Calibration board positioning point initial value positioning method based on deep learning

Abstract

The invention relates to the technical field of image processing, solves the technical problem that a positioning point on a calibration plate is difficult to effectively recognize in a complex scene in the prior art, and particularly relates to a calibration plate positioning point initial value positioning method based on deep learning, which comprises the following steps: acquiring a calibration plate image, and constructing a positioning point detection model; defining a loss function, and pre-training the positioning point detection model; obtaining a calibration plate area and a detection result after the position of the calibration plate is segmented in the calibration plate image; judging whether the initial positioning point is located in the calibration plate area or not according to the detection result; and marking and correcting the target positioning point in the calibration plate area to obtain a positioned calibration plate image. According to the method, rapid and accurate positioning of the positioning points on the calibration plate can be realized, related results can be given to a traditional image as positioning initial values and are directly used for subsequent traditional image processing, the whole calibration process is accelerated, and meanwhile, the calibration point identification precision and effect are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and particularly relates to a method for initially positioning calibration plate positioning points based on deep learning. Background Art

[0002] The detection of positioning points on a calibration plate is a key step to ensure that PIV and DIC systems can accurately perform tasks, which directly affects the performance of the system and the wide range of applications. In the prior art, such as the Chinese invention patent with the publication number CN114862866B and the invention name of a method, device, computer device, and storage medium for detecting a calibration plate, this application uses a semantic segmentation model, and subsequently many conditions need to be set to screen and obtain the corresponding calibration plate area, and the corresponding corner points are obtained by using polygon fitting and a homography matrix for the calibration plate, resulting in a relatively complex process. Another example is the Chinese invention patent with the publication number CN117765091A and the invention name of a method and system for calibrating external parameters of a camera based on identifying key points of a target. This application uses a target detection method to obtain the position of the calibration plate, then removes the background in the calibration plate, crops the image of the calibration plate, and further outputs the coordinates of the black and white corner points relative to the cropped calibration plate image.

[0003] However, in the actual use process, due to some complex scenarios (such as uneven illumination, calibration plate contamination, background circle interference, defocus, overexposure, small proportion of the calibration plate, etc.), the calibration plate positioning points may not be recognized; it is difficult to adapt to all scenarios with a unified set of parameters and processes using traditional image processing algorithms; using traditional image processing methods, the entire process is slightly time-consuming. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention provides a method for initially positioning calibration plate positioning points based on deep learning, which solves the technical problem that it is difficult to effectively recognize the positioning points on the calibration plate in the face of complex scenarios in the prior art.

[0005] To solve the above technical problems, the present invention provides the following technical solution: A method for initially positioning calibration plate positioning points based on deep learning, the method comprising the following steps:

[0006] S1. Obtain any calibration plate image I containing a calibration plate, and construct a positioning point detection model for eliminating background circle interference and long-tail effect in the calibration plate image I;

[0007] S2. Define a loss function L DFL , and use the loss function L DFL to pre-train the positioning point detection model;

[0008] S3. Use the calibration board image I as the input of the pre-trained positioning point detection model to obtain the calibration board area after segmenting the position of the calibration board in the calibration board image I, and multiple initial positioning points P i , and the detection results of the positioning point category and coordinates;

[0009] S4. Judge whether the initial positioning point P i is located within the calibration board area; if so, mark the initial positioning point P i as the target positioning point D i ; if not, exclude the initial positioning point P i ;

[0010] S5. Perform annotation correction on the target positioning point D i located within the calibration board area to obtain the calibrated calibration board image I'.

[0011] Further, the positioning point detection model uses the YOLOv8 network structure as the backbone network and introduces the FPN module and the attention module to perform region segmentation on the calibration board image I and output the detection results.

[0012] Further, the positioning point detection model includes a Conv + C2f module for feature extraction of the calibration board image I, and an FPN module for fusing context information of the features extracted by the Conv + C2f module to obtain multi-scale features;

[0013] And, a C2f + Upsample module for fusing multi-scale features and upsampling for feature enhancement, and an attention module for performing attention weighting on the fused features output by the C2f + Upsample module by introducing an attention mechanism.

[0014] Further, the positioning point detection model also includes a Detect / Seg Head module for outputting the calibration board area and the detection results.

