Method and device for positioning cross directrix of display screen image and storage medium

Abstract

The invention discloses a method and a device for positioning a cross directrix of a display screen image and a storage medium, which are used for improving the accuracy of positioning the cross directrix. Acquiring a display screen image and a saddle point image through an acquisition camera; performing preliminary distortion correction on the cross directrix of the display screen image through the saddle point on the saddle point image; performing directrix analysis on the display screen image through the deep convolutional neural network model to generate directrix detection data; performing edge detection on the display screen image according to frame information in the directrix detection data, performing straight line detection, and determining directrix intersection point information; clustering slope features in the alignment line intersection point information by adopting a weighted clustering analysis method to generate a straight line category and a clustering center of the straight line category; generating center position information of the cross directrix according to the plurality of straight line categories and the clustering center; and calculating the center coordinate of the target directrix according to the center position information, the frame information in the directrix detection data and the confidence degree information.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of display screen detection, and in particular, to a method for positioning a crosshair of a display screen image. Background Art

[0002] With the continuous improvement of display screen development technology, new display screens emerge in an endless stream, and the structure of new display screens is becoming more and more precise and complex, which makes the defect detection link of new display screens more and more complex. When a new display screen enters the defect detection, it needs to be aligned and positioned first. In the active alignment process of the standard display engine, the crosshair technology is widely used in the installation and assembly stages as a key reference mark. In this process, the display quality requirements for the crosshair at a predetermined position of the new display screen are extremely high.

[0003] However, due to the existence of display defects and screen body defects of the new display screen, problems such as local fracture, missing or adhesion of the crosshair may occur, thus affecting the detection accuracy and stability. Moreover, the number of layers of the display screen is increasing continuously, and the complex film structure also causes various new display errors of the crosshair.

[0004] The current detection and positioning of the crosshair mainly rely on traditional image processing technologies, such as line detection based on the Hough transform and template matching. These methods can achieve effective detection under a display screen with a simple hierarchical structure, but in a new display screen, when the crosshair is prone to local missing or noise interference, the positioning accuracy is significantly reduced, that is, the accuracy of positioning the crosshair is reduced. Summary of the Invention

[0005] The present application discloses a method, device and storage medium for positioning a crosshair of a display screen image, which are used to improve the accuracy of positioning the crosshair.

[0006] In a first aspect, an embodiment of the present application provides a method for positioning a crosshair of a display screen image, including: Input a designed crosshair screen and a checkerboard screen into a target display screen, and collect a display screen image and a saddle point image through a collection camera. There are several crosshairs on the display screen image, and several saddle points on the saddle point image; Perform preliminary distortion correction on the crosshair of the display screen image through the saddle points on the saddle point image; Perform crosshair analysis on the display screen image through a deep convolutional neural network model to generate crosshair detection data, and the crosshair detection data includes border information and confidence information; Perform edge detection on the display screen image according to the border information in the crosshair detection data, and then perform line detection to determine the crosshair intersection information; Cluster the slope features in the crosshair intersection information using the weighted clustering analysis method to generate several line categories and the clustering centers of the line categories; Generate the center position information of the crosshair based on the several line categories and the clustering centers; Calculate the target crosshair center coordinates based on the center position information, the border information and the confidence information in the crosshair detection data.

[0007] Optionally, the steps of performing preliminary distortion correction on the crosshair of the display screen image through the saddle points on the saddle point image include: Locate the saddle points on the saddle point image by the adaptive multi-scale feature fusion method; When there is a situation of missing saddle points, perform polynomial fitting on the already determined saddle points to generate several fitting curves; Perform interpolation processing and intersection calculation on the missing saddle points according to the several fitting curves to complete the missing saddle points; Generate a homography matrix based on all the completed saddle points and the reference saddle points on the standard template, and perform preliminary distortion correction on the crosshair of the display screen image through the homography matrix.

[0008] Optionally, the steps of performing edge detection on the display screen image according to the border information in the crosshair detection data, and then performing line detection to determine the crosshair intersection information include: Perform region screening on the display screen image according to the border information in the crosshair detection data; Calculate the gradients of the display screen image after region screening in the horizontal and vertical directions to extract potential edge information; Screen the local maxima along the gradient direction according to two preset thresholds, and only retain the significant edge points to remove the non-critical edges; Perform line detection on the display screen image after edge detection according to the Hough transform to determine the crosshair intersection information.

[0009] Optionally, the crosshair detection data further includes type information; After the step of performing crosshair analysis on the display screen image through the deep convolutional neural network model to generate the crosshair detection data, the positioning method further includes: Perform region annotation on the initial crosshair on the display screen image to generate a reference crosshair border, and associate a reference feature label of the crosshair with the reference crosshair border, and the reference feature label includes a type feature label, a position feature label and a confidence feature label; Train the deep convolutional neural network model according to the border information, the type information, the confidence information, the type feature label, the position feature label, the confidence feature label and a preset composite loss function.

[0010] Optionally, after the step of calculating the center coordinates of the target reticle based on the center position information, the border information and the confidence information in the reticle detection data, the positioning method further includes: Construct an ROI region for detecting the cross reticle according to the center coordinates of the target reticle; Perform gray projection processing on the ROI region to generate a gray projection sequence; Generate dynamic gray projection change threshold information based on the defect projection fluctuation law; Perform cross reticle defect detection on the gray projection sequence according to the dynamic gray projection change threshold information to generate a first reticle detection result.

[0011] Optionally, after the step of constructing an ROI region for detecting the cross reticle according to the center coordinates of the target reticle, the positioning method further includes: Construct image hierarchies with different resolutions on the ROI region, and perform multi-scale analysis on the cross reticle using a pyramid structure to generate a second reticle detection result.

[0012] Optionally, after performing gray projection processing on the ROI region to generate a gray projection sequence, the positioning method further includes: Perform gray projection and defect detection on the gray projection sequence at different scales respectively to generate a third reticle detection result.

