output by the YOLO model.

[0034] Step 5: Verification of Method Validity In this invention, the data of an air-breathing aircraft intake at Mach number 4 are used as the research object.

[0035] Comparative experiment: Compared with the YOLO fourth version model (without using kmeans prior box clustering optimization and without using data augmentation strategy), the algorithm was compared on the same test set, and the results are shown in Table 1.

[0036] Table 1: Model Performance Comparison As shown in Table 1, the method of this invention achieves an average accuracy (AP) of 98.3% (compared to only 92.1% for the original YOLO model) and a detection speed of 64 FPS (compared to only 33 FPS for the original YOLO model). Therefore, this invention balances detection accuracy and real-time performance, significantly outperforming the original YOLO model before optimization, and fully meets the engineering requirements for high-frequency detection of key structures in the flow field of aircraft.

[0037] The above solution is merely an illustration of a preferred example and is not limited thereto. When implementing this invention, appropriate substitutions and / or modifications can be made according to the user's needs.

[0038] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for the present invention. Other modifications can be readily made by those skilled in the art. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and examples shown and described herein.

Claims

1. A method for detecting key structures and extracting feature points in a flow field based on an optimized YOLO model, characterized in that, include: S1. In wind tunnel testing, a high-speed schlieren acquisition system is used to acquire time-series schlieren images of the flow field of an air-breathing aircraft under different operating conditions. S2. By preprocessing the flow field time-series schlieren image, a data set corresponding to the strong shock wave of the isolation section of the air-breathing aircraft is obtained, and the data set is divided into training set, validation set and test set according to the proportion. S3. Construct an optimized YOLO object detection model, and iteratively train and validate the YOLO object detection model using the training set and validation set. After training, the model parameters with the highest average accuracy (AP) in the validation set are taken as the optimal weight model. S4. Use the optimal weight model to complete the target detection and feature point extraction of the test set or the schlieren image to be detected; The optimization of the YOLO object detection model includes: prior box clustering optimization based on the k-means algorithm, lightweight adaptation adjustment of the original YOLO model network structure, and optimization selection of the loss function of the YOLO object detection model.

2. The method for detecting key structures and extracting feature points in a flow field based on an optimized YOLO model as described in claim 1, characterized in that, In S1, the flow field time-series schlieren image is acquired by collecting 3,000 original images for each working condition, which can completely capture the dynamic evolution of the strong shock wave, with the acquisition frame rate set to 9,000 FPS.

3. The method for detecting key structures and extracting feature points in a flow field based on an optimized YOLO model as described in claim 1, characterized in that, In S2, the preprocessing includes: S20. Crop out isolated segments from the original image, and the resolution of the cropped isolated segments is uniformly 202×59 pixels; S21. Use the labeling tool LabelImg to label 300 randomly selected isolation segment images as labeling samples. The labeling content includes: the coordinates of the minimum bounding rectangle detection box of the strong shock wave target and the category label. The minimum bounding rectangle detection box should completely wrap around the core area of ​​the strong shock wave and fit the edge of the strong shock wave. The category label is strong shock wave in the isolation segment. S22. An image enhancement strategy is used to enhance the labeled samples nine times to obtain a set of 3,000 enhanced labeled images. The image enhancement strategy includes the following enhancement methods: adding Gaussian noise with a standard deviation of 0.02 to 0.08, and adjusting brightness, contrast, and saturation with adjustment factors of 0.7 to 1.

3. The image enhancement strategy refers to applying each enhancement method independently or in combination.

4. The method for detecting key structures and extracting feature points in a flow field based on an optimized YOLO model as described in claim 1, characterized in that, Prior box clustering optimization based on the kmeans algorithm refers to: based on the narrow shape and size distribution of the strong shock wave in the isolation segment, after extracting the width and height data corresponding to all labeled boxes in the training set, the kmeans clustering algorithm is used to perform clustering analysis on the width and height data. In the clustering process, the intersection-union ratio (IoU) between the bounding boxes and the cluster centers was used as the distance metric. The elbow rule was used to verify the clustering error under different numbers of clusters k. It was found that the matching degree between the prior boxes and the size of the strong shock wave target was the highest when k=6. The corresponding six sets of highest priority prior box sizes were obtained as follows: (54,136), (62,144), (60,151), (65,148), (65,151), (70,159). Replace the default prior boxes of the YOLO model with the six sets of highest priority bounding box sizes to obtain the initial anchor points that fit the target.

5. The method for detecting key structures and extracting feature points in a flow field based on an optimized YOLO model as described in claim 1, characterized in that, The lightweight adaptation adjustment of the original YOLO model network structure refers to: The backbone network in the YOLO network is configured to include: 2 CBL layers + 3 CSP layers + 1 CBL layer; The feature fusion module for the neck structure in the YOLO network is set to be derived from the last layer CBL and the third-to-last layer CSP of the backbone network. The YOLO network head structure retains two scales of detection heads, with three prior boxes set for each scale.

6. The method for detecting key structures and extracting feature points in a flow field based on an optimized YOLO model as described in claim 1, characterized in that, The loss function selected for the YOLO object detection model is the CIoU loss function.

7. The method for detecting key structures and extracting feature points in a flow field based on an optimized YOLO model as described in claim 1, characterized in that, In S4, for the detection box coordinates (x1, y1, x2, y2) of the strong shock wave target in each image, the feature point selection is a coordinate point (x1, y2) that can reflect the radial position change of the strong shock wave within the isolation section.

Images