Corner point recognition method and computer readable storage medium, system

Abstract

The present application relates to a corner point recognition method and a computer readable storage medium and a system, one aspect of which is to build a corner point recognition model outputting corner point coordinates and confidence rates, take the confidence rate as a first evaluation index, judge whether the confidence rate of the corresponding corner point recognition model is greater than a first set threshold, if yes, retain the corner point coordinates output by the corner point recognition model, if not, discard the corner point coordinates output by the corner point recognition model, so as to screen the corner point coordinates output by the multi-corner point model in the combined model; another aspect is to determine the weight of each corner point recognition model through a neural network method, combine the screened corner point coordinates and the corresponding weight, and calculate the corner point coordinates of the image to be recognized. The whole is a multi-model fusion corner point recognition method, which can consider globally and analyze comprehensively to improve the accuracy, precision and recognition efficiency of corner point recognition.

Description

Technical Field

[0001] This invention relates to the field of image recognition, and in particular to a corner recognition method using multi-model fusion. Background Technology

[0002] Steel plate cutting is a crucial step in steel plate processing. Taking laser cutting as an example, the most critical issue during the cutting process is determining the relative position of the laser cutting blade to the steel plate. In existing technologies, the laser cutting blade is mounted above the cutting table via a movable slide rail, allowing it to move in a two-dimensional plane. The steel plate simply needs to be placed on the cutting table, and the cutting operation can be achieved by controlling the movement of the laser cutting blade.

[0003] However, determining the coordinates of the steel plate corners, using the coordinate system of the cutting table as a reference, to establish the initial position for laser cutting is a crucial step affecting the accuracy of subsequent cutting. Traditional methods typically use an industrial camera to photograph the steel plate, then process the image and use a specific algorithm to obtain the corner positions within the image. Coordinate transformation is then used to determine the relative position of the steel plate and the cutting blade. However, the lighting environment in steel plate production is usually complex and variable, significantly impacting the effectiveness of the corner detection algorithm. This results in low accuracy, significant errors, and a low success rate for traditional methods. Therefore, improving the accuracy of corner recognition is a pressing technical problem in this field. Summary of the Invention

[0004] To improve the accuracy of corner detection, this invention provides a multi-model fusion corner detection method, comprising:

[0005] S1: Construct and train multiple corner recognition models that take the image to be recognized as input and output the corner coordinates and confidence rate of the image to be recognized;

[0006] S2: Based on the neural network method, determine the weights of each corner recognition model;

[0007] S3: Obtain the current image to be recognized;

[0008] S4: Input the current image to be recognized into each corner recognition model, and output the corner coordinates and confidence rate of each corner recognition model;

[0009] S5: Determine whether the confidence rate output by each corner recognition model is greater than the first set threshold. If it is greater, retain the corner coordinates of the corresponding corner recognition model. If it is not greater, discard the corner coordinates of the corresponding corner recognition model.

[0010] S6: Calculate the corner coordinates of the image to be recognized based on the retained corner coordinates and the weights of the corresponding corner recognition model.

[0011] Further, in step S6, the corner coordinates of the image to be identified are calculated using formulas (1)-(2);

[0012] X=∑ i∈A x i *w i (1)

[0013] Y = ∑ i∈A y i *w i (2)

[0014] Where A is the set of retained corner coordinates; (x i ,y i ) represents the coordinates of the i-th corner point in the retained set of corner point coordinates; w i is the weight of the corner recognition model corresponding to the coordinates of the i-th corner point; (X,Y) are the corner coordinates of the image to be recognized.

[0015] Furthermore, step S5 also includes:

[0016] Calculate the average of the retained corner coordinates;

[0017] Determine whether the difference between the coordinates of each retained corner point and the average value is greater than a second set threshold;

[0018] Otherwise, retain the corner coordinates of the corresponding corner recognition model; otherwise, discard the corner coordinates of the corresponding corner recognition model.