[0013] In a second aspect, an embodiment of the present application provides a positioning device for a cross reticle of a display screen image, including: An acquisition unit, configured to input a designed cross reticle image and a checkerboard image into a target display screen, and acquire a display screen image and a saddle point image through an acquisition camera. There are several cross reticles on the display screen image, and several saddle points on the saddle point image; A correction unit, configured to perform preliminary distortion correction on the cross reticle of the display screen image through the saddle points on the saddle point image; A first generation unit, configured to perform reticle analysis on the display screen image through a deep convolutional neural network model to generate reticle detection data, where the reticle detection data includes border information and confidence information; A determination unit, configured to perform edge detection on the display screen image according to the border information in the reticle detection data, and then perform line detection to determine the reticle intersection information; A second generation unit, configured to cluster the slope features in the reticle intersection information by using a weighted clustering analysis method to generate several line categories and the clustering centers of the line categories; A third generation unit, configured to generate the center position information of the cross reticle according to the several line categories and the clustering centers; A calculation unit, configured to calculate the center coordinates of the target reticle according to the center position information, the border information and the confidence information in the reticle detection data.

[0014] Optionally, the correction unit includes: Locate the saddle points on the saddle point image by the adaptive multi-scale feature fusion method; When there is a situation of missing saddle points, perform polynomial fitting on the already determined saddle points to generate several fitting curves; Interpolate and calculate the intersection points of the missing saddle points according to the several fitting curves to complete the missing saddle points; Generate a homography matrix based on all the completed saddle points and the reference saddle points on the standard template, and perform preliminary distortion correction on the crosshair of the display screen image through the homography matrix.

[0015] Optionally, the determination unit includes: Perform regional screening on the display screen image according to the border information in the crosshair detection data; Calculate the gradients of the display screen image after regional screening in the horizontal and vertical directions to extract potential edge information; Screen the local maximum values along the gradient direction according to two preset thresholds, and only retain the significant edge points to remove the non-critical edges; Perform line detection on the display screen image after edge detection according to the Hough transform to determine the crosshair intersection information.

[0016] Optionally, the crosshair detection data further includes type information; After the first generation unit, the positioning device further includes: A fourth generation unit, configured to perform regional annotation on the initial crosshair on the display screen image, generate a reference crosshair border, and associate a reference feature label of the crosshair with the reference crosshair border, where the reference feature label includes a type feature label, a position feature label, and a confidence feature label; A training unit, configured to train a deep convolutional neural network model according to the border information, type information, confidence information, type feature label, position feature label, confidence feature label, and a preset composite loss function.

[0017] Optionally, after the calculation unit, the positioning device further includes: A detection unit, configured to construct an ROI region for detecting the crosshair according to the target crosshair center coordinates; A fifth generation unit, configured to perform gray projection processing on the ROI region to generate a gray projection sequence; A sixth generation unit, configured to generate dynamic gray projection change threshold information according to the defect projection fluctuation rule; A seventh generation unit, configured to perform crosshair defect detection on the gray projection sequence according to the dynamic gray projection change threshold information to generate a first crosshair detection result.

[0018] Optionally, after the second detection unit, the positioning device further includes: An eighth generation unit, configured to build image hierarchies with different resolutions on the ROI region, perform multi-scale analysis on the crosshair using a pyramid structure, and generate a second crosshair detection result.

[0019] Optionally, after the fifth generation unit, the positioning device further includes: A ninth generation unit, configured to perform gray-scale projection and defect detection on the gray-scale projection sequences at different scales respectively, and generate a third crosshair detection result.

[0020] In a third aspect, an embodiment of the present application provides a positioning device for a crosshair in a display screen image, including: a processor, a memory, an input / output unit, and a bus; the processor is connected to the memory, the input / output unit, and the bus; the memory stores a program, and the processor calls the program to execute the positioning method as described in the first aspect and any optional positioning method of the first aspect.

[0021] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, on which a program is stored, and when the program is executed on a computer, it executes the positioning method as described in the first aspect and any optional positioning method of the first aspect.

[0022] It can be seen from the above technical solutions that the embodiments of the present application have the following advantages: In the present application, first, the designed crosshair screen and checkerboard screen are input into the target display screen, and the display screen image and saddle point image are collected by a collection camera. There are several crosshairs on the display screen image, and several saddle points on the saddle point image. The crosshairs on the display screen image are initially corrected for distortion by the saddle points on the saddle point image. Then, the display screen image is analyzed for crosshairs through a deep convolutional neural network model, the crosshair features in the display screen image are extracted, the information of the crosshairs is determined, and crosshair detection data is generated. The crosshair detection data includes border information and confidence information. Edge detection is performed on the display screen image according to the border information in the crosshair detection data, and then line detection is performed to determine the crosshair intersection information. A weighted clustering analysis method is used to cluster the slope features in the crosshair intersection information to generate several line categories and the clustering centers of the line categories. The center position information of the crosshair is generated according to the several line categories and the clustering centers. The target crosshair center coordinates are calculated according to the center position information, the border information and the confidence information in the crosshair detection data.

[0023] By using a deep convolutional neural network model, the deep convolutional neural network model is made to learn the structural features of different types of crosshairs, enabling detection even in cases of partial absence. At the same time, combined with a weighted clustering analysis algorithm designed for the respective new defects that the crosshairs may have, and further intersection optimization is carried out. Finally, the target crosshair center coordinates are calculated based on the center position information, the border information in the crosshair detection data, and the confidence information, further improving the accuracy of crosshair positioning. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] To more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0025] Figure 1 Schematic diagram of an embodiment of the method for positioning the crosshair in the display screen image of the present application; Figure 2 Schematic diagram of an embodiment of the method for generating crosshair correction in the present application; Figure 3 Schematic diagram of an embodiment of the method for determining crosshair intersection information in the present application; Figure 4 Schematic diagram of an embodiment of the method for training the deep convolutional neural network model in the present application; Figure 5 Schematic diagram of an embodiment of the method for crosshair defect detection in the present application; Figure 6 Schematic diagram of another embodiment of the method for crosshair defect detection in the present application; Figure 7 Schematic diagram of another embodiment of the method for crosshair defect detection in the present application; Figure 8 Schematic diagram of an embodiment of the device for positioning the crosshair in the display screen image of the present application; Figure 9 Schematic diagram of another embodiment of the device for positioning the crosshair in the display screen image of the present application; Figure 10 Schematic diagram of a result of the saddle point position extracted by the traditional straight line fitting method; Figure 11 Schematic diagram of a result of the saddle point position extracted by the present application; Figure 12 Schematic diagram of a display screen image of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0026] In the following description, specific details such as specific system architectures, technologies, etc. are presented for purposes of illustration and not limitation, so as to provide a thorough understanding of the embodiments of the present application. However, those skilled in the art should understand that the present application can also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid unnecessary details from obscuring the description of the present application.