[0015] Further, the attention module includes an SE module for compressing the global information of the fused features to capture the global features of each channel, calculating the weights of the global features to generate the attention weights of each channel, and finally weighting the global features of each channel according to the attention weights to generate the detection results;

[0016] And, an AttentionU-Net module for dynamically adjusting the attention degree of different regions in the fused features and focusing on the calibration board area to suppress the interference of irrelevant backgrounds.

[0017] Further, the expression of the loss function L DFL is:

[0018] LDFL = α·L cls + β·L reg + γ·L dynamic

[0019] Wherein, L cls is the classification loss; L reg is the regression loss; L dynamic is the dynamic freezing loss; α, β, and γ are all weighting coefficients.

[0020] With the above technical solution, the present invention provides an initial value positioning method for calibration plate positioning points based on deep learning, which has at least the following beneficial effects:

[0021] 1. The present invention can achieve "fast and accurate" positioning of the positioning points on the calibration plate, and can give the relevant results to the traditional image as the initial value of positioning, which can be directly used for subsequent traditional image processing, accelerating the entire calibration process, and at the same time improving the recognition accuracy and effect of the calibration points.

[0022] 2. The present invention can directly obtain the corresponding calibration plate area without additional operations, and the process is relatively simple. At the same time, the positioning point detection model is used to directly output the corresponding positioning point coordinates and categories, and the output information is more and more accurate.

[0023] 3. The present invention can more accurately determine the position of the calibration plate, and at the same time can exclude some interferences of the background circle relatively close to the calibration plate. By using the method of combining calibration plate segmentation and positioning point detection, the coordinates of the positioning points relative to the original image can be directly output, and the process is relatively simple. Description of the Drawings

[0024] The drawings described herein are used to provide a further understanding of the present application, and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application, and do not constitute an improper limitation of the present application. In the drawings:

[0025] Figure 1 is a schematic diagram of a calibration plate with background circle interference in the present invention;

[0026] Figure 2 is a schematic diagram of the network structure of the positioning point detection model in the present invention;

[0027] Figure 3 is a schematic diagram showing the long-tail effect of the proportion of positioning points in the DIC calibration plate in the present invention;

[0028] Figure 4 is a schematic diagram showing the long-tail effect of the proportion of positioning points in the PIV calibration plate in the present invention;

[0029] Figure 5 is a schematic diagram of the DIC simulation calibration plate in the present invention;

[0030] Figure 6 Schematic diagram of the PIV simulation calibration board in the present invention;

[0031] Figure 7 Schematic diagram of the detection result of the DIC calibration board with long-tail effect in the present invention;

[0032] Figure 8 Schematic diagram of the detection result of the PIV calibration board with long-tail effect in the present invention. Detailed implementation manners

[0033] To make the above objects, features and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific implementation manners. Thereby, the implementation process of how the present application uses technical means to solve technical problems and achieve technical effects can be fully understood and implemented accordingly.

[0034] The detection of the positioning points on the calibration board plays a crucial role in the fields of PIV and DIC. Its main purposes and significance can be summarized as follows:

[0035] Establish an accurate geometric model: By detecting the positioning points on the calibration board, it helps to establish an accurate geometric model from the three-dimensional space to the two-dimensional image plane, which is conducive to achieving precise measurement and positioning in subsequent vision tasks.

[0036] Correct lens distortion: The lens will produce distortion during imaging, such as barrel distortion or pincushion distortion. Through the detection of positioning points, these distortions can be identified and corrected, thereby obtaining more accurate image data.

[0037] Improve measurement accuracy: In the PIV and DIC systems, by accurately detecting the positioning points on the calibration board, the measurement accuracy of the size, shape and position of the object can be improved, which is crucial for subsequent precision engineering applications.

[0038] Support complex vision algorithms: The detection of the positioning points on the calibration board is the basis of many advanced vision algorithms, such as stereo vision, 3D reconstruction and motion tracking, etc.

[0039] Enhance the robustness of the system: By using different types and designs of calibration boards (such as checkerboards, dot grids, etc.), the robustness of the vision system can be enhanced under various environmental conditions.

[0040] Provide calibration parameters: The positioning points detected during the calibration process are used to calculate the internal parameters (such as focal length, principal point coordinates and distortion coefficients) and external parameters (such as rotation and translation matrices) of the camera, and these parameters are crucial for the accurate interpretation of the image.