[0019] Furthermore, step S5 also includes:

[0020] If the number of retained corner coordinates is greater than the third set threshold, proceed to step S6 to calculate the corner coordinates of the image to be identified; if it is not greater than the third set threshold, it is determined that there are no corners in the current image to be identified.

[0021] Further, step S2 includes:

[0022] S21: Acquire the image to be identified with the corner coordinates already marked, input each corner recognition model, and output the corner coordinates and confidence rate of each corner recognition model;

[0023] S22: Determine the weight of each corner recognition model based on the corner coordinates and confidence rate output by each corner recognition model, as well as the coordinates of the marked corners on the image to be recognized.

[0024] Furthermore, it also includes step S7, which includes:

[0025] Determine whether the confidence rate output by the corner recognition model corresponding to each retained corner coordinate is greater than a fourth set threshold; wherein the fourth set threshold is greater than the first set threshold.

[0026] If so, the corner coordinates output by the corner recognition model corresponding to a confidence rate greater than the fourth set threshold are defined as weighted coordinates;

[0027] If not, the corner coordinates output by the corner recognition model corresponding to a confidence rate not greater than the fourth set threshold are defined as unweighted coordinates.

[0028] The corner coordinates of the image to be identified are updated based on the weighted coordinates, unweighted coordinates, and the weights of the corresponding corner recognition model.

[0029] Furthermore, in step S7, the corner coordinates of the image to be identified are updated using formulas (3)-(4);

[0030] X ’ = P*∑ i∈B x i *w i +∑ i∈C x i *w i (3)

[0031] Y ’ = Q*∑ i∈B y i *w i +∑ i∈C y i *w i (4)

[0032] Where B is the retained weighted coordinate set; C is the retained unweighted coordinate set; P and Q are the weighting coefficients of the x and y coordinates, respectively; (X', Y') are the corner coordinates of the updated image to be identified.

[0033] Furthermore, the weighting coefficient P of the horizontal axis is the horizontal axis with the highest confidence rate divided by the average of all horizontal axes in the weighted coordinate system; the weighting coefficient Q of the vertical axis is the vertical axis with the highest confidence rate divided by the average of all vertical axes in the weighted coordinate system.

[0034] On the other hand, the present invention also provides a computer-readable storage medium having a computer program stored thereon for performing any of the corner recognition methods described above.

[0035] On the other hand, the present invention also provides a computer system including any of the above-described computer-readable storage media and one or more processors;

[0036] A processor is configured to run computer programs.

[0037] The corner recognition method, computer-readable storage medium, and system of the present invention, on the one hand, construct a corner recognition model that outputs corner coordinates and a confidence rate. Using the confidence rate as the first evaluation index, it determines whether the confidence rate of the corresponding corner recognition model is greater than a first preset threshold. If it is greater, the corner coordinates output by that corner recognition model are retained; if not, the corner coordinates output by that model are discarded. This process filters the corner coordinates output by multiple corner recognition models in a combined model. On the other hand, it uses a neural network method to determine the weights of each corner recognition model, combining the filtered corner coordinates with their corresponding weights to calculate the corner coordinates of the image to be recognized. Overall, it is a multi-model fusion corner recognition method that can consider the overall picture and perform comprehensive analysis to improve the accuracy, precision, and efficiency of corner recognition. Attached Figure Description

[0038] Figure 1 This is a flowchart of an embodiment of the corner recognition method of the present invention;

[0039] Figure 2 This is a schematic diagram illustrating the application of the corner recognition method of the present invention in steel plate corner recognition. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0041] It should be noted that if the embodiments of the present invention involve directional indications, such as up, down, left, right, front, back, etc., these directional indications are only used to explain the relative positional relationships and movement of the components in a specific posture. If the specific posture changes, the directional indications will also change accordingly. Furthermore, if the embodiments of the present invention involve descriptions such as "first," "second," "S1," "S2," "step one," "step two," etc., these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance, or implicitly indicating the number of technical features indicated or the order of method execution. Those skilled in the art will understand that anything that does not violate the inventive concept and is within the scope of the present invention should be included in the protection scope of the present invention.