[0027] It should be understood that when used in the specification of the present application and the appended claims, the term "comprising" indicates the presence of the described features, wholes, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components, and / or their combinations.

[0028] It should also be understood that the term "and / or" as used in the specification of the present application and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0029] As used in the specification of the present application and the appended claims, the term "if" can be interpreted as "when" or "once" or "in response to determining" or "in response to detecting" depending on the context. Similarly, the phrase "if determined" or "if detected [the described condition or event]" can be interpreted as meaning "once determined" or "in response to determining" or "once detected [the described condition or event]" or "in response to detecting [the described condition or event]" depending on the context.

[0030] In addition, in the description of the specification of the present application and the appended claims, the terms "first", "second", "third", etc. are only used for distinguishing descriptions and cannot be understood as indicating or implying relative importance.

[0031] Reference to "one embodiment" or "some embodiments" or the like described in the specification of the present application means that a specific feature, structure, or characteristic described in connection with that embodiment is included in one or more embodiments of the present application. Thus, statements such as "in one embodiment", "in some embodiments", "in other some embodiments", "in still other embodiments", etc. that appear in different places in this specification are not necessarily all referring to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "comprising", "including", "having", and their variants all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0032] In the prior art, due to the existence of display defects and panel defects in the new display screen, problems such as local fracture, missing, or adhesion may occur in the crosshair, thus affecting the accuracy and stability of its detection. Moreover, the number of layers of the display screen is increasing continuously, and the complex film structure also causes various new display errors in the crosshair.

[0033] The current detection and positioning of the crosshair mainly rely on traditional image processing techniques, such as line detection based on the Hough transform and template matching. These methods can achieve effective detection under the display screen with a simple hierarchical structure. However, in the new display screen, when the crosshair is prone to local missing or noise interference, its positioning accuracy significantly decreases, that is, the accuracy of positioning the crosshair is reduced.

[0034] Based on this, the present application discloses a method, device, and storage medium for positioning the crosshair of a display screen image, which are used to improve the accuracy of positioning the crosshair.

[0035] Next, the technical solutions in the present application will be described clearly and completely in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0036] The method of the present application can be applied to a server, device, terminal, or other devices with logical processing capabilities. In this regard, the present application makes no limitation. For the convenience of description, the following description will be made taking the execution entity as the terminal as an example.

[0037] Please refer to Figure 1 , an embodiment of a method for positioning the crosshair of a display screen image provided by the present application includes: 101. Input the designed crosshair image and checkerboard image into the target display screen, and collect the display screen image and saddle point image through the acquisition camera. There are several crosshairs on the display screen image, and several saddle points on the saddle point image; In this embodiment, the terminal will generate a crosshair image according to the detection items to be performed on the target display screen, and generate a checkerboard image for correcting the distortion of the crosshair. The checkerboard image shows double-color alternating grids, and the gray-level difference between the double-color grids is designed to be large enough to generate sufficient saddle points. The target display screen sequentially displays these two images, and allows the acquisition camera to perform image acquisition to generate the display screen image and the saddle point image.

[0038] In this embodiment, the terminal collects image data containing multiple crosshairs, ensuring that the data set includes different lighting conditions, different angles, and different occlusion situations. The data set contains diverse crosshair sizes, orientations, partial occlusions, etc., ensuring that the subsequent deep convolutional neural network model can learn rich enough crosshair features.

[0039] When necessary, the bounding box of each crosshair in the display screen image can be manually annotated or machine-annotated. The annotation format for each crosshair is [category, x1, y1, x2, y2], where (x1, y1) and (x2, y2) represent the upper left and lower right coordinates of the bounding box, respectively.

[0040] 102. Perform preliminary distortion correction on the crosshairs of the display screen image through the saddle points on the saddle point image; Since the target display screen is prone to being affected by the sampling camera lens during the shooting process, and the optical imaging system is inevitably affected by lens processing errors and assembly deviations during manufacturing and installation, the crosshairs may shift, deform, and tilt during the imaging process, which will have an adverse impact on subsequent precise positioning and detection accuracy. Moreover, problems such as the position offset and warping of the target display screen further cause new distortions of the crosshairs on the target display screen during the acquisition process.

[0041] To solve the above distortion problems, the embodiment of the present application uses a checkerboard image with saddle points. The checkerboard image is generated according to the structure and resolution of the target display screen. In the checkerboard image, a sufficient number of saddle points are created by setting a high-gap gray scale. A point pair set is constructed based on the correspondence between the detected saddle points and the ideal saddle point coordinates to achieve high-precision calibration, thereby effectively correcting the distortion of the crosshairs on the display screen image and improving the imaging accuracy. The specific correction steps will be further described in the subsequent embodiments.

[0042] 103. Analyze the crosshairs of the display screen image through a deep convolutional neural network model to generate crosshair detection data, where the crosshair detection data includes bounding box information and confidence information; In the embodiment of the present application, a deep learning object detection model (YOLOv9) is used to detect all the crosshairs in the input display screen image to generate crosshair detection data, that is, output the bounding box information, confidence information, and category labels of the target crosshairs. Then, by extracting the bounding box information, the central position of the crosshairs can be initially estimated, thus completing the rough positioning.

[0043] Use the YOLOv9 model for object detection because it has high detection speed and good accuracy. The YOLOv9 model uses a deep convolutional neural network (CNN) to extract the features of the crosshairs from the input display screen image, and predicts information such as the position (bounding box), category, and confidence of the target through a regression method.

[0044] During the object detection process, the YOLOv9 model outputs the corresponding bounding box coordinates and confidence scores for each detected crosshair. The bounding box is represented in the format [x1, y1, x2, y2], where (x1, y1) is the upper left corner coordinate of the target area, and (x2, y2) is the lower right corner coordinate. In addition, the YOLOv9 model also outputs the confidence score of the target, which is a value between 0 and 1, indicating the confidence level of the YOLOv9 model in the detection result. The higher the confidence, the more reliable the model's determination of the existence of the target.

[0045] 104. Perform edge detection on the display screen image according to the bounding box information in the crosshair detection data, and then perform line detection to determine the crosshair intersection information; In this embodiment, the terminal performs edge detection on the display screen image according to the bounding box information in the crosshair detection data, and then performs line detection to determine the crosshair intersection information. The specific method will be described in subsequent embodiments.