[0041] In recent years, deep learning technology has developed rapidly, and its application in the industrial field has become increasingly widespread. For the detection of positioning points on the calibration plate, there are several advantages:

[0042] ① Compared with traditional image processing, deep learning has better universality when processing complex and wide-ranging scenes, and can save the complex parameter adjustment process of traditional image processing;

[0043] ② In recent years, many deep learning algorithms with both high speed and high accuracy have emerged. When the computing power of the platform is sufficient, these algorithms have surpassed the speed of traditional image processing to a certain extent.

[0044] However, deep learning algorithms are not good at solving high-precision tasks, and it is impossible to achieve sub-pixel level accuracy for positioning point detection using deep learning algorithms alone. In the task of positioning point detection on PIV and DIC calibration plates, it cannot completely replace the previously integrated image processing algorithms. However, deep learning also has its own advantages. It can perform rough positioning of positioning points in various scenarios. At the same time, by designing and using relatively lightweight algorithms, deep learning can run slightly faster than traditional image processing. Therefore, using deep learning calibration plate rough positioning combined with traditional image processing sub-pixel recognition can theoretically achieve "fast and good" in positioning point detection tasks.

[0045] In summary, in the face of the technical defects existing in the prior art, the calibration plate positioning point initial value positioning method proposed in this application uses instance segmentation to directly obtain the corresponding calibration plate area without the need for additional operations. The process is relatively simple. At the same time, the positioning point detection model is used to directly output the corresponding positioning point coordinates and categories, and the output information is more and more accurate.

[0046] The location of the calibration plate is accurately segmented using the positioning point detection model. Compared with the prior art CN117765091A, the location of the calibration plate can be determined more accurately, and some interference from the background circle that is close to the calibration plate can be eliminated. By combining the calibration plate segmentation with the positioning point detection, the coordinates of the positioning point relative to the original image can be directly output, and the process is relatively simple.

[0047] Please refer to Figures 1 - 8 This embodiment proposes a method for initializing the calibration plate positioning points based on deep learning, which can achieve "fast and accurate" positioning of the positioning points on the calibration plate. The relevant results can be given to the traditional image as the initial positioning value, which can be directly used for subsequent traditional image processing, accelerating the entire calibration process, and improving the accuracy and effect of calibration point recognition. The method includes the following steps:

[0048] S1. Obtain a calibration board image I containing any calibration board, and construct a positioning point detection model for eliminating the interference of background circles and long-tail effects in the calibration board image I.

[0049] The positioning point detection model uses the YOLOv8 network structure as the backbone network, and introduces an FPN module and an attention module to perform regional segmentation on the calibration board image I and output the detection results.

[0050] As Figure 1 shown in the background circle interference phenomenon, the positioning points on the calibration board in some images account for a relatively small proportion of the pixels in the whole image. In the non-calibration board area of the image, there are some "similar positioning points".

[0051] For the above problems, if only the target detection method is used once to directly output the positions of the positioning points on the calibration board, it undoubtedly poses high requirements on the performance of the detection network. The resolution of DIC images is relatively large. In the form of single-stage end-to-end detection, a relatively large input resolution of the detection network is required, which undoubtedly also brings great troubles to the deployment and use of the model in terms of speed. In addition, since the background circles and the circular positioning points in the image are basically the same in appearance and have very little difference in feature representation, it is very difficult for the network model to learn the difference between the two. Therefore, directly outputting the positions of the positioning points on the calibration board by only one target detection method is not a good solution to solve this problem.

[0052] To effectively solve the above-mentioned technical problems, in this embodiment, a new positioning point detection model is constructed, including a Conv + C2f module for feature extraction of the calibration board image I, and an FPN module for fusing context information of the features extracted by the Conv + C2f module to obtain multi-scale features. A C2f + Upsample module for fusing multi-scale features and performing upsampling for feature enhancement, and an attention module for performing attention weighting on the fused features output by the C2f + Upsample module by introducing an attention mechanism; and a Detect / SegHead module for outputting the calibration board area and detection results.

[0053] The attention module includes an SE module for compressing the global information of the fused features to capture the global features of each channel, calculating the weights of the global features to generate the attention weights of each channel, and finally weighting the global features of each channel according to the attention weights to generate the detection results; and an AttentionU-Net module for dynamically adjusting the attention degree of different regions in the fused features and focusing on the calibration board area to suppress irrelevant background interference.