[0042] like Figure 1 As shown, the corner recognition method based on multi-model fusion of the present invention is presented, including:

[0043] S1: Construct and train multiple corner recognition models that take the image to be recognized as input and output the corner coordinates and confidence rate of the image to be recognized;

[0044] Specifically, such as Figure 2 As shown, the corner recognition of steel plates in industrial scenarios can be used as an example for explanation. A corner recognition model is constructed with the steel plate image as input and the corner coordinates and the confidence rate of the corner coordinates, i.e. the accuracy, of the steel plate in the image as output.

[0045] More specifically, the corner recognition model can be, but is not limited to, building different models in the YOLO series, such as YOLOV5, YOLOV6, and YOLOV7 models. Taking the YOLO object detection model as an example, it can be, but is not limited to, setting the YOLO network to detect a certain pixel. For example, for a target image region of 25*25 pixels, the center coordinates of the target image region are used as the corner coordinates of the image to be recognized.

[0046] More specifically, optional but not limited to first using public datasets in the network, such as the COCO dataset and VOC dataset, to pre-train the model and obtain a pre-trained model, and then combining the idea of ​​transfer training, on the basis of the pre-trained model, inputting the labeled target industrial dataset, that is, the image to be identified with labeled corner coordinates, to train and obtain multiple different corner recognition models, and assigning numbers to these models.

[0047] In this embodiment, by using public datasets and the concept of transfer training, a labeled target industrial dataset (i.e., an image to be identified with labeled corner coordinates) is input into the pre-trained model. Multiple corner recognition models can be trained using a small industrial dataset, which is a simple, efficient, and accurate model building and training method.

[0048] S2: Based on a neural network method, determine the weights of each corner recognition model; specifically, optional but not limited to, the following:

[0049] S21: Acquire the image to be identified with the corner coordinates already marked, input each corner recognition model, and output the corner coordinates and confidence rate of each corner recognition model;

[0050] S22: Determine the weight of each corner recognition model based on the corner coordinates and confidence rate output by each corner recognition model, as well as the coordinates of the marked corners on the image to be recognized.

[0051] More specifically, optionally but not limited to, using untrained labeled industrial datasets (i.e., images of the target image with labeled corner coordinates) as input into the aforementioned corner recognition models, corner recognition is performed. Then, the outputs of each model (predicted corner coordinates and confidence rates) and the labeled corner coordinates (true corner coordinates) on the target image are combined to form a new dataset. This dataset is then recursively converged using a genetic algorithm or AdaBoost algorithm, iteratively reducing the loss function to determine the weights of each model in the combined model. More specifically, optionally but not limited to, these weights are combined to form a weight vector W, which is stored according to the corner recognition model's number.

[0052] S3: Obtain the current image to be recognized;

[0053] Specifically, taking the corner recognition of steel plates in industrial scenarios as an example, it is optional, but not limited to, using an industrial camera to take photos of the steel plates and sending the photos to each corner recognition model; specifically, it is also optional, but not limited to using base64 encoding to encode and encrypt the current image to be recognized on the client side, and then distributing the encoded data to each corner recognition model server in the form of parameters.

[0054] S4: Input the current image to be recognized into each corner recognition model, and output the corner coordinates and confidence rate of each corner recognition model;

[0055] Specifically, each corner recognition model server can optionally, but is not limited to, receive the current image to be recognized captured in S3, receive the data and decode it, input the decoded image into the corner recognition model for corner recognition, and output the corner coordinates and confidence rate of each corner recognition model.

[0056] S5: Determine whether the confidence rate output by each corner recognition model is greater than the first set threshold. If it is greater, retain the corner coordinates of the corresponding corner recognition model. If it is not greater, discard the corner coordinates of the corresponding corner recognition model.