[0046] 105. Use the weighted clustering analysis method to cluster the slope features in the crosshair intersection information, and generate several line categories and the clustering centers of the line categories; In this embodiment, the terminal uses the weighted clustering analysis method to cluster the slope features in the crosshair intersection information, and generates several line categories and the clustering centers of the line categories.

[0047] Specifically, during the detection process of the crosshair intersections, the clustering analysis method is used to optimize the line intersections detected by the Hough transform to improve the positioning accuracy.

[0048] Traditional clustering algorithms (such as the K-Means algorithm) assume that all features contribute equally, that is, traditional methods are suitable for situations where the feature contributions are relatively close. However, in the actual detection application of the display screen crosshairs, due to the increasing precision and complexity of the display screen structure, there are many crosshair features (including normal crosshair design features and many defect features), resulting in large differences in the importance of different features. Therefore, this application adopts a clustering analysis algorithm combined with weighting to enhance the clustering effect, especially when some crosshair features have a greater impact on the clustering result.

[0049] In the weighted clustering analysis method, feature weights are introduced to adjust the distance calculation, making the influence of important features on the clustering result greater. In this embodiment, the calculation formula of the weighted Euclidean distance is designed as follows:

[0050] where \(x=(x_1,x_2,\cdots,x_n)\) represents the feature vector of the data point (the data point in the directrix intersection information), \(n\) represents the total number of features, and \(c=(c_1,c_2,\cdots,c_n)\) represents the clustering center. \(w_i\) is the weight of feature \(i\), ensuring that important features occupy a greater proportion in the calculation process. The weights of each feature are not similar. The calculation method of the specific crosshair feature weights will be described in subsequent embodiments.

[0051] 106. Generate the center position information of the crosshair according to several straight line categories and the clustering center; In the detected straight line set, the weighted K-means clustering algorithm is applied based on its slope feature to achieve adaptive classification. For each crosshair, theoretically, it should contain two perpendicular straight lines (horizontal line and vertical line). The slopes of the straight lines are clustered by the weighted K-means algorithm, divided into two categories, and the center of each category is calculated to determine the average slope representing the category. During the clustering process, the slope feature is used as the main basis, and different straight lines are assigned to the corresponding categories according to their direction attributes. Finally, the center straight line parameters are extracted from each clustering category, and the intersection coordinates of the two straight lines are solved to obtain the center position of the crosshair.

[0052] For each pair of horizontal and vertical straight lines in the clustering results, we can calculate their intersection. Assume the equation of the horizontal straight line is and the equation of the vertical straight line is , and the intersection coordinates \((x,y)\) (i.e., the crosshair center coordinates) are obtained by solving the simultaneous equations.

[0053] 107. Calculate the target crosshair center coordinates according to the center position information, the border information and the confidence information in the directrix detection data.

[0054] In this embodiment, during the process of detecting the crosshair, the deep learning model outputs the bounding box and its confidence of each crosshair, and the adaptive intersection detection method based on the weighted clustering algorithm extracts the intersection coordinates of each crosshair and uses them as the center points. In the embodiment of this application, a weighted fusion strategy is adopted to synchronously analyze the two results to further improve the detection accuracy. Specifically, using the confidence information of YOLOv9, the confidence information is used as the weight parameter to perform weighted calculation on the intersection coordinates of the crosshair. The mathematical expression is as follows:

[0055] Among them, is the center position of the crosshair detected by YOLOv9. is the intersection position detected by weighted clustering detection. is the confidence value of YOLOv9, which reflects YOLOv9's detection confidence in this crosshair.

[0056] If the confidence of the YOLOv9 model is lower than the preset threshold, it is more inclined to rely on the clustering-based detection results, thereby enhancing the robustness of the algorithm. Finally, the output of the two methods is fused by weighted average to obtain the high-precision center coordinates (x, y) of each crosshair.

[0057] In this application, the designed crosshair image and checkerboard image are first input into the target display screen, and the display screen image and saddle point image are collected by the acquisition camera. There are several crosshairs on the display screen image and several saddle points on the saddle point image. The crosshairs on the display screen image are initially corrected for distortion by the saddle points on the saddle point image. Then, the display screen image is analyzed for crosshairs by a deep convolutional neural network model, the crosshair features in the display screen image are extracted, the crosshair information is determined, and crosshair detection data is generated. The crosshair detection data includes border information and confidence information. Edge detection is performed on the display screen image according to the border information in the crosshair detection data, and then line detection is performed to determine the crosshair intersection information. The weighted clustering analysis method is used to cluster the slope features in the crosshair intersection information to generate several line categories and the clustering centers of the line categories. The center position information of the crosshair is generated according to the several line categories and the clustering centers. The target crosshair center coordinates are calculated according to the center position information, the border information and the confidence information in the crosshair detection data.

[0058] By using a deep convolutional neural network model, the deep convolutional neural network model learns the structural features of different types of crosshairs, and can still achieve detection even in the case of partial missing. At the same time, combined with the weighted clustering analysis algorithm designed for the possible respective new defects of the crosshairs, and further intersection optimization is performed. Finally, the target crosshair center coordinates are calculated according to the center position information, the border information and the confidence information in the crosshair detection data, further improving the accuracy of crosshair positioning.

[0059] Please refer to Figure 2 , this application provides an embodiment of a method for generating crosshair correction, including: 201. Locate the saddle points on the saddle point image by the adaptive multi-scale feature fusion method; Due to the inevitable influence of lens processing errors and assembly deviations during the manufacturing and installation of the optical imaging system, the alignment line may shift, deform, and tilt during the imaging process, thus having an adverse impact on subsequent precise positioning and detection accuracy.

[0060] In this embodiment, the terminal locates the saddle point on the saddle point image through the adaptive multi-scale feature fusion method. In the traditional method, the Hessian matrix of a single scale cannot adapt to the corner features (saddle point features) of the checkerboard in the checkerboard image because the corner features exist in different scales. In the embodiment of the present application, the calculation of the multi-scale Hessian matrix for the saddle point features is introduced to enhance the adaptability to saddle point images of different resolutions.