[0054] As Figure 2Schematic diagram of the network structure of the positioning point detection model shown. The Conv+C2f module is an existing module in the YOLOv8 network. The Conv+C2f module is an important feature extraction unit, mainly used to enhance the feature expression ability. Conv represents the convolutional layer, and C2f combines the idea of cross-stage partial connection and the strategy of feature fusion, aiming to improve the network's expression ability and reduce redundant calculations.

[0055] The FPN (Feature Pyramid Network) module is a module added to the YOLOv8 network in this embodiment. The FPN module performs excellent in multi-scale object detection through multi-level context information fusion, especially suitable for detection tasks where large and small targets coexist.

[0056] The C2f+Upsample module is a key unit for multi-scale feature fusion and upsampling in the YOLOv8 network, especially playing an important role in the FPN (Feature Pyramid Network) module. It combines the feature extraction and fusion capabilities of C2f (Cross Stage Partial Networks with Fusion) and the resolution restoration function of Upsample, mainly used to upsample low-resolution feature maps to higher resolutions for processing targets of different scales.

[0057] The SE module is a module added to the YOLOv8 network in this embodiment. The SE module mainly consists of three steps: Squeeze (compression), Excitation (excitation), and Recalibration (re-weighting). By introducing the channel attention mechanism, the expression ability of the positioning point detection model for different channel features is enhanced. The global features of each channel are captured through global information compression in the Squeeze stage, the attention weights of each channel are generated through weight calculation in the Excitation stage, and finally the global features are weighted in the Recalibration stage to enhance the feature expression of important channels, thereby outputting the coordinate box, that is, including the initial positioning point P i , and the detection results of the positioning point category and coordinates.

[0058] The AttentionU-Net module weights the fused features through attention weights, enabling the positioning point detection model to automatically learn to focus on more important feature regions and suppress the interference of irrelevant background regions. By introducing the attention mechanism, this module can dynamically adjust the degree of attention to different regions in the fused features, enabling the positioning point detection model to effectively focus on important calibration plate regions and suppress the interference of irrelevant backgrounds. This enables the AttentionU-Net module to significantly improve the segmentation accuracy when dealing with tasks of small targets, complex backgrounds, or high noise.

[0059] The Detect / Seg Head module is the output head for detection and segmentation respectively. The detection output is the rectangular box coordinate position of each target (calibration point), and the segmentation output is a polygonal mask area.

[0060] The positioning point detection model proposed in this embodiment can handle features and targets of different scales, improve the accuracy of region segmentation and target detection, and effectively solve the problems of small proportion of calibration board and interference of background circles. At the same time, compared with the method of only performing target detection once, the positioning point detection model can be trained using a smaller input resolution, greatly improving the overall detection time and running efficiency.

[0061] S2. Define the loss function L DFL , and use the loss function L DFL to pre-train the positioning point detection model. The expression of the loss function L DFL is:

[0062] L DFL = α·L cls + β·L reg + γ·L dynamic

[0063] In the formula, L cls is the classification loss, which is used for the prediction of target categories. In this embodiment, the cross-entropy loss is used; L reg is the regression loss, which is used for the prediction of target bounding boxes. In this embodiment, the IoU loss is used; L dynamic is the dynamic freezing loss, which is used to control which layer weights need to be frozen or the selection of frozen layers during training; α, β, and γ are all weighting coefficients, satisfying α + β + γ = 1, and are usually adjusted according to the actual situation, and no specific limitation is made here.

[0064] In this embodiment, by using the loss function L DFL during the pre-training of the positioning point detection model, the class imbalance problem is mainly solved. By balancing the importance of different classes, the attention of the positioning point detection model to minority classes is improved, so that the positioning point detection model will not reduce the attention to targets with a small class proportion due to the long-tail effect during the pre-training process.

[0065] Such as Figure 3 and Figure 4As shown in the figure, they are respectively the schematic diagrams of the long-tail effect presented by the proportion of positioning points in the DIC and PIV calibration plates. Compared with the circular positioning points on the calibration plate, the number and proportion of triangular, rectangular, and concentric circular positioning points are relatively small, which belongs to the typical "long-tail effect" problem. For this problem, simply increasing the calibration plate data in the actual scenario will only increase the sample numbers of triangles, rectangles, and concentric circles. At the same time, the sample number of circular positioning points also increases synchronously, and it cannot change the sample proportion of triangles, rectangles, and concentric circles. The "long-tail effect" still exists.