[0057] Specifically, a first threshold may be determined, but is not limited to, based on attributes such as corner recognition accuracy and the size of the current image to be recognized. If the confidence rate of a corner recognition model is greater than the first threshold, the corner coordinates predicted by that model are considered reliable and are retained. If the confidence rate of a corner recognition model is not greater than the first threshold, the corner coordinates predicted by that model are considered unreliable and are discarded. More specifically, the retained corner coordinates, their corresponding confidence rates, and the corner recognition model number may be integrated into JSON data and returned to the client.

[0058] S6: Calculate the corner coordinates of the image to be recognized based on the retained corner coordinates and the weights of the corresponding corner recognition model.

[0059] This embodiment presents a corner recognition method based on multi-model fusion. On one hand, it constructs corner recognition models that output corner coordinates and confidence rates. Using the confidence rate as the first evaluation metric, it determines whether the confidence rate of the corresponding corner recognition model is greater than a first set threshold. If it is greater, the corner coordinates output by that model are retained; otherwise, they are discarded. This process filters the corner coordinates output by multiple corner models in the combined model. On the other hand, it uses a neural network method to determine the weights of each corner recognition model. By combining the filtered corner coordinates with their corresponding weights, the corner coordinates of the image to be recognized are calculated. This multi-model fusion corner recognition method can consider the overall picture and provide comprehensive analysis to improve the accuracy, precision, and efficiency of corner recognition. Taking steel plate recognition in an industrial setting as an example, improving the accuracy, precision, and efficiency of corner recognition can enhance the accuracy and speed of corner positioning in agricultural environments, thereby improving the accuracy of steel plate cutting, the utilization rate of steel plate materials, product quality, and production efficiency. Based on this, in step S1, it is preferable to use the YOLO series architecture corner recognition model, which has low computational cost, fast recognition speed, and can be applied to devices with low computing power; it is preferable to use public datasets and transfer training ideas, which can train a better and more accurate corner recognition model with a smaller dataset, especially suitable for working conditions such as limited training data and new production lines.

[0060] Preferably, in step S6, the corner coordinates of the image to be identified can be calculated using formulas (1)-(2), but are not limited to this method.

[0061] X=∑ i∈A x i *w i (1)

[0062] Y = ∑ i∈A y i *w i (2)

[0063] Where A is the set of retained corner coordinates; (x i ,y i ) represents the coordinates of the i-th corner point in the retained set of corner point coordinates; w i Let (X,Y) be the weights of the corner recognition model corresponding to the coordinates of the i-th corner point; (X,Y) be the corner coordinates of the image to be recognized. Specifically, it is optional, but not limited to, setting the first threshold to K; if k i If the value is greater than K, then the corner coordinates of the i-th corner recognition model are retained and stored in the retained corner coordinate set A; if the value is not greater than K, then they are discarded; based on the retained corner coordinates and the weight of the corresponding corner recognition model, the corner coordinates (X, Y) of the image to be recognized are calculated using formulas (1)-(2).

[0064] Preferably, in another preferred embodiment:

[0065] Step S5 may also optionally include, but is not limited to:

[0066] Calculate the average of the retained corner coordinates;

[0067] Determine whether the difference between the coordinates of each retained corner point and the average value is greater than a second set threshold; specifically, it may be optional but not limited to calculating whether the difference between the coordinates of each retained corner point and the average value is greater than the second set threshold; or calculating whether the ratio of the coordinates of each retained corner point to the average value is greater than the second set threshold.

[0068] If not, retain the corner coordinates of the corresponding corner recognition model; if yes, discard them.

[0069] Specifically, similarly, a second threshold is set. If the difference between the coordinates of a corner point and the average value is not greater than the second threshold, it indicates that the corner point coordinates are not special corner points, such as corner points with scratches or abrasions, and are retained within the effective corner point range. If the difference is greater than the second threshold, it indicates that the corner point coordinates are special corner points, such as corner points with scratches or abrasions, and are not corner points that we need to locate normally for corner point identification. They are invalid corner points and are discarded.