[0061] Define the Hessian matrices of different scales:

[0062] Where: , , are the second-order partial derivatives after Gaussian smoothing:

[0063]

[0064]

[0065] Where I is the saddle point image, x and y are the coordinates in the saddle point image, is the standard Gaussian kernel:

[0066] Calculate the eigenvalues of the Hessian matrix of different scales and take the maximum response point as the saddle point:

[0067] Thus, the position of the saddle point is initially determined. In the embodiment of the present application, the eigenvalues of the Hessian matrix are calculated at different scales to obtain the response maps, and then the maximum response points are found as the "saddle points". The eigenvalues of the Hessian matrix represent the strength (curvature) of the second-order change in the local area, that is, the large eigenvalues indicate a fast gradient change, and thus can represent the significant image structure (saddle point). The so-called "response value" of the response map is constructed in the following way: 1. Determinant , and are the eigenvalues of the determinant .

[0068] 2. Eigenvalue combinations (such as the approximate determinant used in SURF).

[0069] 3. Geometric interpretation of eigenvalues (such as the Frangi filter).

[0070] Therefore, each pixel value on the "response map" is essentially a certain combination of eigenvalues, that is, the response value is the Hessian eigenvalue.

[0071] In fact, the "maximum response point" does not refer to the only maximum value in the entire saddle point image, but the "local maximum point", which is locally maximum both spatially and in scale.

[0072] The detection process is actually to first obtain a response map Ri(x, y) at each scale σ¡, then construct a three-dimensional response space: R(x, y, σ). Next, in this response space, find the points that are locally maximum in the space (x, y) and also locally maximum in the scale direction (compared to the two scales above and below). All points that satisfy the maximum value are considered "candidate saddle points". Therefore, "multiple saddle points" come from multiple local maximum points in the response space, and each local maximum point corresponds to a truly significant image structure (such as the center of a checkerboard).

[0073] 202. When there is a situation of missing saddle points, perform polynomial fitting on the already determined saddle points to generate several fitting curves; When there is a situation of missing saddle points, the terminal performs polynomial fitting on the already determined saddle points to generate several fitting curves. Specifically, first determine a set of n data points from the existing saddle points , and use polynomial fitting:

[0074] Construct the fitting matrix equation:

[0075]

[0076]

[0077] Then the terminal solves the equation:

[0078] where is a 3x3 matrix, is the polynomial coefficient vector to be solved. After solving, the expression of the fitting curve can be obtained.

[0079] 203. Interpolate and calculate the intersection points of the missing saddle points according to several fitting curves to complete the missing saddle points; Next, the terminal performs interpolation processing and intersection calculation on the missing saddle points based on several fitting curves to complete the missing saddle points. Specifically, the terminal determines the number of data points (missing saddle points) to be extrapolated to the left and right, generates new independent variable x values according to the set step size (usually the length of the checkerboard grid), and then calculates the corresponding y values using the fitting equation constructed in the previous step to obtain a set of uniformly distributed row and column coordinates. Based on this, two regions of fitting curves can be constructed, and the intersection coordinates can be further solved to update the matrix data. Finally, accurate extraction of all saddle points on the checkerboard grid is achieved.

[0080] 204. Generate a homography matrix based on all the completed saddle points and the reference saddle points on the standard template, and perform preliminary distortion correction on the crosshairs of the display screen image through the homography matrix.

[0081] The terminal generates a homography matrix based on all the completed saddle points and the reference saddle points on the standard template, and performs preliminary distortion correction on the crosshairs of the display screen image through the homography matrix.

[0082] Through step 203, all saddle points in the saddle point image can be obtained and their coordinates can be determined , assuming the standard coordinates of the checkerboard grid are , and the relationship between the two can be expressed as:

[0083] where H is the perspective transformation matrix to be obtained, as shown below:

[0084] Based on the established correspondence between the determined saddle points and the ideal checkerboard grid points, a set of matching points (i.e., the saddle point matrix and the standard point matrix) can be constructed, and high-precision geometric correction can be performed using this relationship to effectively compensate for image distortion and improve the overall calibration accuracy.

[0085] As Figure 10 shown, there are certain deviations in the positions of the saddle points extracted by the traditional line fitting method; while Figure 11 shows the checkerboard grid saddle points extracted based on the method of the present invention. It can be clearly seen from this that the method used in this embodiment is significantly superior to the traditional line fitting method in terms of saddle point positioning accuracy.

[0086] Since the structure of the target display screen itself is becoming more and more complex, both the saddle point image and the display screen image have distortions. After the distortion correction of the display screen image, the position of the crosshairs has been initially adjusted. However, due to various defects and structural improvements in different types of display screens (especially the improvements in the circuit area on the pixel layer), the crosshairs used in the detection of each display screen are not the same. In this embodiment, the weight of the crosshair feature It is determined by the structural types (structural complexity) of different detection regions of the display screen, the crosshairs, the adaptation value of the structure, and the distance between the crosshairs and the sampling center.

[0087] Assume that there are 7 crosshairs on a target display screen. Please refer to Figure 12 , Figure 12 is the display screen image in this embodiment. There are also 7 crosshairs shown on the display screen image. The crosshairs can be marked by numbers. The crosshairs at the edge part are more important. When a complex structure such as a circuit region is set below the pixel layer of the target display screen, the crosshairs on this circuit structure are more important. The lower the adaptation degree of the crosshair type to the structure, the higher the importance. The meaning of the adaptation degree is that when the crosshair is located on such a structure, the higher the regional detection accuracy, the larger the adaptation degree is designed; conversely, the smaller the adaptation degree is designed. It is a value greater than 0 and less than 1. Usually, the crosshair type with the largest adaptation degree is selected for detection. The calculation formula is as follows:

[0088] In this embodiment, represents the structural weight of the crosshair, represents the structural complexity level parameter of the region where the crosshair is located, represents the distance ratio value between the standard position where the crosshair is located and the center of the acquisition camera (the farther the distance, the larger the ratio value), represents the structure adaptation degree parameter, represents the guaranteed value. The guaranteed value is used to prevent the structural weight of the crosshair from being lower than the preset threshold, and it will only be used when the structural weight of the crosshair is lower than the preset threshold. , , are the weighted values of the structural complexity level parameter, the distance ratio value, and the structure adaptation degree parameter respectively, which are usually designed according to the content of the detection item and will not be elaborated here.