[0066] To solve the problem of the "long-tail effect", this embodiment also provides a new method, that is: by using Blender software and writing python scripts, artificially generate some "false calibration plates" to increase the sample proportion of triangles, rectangles, and concentric circles. Specifically as follows:

[0067] Create a calibration plate:

[0068] Open Blender and create a new project. Create a rectangular plane or square as the basis of the calibration plate and adjust its size to meet the calibration requirements.

[0069] Add circular marks:

[0070] In the 3D view, use the "Add" menu to create a circle. Adjust the radius size of the circle, multiple circles can be set, and they can be distributed on the calibration plate through the copy function.

[0071] Add triangular marks:

[0072] Select "Mesh" > "Circle" in the "Add" menu, and then convert the circle to a triangle in the edit mode. Use the "Add" > "Transform" function to adjust the position and angle to ensure they are evenly distributed.

[0073] Add rectangular marks:

[0074] Use "Add" > "Mesh" > "Rectangle" to create a rectangle, and adjust its aspect ratio to ensure coordination with the layout and size of other graphics.

[0075] Add concentric circles:

[0076] Use the circle tool to create circles with different radii multiple times, and use the offset and rotation tools to create the effect of concentric circles.

[0077] Materials and textures:

[0078] Add different materials to different graphics to make them more prominent in the simulation. You can choose to add different colors to each graphic, or use the commonly used high-contrast black and white color scheme for calibration plates.

[0079] Simulated camera:

[0080] Set multiple virtual cameras to simulate photographing the calibration board from different angles. By adjusting parameters such as the focal length and viewing angle of the camera, ensure that different shooting angles are applicable to subsequent calibration tasks.

[0081] Rendering and saving:

[0082] Use the rendering function of Blender to perform image rendering to generate calibration images at different angles. Save the rendering results to ensure that the output image resolution meets the requirements of subsequent calibration algorithms.

[0083] As Figure 5 and Figure 6 shown, they are respectively schematic diagrams of DIC and PIV simulation calibration boards. It can be seen from the figure that by adopting this method, the "long-tail effect" problem in the positioning point detection process is basically solved.

[0084] For this phenomenon of the long-tail effect, this embodiment also proposes another solution. Since in the actual calibration board images, the proportions and quantities of triangles, rectangles, and concentric circles are all small, it is particularly important to improve the recall rates of these categories. Therefore, in actual engineering, combined with the F1-Score curve graph, set lower detection confidence levels for these categories with smaller proportions, effectively improving the recall rates of the positioning points of these categories.

[0085] The F1-Score curve graph is a graph used to evaluate the performance of a classification model, showing the changes in F1-Score at different thresholds. F1-Score is the harmonic mean of Precision and Recall, and it is particularly useful in imbalanced class datasets because it takes into account both the precision and recall capabilities of the model. The F1-Score curve graph can help select the optimal classification threshold, that is, the threshold that can balance precision and recall. In imbalanced data, choosing the appropriate threshold can effectively prevent the model from being biased towards the majority class. The expression of F1-Score is:

[0086]

[0087] In the formula, Precision represents precision; Recall represents recall.

[0088] As Figure 7 and Figure 8 shown, they are respectively schematic diagrams of the detection results of DIC and PIV calibration boards with the long-tail effect. It can be seen from the result diagrams that the three methods proposed in this embodiment can all effectively eliminate the long-tail effect faced by the positioning point detection model during the detection process, so as to be able to handle features and targets of different scales and improve the accuracy of region segmentation and target detection.

[0089] S3. Use the calibration plate image I as the input of the pre-trained positioning point detection model to obtain the calibration plate area after segmenting the position of the calibration plate in the calibration plate image I, and multiple initial positioning points P i , and the detection results of the positioning point category and coordinates.

[0090] S4. Determine whether the initial positioning point P i is located within the calibration plate area according to the detection results; if so, mark the initial positioning point P i as the target positioning point D i ; if not, exclude the initial positioning point P i .

[0091] In this embodiment, it is determined whether the detected initial positioning point P i is within the calibration plate area segmented from the calibration plate, thereby eliminating the interference of the background circle. First, the region segmentation of the calibration plate image I can be completed through the positioning point detection model to find the mask position of the calibration plate in the image, and then the initial positioning point P i in the mask region is detected. The initial positioning point P i not in the mask region will be excluded, while the initial positioning point P i in the mask region is the required target positioning point D i .