[0070] In this embodiment, an average value is added as a second evaluation index to determine whether the difference between the retained corner coordinates and the average value is greater than a second set threshold. A secondary screening is then performed on the corner coordinates predicted by all corner recognition models to further improve the accuracy of corner recognition. More specifically, optional but not limited to the client receiving and parsing JSON data sent by each corner recognition model server, after receiving all corner recognition results, the received data is sorted according to the model number, and the average value of the retained corner coordinates is calculated. If the corner coordinates identified by a certain corner recognition model are too far from the average value, these corner coordinates are considered invalid and are deleted a second time. The corner coordinates retained in step S6, i.e., the corner coordinates in the retained corner coordinate set A, are corner coordinates after secondary screening, and their accuracy, precision, and recognition efficiency are all superior.

[0071] More preferably, in another preferred embodiment:

[0072] Step S5 may also optionally include, but is not limited to:

[0073] If the number of retained corner coordinates is greater than the third set threshold, proceed to step S6 to calculate the corner coordinates of the image to be identified; if it is not greater than the third set threshold, it is determined that there are no corners in the current image to be identified.

[0074] Specifically, the third threshold can be determined, but is not limited to, based on attributes such as corner recognition accuracy, the size of the current image to be recognized, and the number of corner recognition models. For example, the third threshold can be, but is not limited to, half of the total number of corner recognition models. After one or / and two screenings in step S5, if the number of retained corner coordinates exceeds the third threshold, it indicates that the reliable corner coordinates are in the majority and can be adopted, proceeding to step S6 to calculate the corner coordinates determined by multi-model fusion; if the number of retained corner coordinates does not exceed the third threshold, it indicates that the reliable corner coordinates are in the minority and cannot be adopted, determining that there are no corners in the image to be recognized, and no calculation is required. More specifically, taking a third threshold of half the total number of corner recognition models as an example, if more than half of the corner recognition models recognize corner coordinates, and all recognized corner coordinates are valid, it indicates that the corner is valid. The weight vector W obtained above is combined with the corner coordinates obtained by each model to perform a linear combination calculation to obtain the final corner coordinates (X, Y), which is then output to the client. If the sum of the number of corners not recognized by the corner recognition model and the number of invalid corner coordinates reaches half or more of the number of corner recognition models, then it is considered that there are no corners in the image to be recognized, and there is no need to calculate the coordinates of the corners.

[0075] More preferably, in another preferred embodiment:

[0076] It also includes step S7, which may include, but is not limited to:

[0077] Determine whether the confidence rate output by the corner recognition model corresponding to each retained corner coordinate is greater than a fourth set threshold; wherein the fourth set threshold is greater than the first set threshold.

[0078] If so, the corner coordinates output by the corner recognition model corresponding to a confidence rate greater than the fourth set threshold are defined as weighted coordinates;

[0079] If not, the corner coordinates output by the corner recognition model corresponding to a confidence rate not greater than the fourth set threshold are defined as unweighted coordinates.

[0080] The corner coordinates of the image to be identified are updated based on the weighted coordinates, unweighted coordinates, and the weights of the corresponding corner recognition model.

[0081] In this embodiment, the fourth threshold can be optionally determined based on, but is not limited to, attributes such as corner recognition accuracy, the size of the current image to be recognized, and the number of corner recognition models. In this preferred embodiment, a fourth threshold is set for the confidence rate. Corner coordinates with a confidence rate greater than the fourth threshold (i.e., higher confidence) are defined as weighted coordinates. Corner coordinates with a confidence rate not greater than the fourth threshold (i.e., those whose confidence exceeds the first threshold but not the fourth threshold) are defined as ordinary coordinates—coordinates that, while not discarded, are not representative or highly reliable. The corner coordinates of the image to be recognized are updated based on the weighted coordinates, unweighted coordinates, and the weights of the corresponding corner recognition models. This further improves the accuracy and precision of corner recognition based on the confidence rate, ensuring that the coordinates are closer to the weighted coordinates output by high-confidence corner recognition models while also fully considering the ordinary coordinates of other corner recognition models, thus balancing the comprehensiveness and specificity of multi-model fusion.