[0089] Please refer to Figure 3 , this application provides an embodiment of a method for determining the intersection information of crosshairs, including: 301. Perform region screening on the display screen image according to the border information in the crosshair detection data; 302. Calculate the gradients of the display screen image after region screening in the horizontal and vertical directions to extract potential edge information; 303. Screen local maxima along the gradient direction according to two preset thresholds, and only retain significant edge points to remove non-critical edges; 304. Perform line detection on the display screen image after edge detection according to the Hough transform to determine the intersection information of the crosshairs.

[0090] Traditional methods usually require complex operations on the entire image, resulting in a large computational overhead during the crosshair center positioning and defect detection processes, thereby reducing the detection efficiency and making it difficult to meet the high real-time requirements. To solve the above problems, the corresponding region is cropped from the display screen image using the crosshair bounding box (bounding box information) detected by the YOLOv9 model to reduce the computational overhead, and edge detection and intersection extraction are only performed within the region of interest.

[0091] The basic process of edge detection is as follows: (1) Calculate the gradient of the display screen image: Calculate the gradients of the image in the horizontal and vertical directions to extract potential edge information.

[0092] (2) Screen the local maximum values along the gradient direction and only retain significant edge points to remove non-critical edges.

[0093] (3) Double-threshold processing: Set two thresholds, a high threshold and a low threshold, to ensure that the detected edges are neither affected by noise nor miss key structures.

[0094] After edge detection is completed, a designed weighted Hough transform is used for line detection. This method converts the image space into a polar coordinate space, making the line detection problem transformed into peak detection in the parameter space. The detection results are represented in polar coordinates, where each point corresponds to a line. By analyzing the set of detected lines, the intersection position of the crosshair can be further determined.

[0095] Please refer to Figure 4 , this application provides an embodiment of a method for training a deep convolutional neural network model. The crosshair detection data also includes type information, including: 401. Perform region annotation on the initial crosshair on the display screen image to generate a reference crosshair bounding box, and associate a reference feature label of the crosshair with the reference crosshair bounding box. The reference feature label includes a type feature label, a position feature label, and a confidence feature label; 402. Train a deep convolutional neural network model according to the bounding box information, type information, confidence information, type feature label, position feature label, confidence feature label, and a preset composite loss function.

[0096] In this embodiment, due to the influence of crosshair defects on the positioning detection of the crosshair, there are usually multiple crosshairs with positioning errors and type judgment errors.

[0097] In this embodiment, in order to enhance the detection effect of the model, a new YOLOv9 model is constructed to detect the target crosshair position. Since the new YOLOv9 model has high detection speed and good accuracy, in this embodiment, the terminal uses the YOLOv9 model to extract features from the input display screen image by using a deep convolutional neural network (CNN), and predicts the position and category of the target through a regression method. When necessary, before inputting the model for detection, the bounding boxes of each crosshair in the display screen image can be manually marked or machine marked, and the format of each crosshair mark is [category, x1, y1, x2, y2], where (x1, y1) and (x2, y2) respectively represent the upper left and lower right coordinates of the bounding box, and then it is analyzed with the result output by the model.

[0098] During the training process, the model optimizes the detection result by generating multiple candidate boxes and comparing them with the true annotation boxes. Specifically, during the training process, the crosshair classification data, confidence, and position are used as multiple loss value items. Specifically, in the embodiment of the present application, the position feature is designed as a combination of the bounding box difference loss and the intersection over union loss. Compared with the defect characteristics of the crosshair, the loss value item of this position feature is more in line with the characteristics of the crosshair, and the loss function is used to measure the error between the prediction result and the true label. The loss function L is as follows:

[0099] Among them, : Classification loss, which measures the error of the model in the classification task, is the weight coefficient of the classification loss. : Object loss, which measures the confidence that the model correctly identifies the object, is the weight coefficient of the object loss. : Bounding box regression loss, which measures the difference between the predicted bounding box and the true bounding box, is the weight coefficient of the bounding box regression loss. : Intersection over union loss, which is used to optimize the overlap degree between the predicted box and the true box. The parameters before each loss value are the corresponding weight parameters, is the weight coefficient of the intersection over union loss.

[0100] Please refer to Figure 5 , an embodiment of a method for detecting crosshair defects provided by the present application includes: 501. Construct an ROI region for detecting crosshairs according to the center coordinates of the target crosshair; According to the foregoing method, the central coordinates (x, y) of the crosshair have been extracted. Taking this center point as a reference, a certain range is expanded along the horizontal and vertical directions respectively to construct a Region of Interest (ROI). The size of the ROI can be reasonably set according to the image resolution and the actual size of the crosshair to ensure the accuracy and robustness of subsequent analysis.

[0101] 502. Perform gray projection processing on the ROI to generate a gray projection sequence; Horizontal gray projection: Within the ROI, calculate the mean value of the pixel gray values of each horizontal scan line to construct a gray projection sequence in the horizontal direction, and save this sequence. This sequence can be used to analyze the integrity of the crosshair and possible defects.

[0102] For the i-th row, the gray projection value is calculated by the formula:

[0103] Where: is the pixel gray value of the x-th column in the i-th row within the ROI.

[0104] is the width of the ROI, that is, the number of pixels in the horizontal range.

[0105] represents the horizontal scan area of this row.

[0106] Vertical gray projection: Similar to the horizontal projection, within the ROI, calculate the mean value of the pixel gray values of each vertical scan line to construct a gray projection sequence in the vertical direction, and save this sequence.

[0107] For the j-th column, the gray projection value is calculated by the formula:

[0108] Where: is the pixel gray value of the y-th row in the j-th column within the ROI.

[0109] is the height of the ROI, that is, the number of pixels in the vertical range.

[0110] represents the vertical scan area of this column.

[0111] 503. Generate dynamic gray projection change threshold information based on the defect projection fluctuation law; 504. Detect the crosshair defects in the gray projection sequence according to the dynamic gray projection change threshold information to generate the first crosshair detection result.

[0112] In this embodiment, the ability to detect various defects of the crosshair is poor. When the crosshair breaks, is missing, or has pixel adhesion, the existing methods are vulnerable to background noise and local morphological changes during the segmentation and recognition processes, resulting in limited defect detection ability and affecting the stability and robustness of the detection.