[0092] S5. Perform annotation correction on the target positioning point D i located within the calibration plate area to obtain the calibrated calibration plate image I';

[0093] Since there may be hundreds of positioning points on a calibration plate image, it belongs to the calibration problem of dense targets. If each point is annotated one by one, even using the current popular SAM large model to guide the target annotation method, it will take a long time for annotation, affecting the progress of the final algorithm development.

[0094] Therefore, in this embodiment, in the application of deep learning, a semi-automatic annotation method is realized, which greatly reduces the annotation time of the dataset and speeds up the iteration speed of the model. By using the positioning point detection model to perform prediction and positioning first, and then manually correcting the position, it is beneficial to accelerate the annotation speed and improve the annotation efficiency. Then, the unannotated dataset is inferred, and the inference results are saved as annotation files in relevant formats. In this way, only a part of the targets with incorrect annotation effects need to be manually modified, greatly improving the annotation efficiency of the dataset. Using this semi-automatic annotation method, as the accuracy and generalization of the training positioning point detection model continue to improve, the annotation quality and annotation efficiency will also be optimized accordingly.

[0095] The method proposed in this embodiment effectively improves the detection rate of positioning points, and provides a very good initial value for the subsequent accurate positioning of positioning points on the PIV and DIC calibration plates. At the same time, combining with the positioning point detection model can significantly reduce the time consumption of the entire positioning point detection process, generally not exceeding 50% of the original process, and preferably improves the operation efficiency of the algorithm.

[0096] Those of ordinary skill in the art can understand that all or part of the steps in implementing the method of the above embodiment can be completed by instructing relevant hardware through a program. Therefore, this application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) containing computer-usable program code.

[0097] The above embodiments have introduced the present invention in detail. Specific examples are used in this article to elaborate on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A method for initial value positioning of calibration plate positioning points based on deep learning, characterized in that: The method comprises the following steps: S1, obtaining any calibration plate image I including the calibration plate, and constructing a positioning point detection model for eliminating background circle interference and long tail effect in the calibration plate image I; S2. Define the loss function L DFL , and use the loss function L DFL Pre-train the anchor point detection model; S3, using the calibration plate image I as the input of the pre-trained positioning point detection model, and obtaining the calibration plate area after the calibration plate position segmentation in the calibration plate image I, and the calibration plate area containing multiple initial positioning points P i , detection results of positioning point categories and coordinates; S4. Determine the initial positioning point P according to the detection results i Is it located in the calibration plate area? If so, the initial positioning point P i Mark as target positioning point D i ; If not, exclude the initial positioning point P i ; S5, the target positioning point D located in the calibration plate area i Perform annotation correction to obtain the positioned calibration plate image I′.

2. The calibration plate positioning point initial value positioning method according to claim 1, characterized in that: The positioning point detection model adopts the YOLOv8 network structure as the backbone network, and introduces the FPN module and the attention module to perform region segmentation on the calibration plate image I and output the detection results.

3. The method for initial value positioning of the positioning point of the calibration plate according to claim 1 or 2, characterized in that: The positioning point detection model includes a Conv+C2f module for extracting features from the calibration plate image I, and an FPN module for fusing context information on the features extracted by the Conv+C2f module to obtain multi-scale features; As well as, a C2f+Upsample module that fuses multi-scale features and upsamples them for feature enhancement, and an attention module that introduces an attention mechanism to perform attention weighting on the fused features output by the C2f+Upsample module.

4. The method for initial value positioning of the positioning point of the calibration plate according to claim 3, characterized in that: The positioning point detection model also includes a Detect / Seg Head module for outputting a calibration plate area and a detection result.

5. The method for initial value positioning of the positioning point of the calibration plate according to claim 3, characterized in that: The attention module includes a SE module for compressing the global information of the fusion features to capture the global features of each channel, calculating the weight of the global features to generate the attention weight of each channel, and finally weighting the global features of each channel according to the attention weight to generate the detection result; And, the AttentionU-Net module dynamically adjusts the attention level of different areas in the fused features and focuses on the calibration plate area to suppress irrelevant background interference.

6. The calibration plate positioning point initial value positioning method according to claim 1, characterized in that: The loss function L DFL The expression is: L DFL =α·L cls +β·L reg +γ·L dynamic Where, L cls is the classification loss; L reg is the regression loss; L dynamic is the dynamic freezing loss; α, β, γ are all weighting coefficients.

Images