[0082] More preferably, in another preferred embodiment:

[0083] In step S7, the corner coordinates of the image to be identified can be updated using formulas (3)-(4), but are not limited to this method.

[0084] X ’ = P*∑ i∈B x i *w i +∑ i∈C x i *w i (3)

[0085] Y ’ = Q*∑ i∈B y i *w i +∑ i∈C y i *w i (4)

[0086] Where B is the retained weighted coordinate set; C is the retained unweighted coordinate set; P and Q are the weighting coefficients of the x and y coordinates, respectively; (X', Y') are the corner coordinates of the updated image to be identified.

[0087] More specifically, the weighting coefficient P of the horizontal axis can be, but is not limited to, the horizontal axis with the highest confidence rate among the weighted coordinates divided by the average of the horizontal axes among all weighted coordinates; the weighting coefficient Q of the vertical axis can be, but is not limited to, the vertical axis with the highest confidence rate among the weighted coordinates divided by the average of the vertical axes among all weighted coordinates.

[0088] For example, if initially 10 corner recognition models are built, outputting 10 sets of corner coordinates and confidence rates, and assuming that after one or two screenings using the first and / or second set thresholds, 7 sets of corner coordinates and confidence rates remain, then it is determined whether the number of these 7 sets exceeds the third set threshold, preferably half of the total number of corner recognition models, i.e., 5. If it exceeds this threshold, it indicates that corners do exist. Further, the remaining 7 sets of corner coordinates and confidence rates are further divided into weighted coordinates and unweighted coordinates according to the fourth set threshold. For example, assuming there are 3 weighted coordinates and 4 unweighted coordinates, the 3 weighted coordinate statistics are multiplied by a weighting coefficient, preferably determined based on the coordinate corresponding to the highest confidence rate among the 3 weighted coordinates and the average of the three weighted coordinates. The 4 unweighted coordinate statistics are calculated in the normal way. Finally, they are summed to determine the updated corner coordinates of the image to be recognized.

[0089] This embodiment provides a preferred method for updating the corner coordinates of the image to be identified in step S7. Instead of simply combining the retained corner coordinates with their corresponding weights to calculate the corner coordinates of the image to be identified, it multiplies the statistical value of the weighted coordinates by a weighting coefficient and then adds it to the statistical value of the unweighted coordinates. Preferably, the weighting coefficient is the coordinate with the highest confidence rate among the weighted coordinates divided by the average of all weighted coordinates. The gain brought by the weighted coordinates is obtained by multiplying the ratio by the statistical value of the weighted coordinates. This update calculation method significantly widens the gap between the weighted coordinates with high confidence rates and the unweighted coordinates with relatively high confidence rates. Experimental comparisons show that it can further improve the accuracy and efficiency of corner recognition. Taking steel plate corner recognition as an example, it significantly improves the accuracy and efficiency of subsequent steel plate cutting, effectively avoiding material waste and improving production quality and efficiency. It is worth noting that the corner recognition method of this invention is only explained using steel plate corner recognition as an example, but it is not limited thereto. Those skilled in the art will understand that it can be used in other fields, but is not limited to these applications. Further examples include package corner recognition for accurate package grabbing, which can be applied to logistics sorting, cargo sorting and other fields.

[0090] On the other hand, the present invention also provides a computer-readable storage medium having a computer program stored thereon for performing any of the corner recognition methods described above.

[0091] On the other hand, the present invention also provides a computer system including the above-described computer-readable storage medium and one or more processors;

[0092] A processor is configured to run computer programs.