[0113] Save the projection sequence values obtained in step 502. Traditional gray projection methods usually rely on a fixed threshold to distinguish normal gray changes from potential defect regions. To improve the stability and adaptability of the detection, an adaptive trend threshold is created in this embodiment, and the threshold for each region is dynamically adjusted through the local mean or local standard deviation of the image. In this way, even if the brightness distribution in the display screen image is different, appropriate thresholds can be applied in different regions, greatly enhancing the robustness of the algorithm. During the detection process, first generate a dynamic gray projection change threshold T (dynamic gray projection change threshold information), which is used to distinguish normal projection fluctuations from potential defect regions.

[0114] In the horizontal and vertical gray projections, if the projection value of a certain region or shows a significant decrease in a certain section of the region, and the decrease amplitude exceeds the set threshold T, it can be determined that there may be potential abnormal points in this region.

[0115] By detecting continuous regions of gray values below the threshold, we can mark the missing or broken parts. To avoid noise interference, during the gray value detection process, the sliding window technique is used to analyze the gray change trend in the local region, so as to ensure that the detected abnormal regions have structural characteristics rather than single-point noise.

[0116] Please refer to Figure 6 , another embodiment of a method for crosshair defect detection provided by this application includes: 601. Build image hierarchies with different resolutions on the ROI region, and use the pyramid structure to perform multi-scale analysis on the crosshair to generate the second crosshair detection result.

[0117] Considering the scale changes of the crosshair in different scenarios, single-scale gray projection analysis may not be sufficient to detect all defects. Therefore, by constructing an image pyramid (Image Pyramid) to process images with different resolutions, the detection ability is enhanced.

[0118] In this embodiment, image pyramid construction is also utilized. Specifically, by constructing image levels with different resolutions, the ROI region of the crosshair is analyzed at multiple scales using the pyramid structure to ensure that the algorithm can adapt to crosshairs of different sizes.

[0119] Please refer to Figure 7 , another embodiment of a method for detecting crosshair defects provided by this application includes: 701. Perform gray projection and defect detection on the gray projection sequences at different scales respectively to generate a third crosshair detection result.

[0120] Considering the scale changes of the crosshair in different scenarios, single-scale gray projection analysis may not be sufficient to detect all defects. Therefore, a multi-scale analysis method is also selected to enhance the detection ability.

[0121] Layer-by-layer gray projection and defect detection, perform gray projection and defect detection respectively at multiple scales of the display screen image, and fuse the detection results of each scale to improve the recognition accuracy of the defect area.

[0122] Please refer to Figure 8 , an embodiment of a positioning device for the crosshair of a display screen image provided by this application includes: An acquisition unit 801, configured to input the designed crosshair picture and checkerboard picture into the target display screen, and acquire the display screen image and the saddle point image through an acquisition camera. There are several crosshairs on the display screen image and several saddle points on the saddle point image; A correction unit 802, configured to perform preliminary distortion correction on the crosshairs of the display screen image through the saddle points on the saddle point image; Optionally, the correction unit 802 includes: Locate the saddle points on the saddle point image through the adaptive multi-scale feature fusion method; When there is a situation of missing saddle points, perform polynomial fitting on the already determined saddle points to generate several fitting curves; Perform interpolation processing and intersection calculation on the missing saddle points according to the several fitting curves to complete the missing saddle points; Generate a homography matrix according to all the completed saddle points and the reference saddle points on the standard template, and perform preliminary distortion correction on the crosshairs of the display screen image through the homography matrix.

[0123] A first generation unit 803, configured to perform crosshair analysis on the display screen image through a deep convolutional neural network model to generate crosshair detection data, where the crosshair detection data includes border information and confidence information; The fourth generation unit 804 is configured to perform regional annotation on the initial crosshair on the display screen image, generate a reference crosshair border, and associate a reference feature label of the crosshair with the reference crosshair border. The reference feature label includes a type feature label, a position feature label, and a confidence feature label; The training unit 805 is configured to train a deep convolutional neural network model according to the border information, type information, confidence information, type feature label, position feature label, confidence feature label, and a preset composite loss function; The determination unit 806 is configured to perform edge detection on the display screen image according to the border information in the crosshair detection data, and then perform line detection to determine the crosshair intersection information; Optionally, the determination unit 806 includes: Performing regional screening on the display screen image according to the border information in the crosshair detection data; Calculating the gradients of the display screen image after regional screening in the horizontal and vertical directions to extract potential edge information; Filtering local maxima along the gradient direction according to two preset thresholds, and only retaining significant edge points to remove non-critical edges; Performing line detection on the display screen image after edge detection according to the Hough transform to determine the crosshair intersection information.

[0124] The second generation unit 807 is configured to cluster the slope features in the crosshair intersection information by using a weighted clustering analysis method to generate several line categories and the clustering centers of the line categories; The third generation unit 808 is configured to generate the center position information of the crosshair according to the several line categories and the clustering centers; The calculation unit 809 is configured to calculate the target crosshair center coordinates according to the center position information, the border information, and the confidence information in the crosshair detection data; The detection unit 810 is configured to construct an ROI region for detecting the crosshair according to the target crosshair center coordinates; The fifth generation unit 811 is configured to perform gray-scale projection processing on the ROI region to generate a gray-scale projection sequence; The sixth generation unit 812 is configured to generate dynamic gray-scale projection change threshold information according to the defect projection fluctuation rule; The seventh generation unit 813 is configured to perform crosshair defect detection on the gray-scale projection sequence according to the dynamic gray-scale projection change threshold information to generate a first crosshair detection result; The eighth generation unit 814 is configured to construct image hierarchies with different resolutions on the ROI region, and perform multi-scale analysis on the crosshair by using a pyramid structure to generate a second crosshair detection result; The ninth generation unit 815 is configured to perform gray projection and defect detection on the gray projection sequences at different scales respectively, and generate a third alignment detection result.

[0125] Please refer to Figure 9 , this application provides a positioning device for the crosshair of a display screen image, including: A processor 901, a memory 902, an input / output unit 903, and a bus 904.

[0126] The processor 901 is connected to the memory 902, the input / output unit 903, and the bus 904.

[0127] The memory 902 stores a program, and the processor 901 calls the program to execute the positioning methods as described in Figure 1 , Figure 2 and Figure 3 , Figure 4 , Figure 5 , Figure 6 and Figure 7 .

[0128] This application provides a computer-readable storage medium, on which a program is stored. When the program is executed on a computer, it executes the positioning methods as described in Figure 1 , Figure 2 and Figure 3 , Figure 4 , Figure 5 , Figure 6 and Figure 7 .

[0129] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the above-described systems, devices, and units can refer to the corresponding processes in the foregoing method embodiments, and will not be described herein again.

[0130] In several embodiments provided by this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division, and there can be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces, and the indirect couplings or communication connections of the devices or units can be in electrical, mechanical, or other forms.

[0131] The unit described as a separation component may or may not be physically separated. The component shown as a unit may or may not be a physical unit, that is, it may be located in one place or distributed across multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of the embodiments of the present application.

[0132] In addition, each functional unit in the various embodiments of the present application may be integrated in a processing unit, may exist physically alone for each unit, or two or more units may be integrated in one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of a software functional unit.

[0133] If the above-mentioned integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, read-only memory), random access memories (RAM, random access memory), magnetic disks, or optical discs that can store program codes.

Claims

1. A method for positioning a crosshair of a display screen image, characterized in that, Including: Input the designed crosshair image and checkerboard image into the target display screen, and collect the display screen image and saddle point image through the acquisition camera. There are several crosshairs on the display screen image, and several saddle points on the saddle point image; Perform preliminary distortion correction on the crosshairs of the display screen image through the saddle points on the saddle point image; Perform crosshair analysis on the display screen image through the deep convolutional neural network model to generate crosshair detection data, and the crosshair detection data includes border information and confidence information; Perform edge detection on the display screen image according to the border information in the crosshair detection data, and then perform line detection to determine the crosshair intersection information; Use the weighted clustering analysis method to cluster the slope features in the crosshair intersection information to generate several line categories and the clustering centers of the line categories; Generate the center position information of the crosshair according to the several line categories and the clustering centers; Calculate the target crosshair center coordinates according to the center position information, the border information in the crosshair detection data, and the confidence information.

2. The positioning method according to claim 1, characterized in that The step of performing preliminary distortion correction on the crosshairs of the display screen image through the saddle points on the saddle point image includes: Locate the saddle points on the saddle point image through the adaptive multi-scale feature fusion method; When there is a situation of missing saddle points, perform polynomial fitting on the already determined saddle points to generate several fitting curves; Perform interpolation processing and intersection calculation on the missing saddle points according to the several fitting curves to complete the missing saddle points; Generate a homography matrix according to all the completed saddle points and the reference saddle points on the standard template, and perform preliminary distortion correction on the crosshairs of the display screen image through the homography matrix.

3. The positioning method according to claim 1, characterized in that The step of performing edge detection on the display screen image according to the border information in the crosshair detection data, and then performing line detection to determine the crosshair intersection information includes: Perform region screening on the display screen image according to the border information in the crosshair detection data; Calculate the gradients of the display screen image in the horizontal and vertical directions after region screening to extract potential edge information; Screen the local maximum values along the gradient direction according to two preset thresholds, and only retain the significant edge points to remove the non-critical edges; Perform line detection on the display screen image after edge detection according to the Hough transform to determine the crosshair intersection information.

4. The positioning method according to claim 1, wherein The crosshair detection data further includes type information; After the step of performing crosshair analysis on the display screen image through the deep convolutional neural network model to generate crosshair detection data, the positioning method further includes: Perform region annotation on the initial crosshairs on the display screen image to generate a reference crosshair border, and associate a reference feature label of the crosshair with the reference crosshair border. The reference feature label includes a type feature label, a position feature label, and a confidence feature label; Train the deep convolutional neural network model according to the border information, the type information, the confidence information, the type feature label, the position feature label, the confidence feature label, and a preset composite loss function.

5. The positioning method according to any one of claims 1 to 4, characterized in that, After the step of calculating the center coordinates of the target reticle according to the center position information, the border information in the reticle detection data, and the confidence information, the positioning method further includes: Constructing an ROI region for detecting the cross reticle according to the center coordinates of the target reticle; Performing gray projection processing on the ROI region to generate a gray projection sequence; Generating dynamic gray projection change threshold information based on the defect projection fluctuation law; Performing cross reticle defect detection on the gray projection sequence according to the dynamic gray projection change threshold information to generate a first reticle detection result.

6. The positioning method according to claim 5, wherein After the step of constructing an ROI region for detecting the cross reticle according to the center coordinates of the target reticle, the positioning method further includes: Constructing image hierarchies with different resolutions on the ROI region, and performing multi-scale analysis on the cross reticle using a pyramid structure to generate a second reticle detection result.

7. The positioning method according to claim 5, characterized in that, After performing gray projection processing on the ROI region to generate a gray projection sequence, the positioning method further includes: Performing gray projection and defect detection on the gray projection sequence at different scales respectively to generate a third reticle detection result.

8. A positioning device for the crosshair of a display screen image, characterized in that, Including: An acquisition unit for inputting a designed cross reticle screen and a checkerboard screen into a target display screen, and acquiring a display screen image and a saddle point image through an acquisition camera. There are several cross reticles on the display screen image, and several saddle points on the saddle point image; A correction unit for preliminarily correcting the distortion of the cross reticles on the display screen image through the saddle points on the saddle point image; A first generation unit for performing reticle analysis on the display screen image through a deep convolutional neural network model to generate reticle detection data, where the reticle detection data includes border information and confidence information; A determination unit for performing edge detection on the display screen image according to the border information in the reticle detection data, and then performing line detection to determine the reticle intersection information; A second generation unit for clustering the slope features in the reticle intersection information using a weighted clustering analysis method to generate several line categories and the clustering centers of the line categories; A third generation unit for generating the center position information of the cross reticle according to the several line categories and the clustering centers; A calculation unit for calculating the center coordinates of the target reticle according to the center position information, the border information in the reticle detection data, and the confidence information.

9. The positioning device according to claim 8, characterized in that The correction unit includes: Locating the saddle points on the saddle point image through an adaptive multi-scale feature fusion method; When there is a situation of missing saddle points, performing polynomial fitting on the already determined saddle points to generate several fitting curves; Performing interpolation processing and intersection calculation on the missing saddle points according to the several fitting curves to complete the missing saddle points; Generating a homography matrix according to all the completed saddle points and the reference saddle points on the standard template, and preliminarily correcting the distortion of the cross reticles on the display screen image through the homography matrix.

10. A computer-readable storage medium, characterized in that, A program is stored on the computer-readable storage medium, and when the program is executed on a computer, it executes the positioning method according to any one of claims 1 to 7.

Images