[0093] The computer-readable storage medium and computer system described above are created based on the corner recognition method described above. Their technical functions and beneficial effects will not be elaborated here. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0094] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A corner recognition method using multi-model fusion, characterized in that, include: S1: Construct and train multiple corner recognition models that take the image to be recognized as input and the corner coordinates and confidence rate of the image to be recognized as output; S2: Based on the neural network method, determine the weights of each corner recognition model; S3: Obtain the current image to be recognized; S4: Input the current image to be recognized into each corner recognition model, and output the corner coordinates and confidence rate of each corner recognition model; S5: Determine whether the confidence rate output by each corner recognition model is greater than the first set threshold. If it is, retain the corner coordinates of the corresponding corner recognition model; otherwise, discard the corner coordinates of the corresponding corner recognition model. Calculate the average value of the retained corner coordinates. Determine whether the difference between each retained corner coordinate and the average value is greater than the second set threshold. Otherwise, retain the corner coordinates of the corresponding corner recognition model; otherwise, discard the corner coordinates of the corresponding corner recognition model. Count the number of retained corner coordinates. If it is greater than the third set threshold, proceed to step S6 to calculate the corner coordinates of the image to be recognized. If it is not greater than the third set threshold, determine that there are no corners in the current image to be recognized. S6: Calculate the corner coordinates of the image to be recognized based on the retained corner coordinates and the weights of the corresponding corner recognition model; S7: Determine whether the confidence rate output by the corner recognition model corresponding to each retained corner coordinate is greater than the fourth set threshold; wherein the fourth set threshold is greater than the first set threshold; if yes, define the corner coordinates output by the corner recognition model corresponding to the confidence rate greater than the fourth set threshold as weighted coordinates; if no, define the corner coordinates output by the corner recognition model corresponding to the confidence rate not greater than the fourth set threshold as unweighted coordinates; update the corner coordinates of the image to be recognized according to the weighted coordinates, unweighted coordinates and the weights of the corresponding corner recognition models.

2. The corner recognition method according to claim 1, characterized in that, In step S6, the corner coordinates of the image to be identified are calculated using formulas (1)-(2); X=∑ i∈A x i *w i (1); Y=∑ i∈A y i *w i (2); Where A is the set of corner coordinates to be retained; (x i ,y i ) represents the coordinates of the i-th corner point in the retained set of corner point coordinates; w i is the weight of the corner recognition model corresponding to the coordinates of the i-th corner point; (X,Y) are the corner coordinates of the image to be recognized.

3. The corner recognition method according to claim 1, characterized in that, Step S2 includes: S21: Acquire the image to be identified with the corner coordinates already marked, input each corner recognition model, and output the corner coordinates and confidence rate of each corner recognition model; S22: Determine the weight of each corner recognition model based on the corner coordinates and confidence rate output by each corner recognition model, as well as the coordinates of the marked corners on the image to be recognized.

4. The corner recognition method according to any one of claims 1-3, characterized in that, In step S7, the corner coordinates of the image to be identified are updated using formulas (3)-(4); X ’ = P*∑ i∈B x i *w i +∑ i∈C x i *w i (3); Y ’ = Q*∑ i∈B y i *w i +∑ i∈C y i *w i (4); Where B is the retained weighted coordinate set; C is the retained unweighted coordinate set; P and Q are the weighting coefficients of the x and y coordinates, respectively; (X', Y') are the corner coordinates of the updated image to be identified.

5. The corner recognition method according to claim 4, characterized in that, The weighting coefficient P for the horizontal axis is the horizontal axis with the highest confidence level divided by the average of all horizontal axes in the weighted coordinate system; the weighting coefficient Q for the vertical axis is the vertical axis with the highest confidence level divided by the average of all vertical axes in the weighted coordinate system.

6. A computer-readable storage medium, characterized in that, It stores a computer program for performing the corner identification method according to any one of claims 1-5.

7. A computer system, characterized in that, Includes the computer-readable storage medium of claim 6 and one or more processors; The processor is configured to run the computer program.

Images