A method and system for identifying left and right directions of vehicle accessories

Abstract

The present invention provides a method and system for identifying the left and right directions of vehicle accessories, the method comprising the following steps: S1: a training model step, including model training for mirror behavior judgment and mirror image judgment of feature objects, and model training for vehicle orientation, central axis, and vehicle accessories recognition; S2: a model recognition step, extracting feature objects of received images for mirror image judgment, identifying vehicle accessories after judging the orientation and central axis of the vehicle in the image, and judging the final left and right orientation of the accessories; S3: summarizing all extracted accessory information and determined accessory directions, and feeding back user results. The present invention can significantly improve the recognition accuracy of mirror images, making the performance of the model more stable in different environments, thereby improving the recognition accuracy, and enabling the model to better learn the characteristics of left-right symmetrical objects during the training process, thereby showing stronger recognition capabilities in practical applications.

Description

Technical Field

[0001] The invention relates to a vehicle accessory identification method, and in particular to a method and system for identifying the left and right directions of a vehicle accessory. Background Art

[0002] With the development of the insurance industry, every vehicle will be insured. In daily life, vehicle accidents always happen. In order to solve the relevant issues of vehicle accidents, insurance companies have launched a set of insurance claims process and system. The process is mainly divided into six links: reporting, investigation, damage assessment, damage verification, claims verification, and settlement.

[0003] (1) In the reporting stage, when a vehicle accident occurs, the person involved in the accident calls the insurance company and tells the insurance company operator their vehicle information, personal information, insurance information, cause of the accident, accident details, etc. The operator will briefly record the information described by the reporter in the insurance claims system.

[0004] (2) Investigation stage: After the report is made, the operator will arrange for the insurance company’s investigators to conduct an on-site investigation at the accident site based on the accident details provided by the reporter. After arriving at the accident site, the investigators will further investigate and collect data on the vehicle information, personnel information, insurance information, cause of the accident, and accident details, and will take comprehensive photos of the accident site, vehicles, personnel, documents, etc. All information will also be entered into the insurance claims system.

[0005] (3) Damage assessment: After the survey, the surveyor will communicate with the reporter to choose which repair shop to repair the vehicle. The repair shop will then inspect the accident vehicle to determine which vehicle parts should be replaced or repaired. During this phase, the insurance company will arrange for new personnel to take charge, called damage assessors. In some cases, the damage assessor can also work part-time as a surveyor. The damage assessor will enter the repair suggestions given by the repair shop into the insurance claim system. The input information includes vehicle information, repair shop information, accessories information, working hours information, external repair information, and auxiliary materials information. The vehicle information mainly includes the brand, model, announcement number, vehicle type, VIN code, etc. The repair shop information mainly includes the name, address, contact number, cooperation, type of repair shop, etc. The accessories information mainly includes the vehicle parts recommended by the repair shop for replacement, vehicle parts code, quantity, unit price, price source, etc. The labor time information mainly includes the vehicle parts recommended by the repair shop for repair, repair type, repair hours, repair hour unit price, total repair hours, among which the repair types are mainly divided into: sheet metal (vehicle parts will be repaired through this type of repair when they are dented, cracked, or broken), paint (vehicle parts are scratched, Stains, and sheet metal will be repaired through this type of repair later), disassembly and assembly (when vehicle parts are replaced or repaired, parts need to be disassembled for processing, and some will have disassembly and assembly repair costs); external repair information mainly includes items that the repair shop itself cannot repair and must be repaired by the vehicle manufacturer or a more professional repair shop. Generally, insurance companies will enter this item separately; auxiliary material information mainly includes glue, coolant, clips and other auxiliary repair materials generated by the repair shop during the repair process; all of the above information is entered into the system by the adjuster, and after entry, it will be reviewed by the next link personnel.

[0006] (4) Damage assessment. After the damage assessment, the insurance company sets up this step to further review the damage assessment data. Some insurance companies may have price assessment and damage assessment, or there may be a situation where there is a single position for price assessment and damage assessment. Price assessment is mainly done by the price assessor who inquires about the price of the accessories based on his own experience or the vehicle factory data, the cooperative sales channels of vehicle accessories, etc., to determine whether the price given by the repair shop has risks. Damage assessment is mainly done by the damage assessor who analyzes the accident scene, accident pictures, pictures of damaged parts, and the logical relationship of parts loss based on his own experience, to determine whether the repair shop deliberately exaggerates the loss, whether the repair must replace the parts, or whether the damaged parts are not caused by the accident, etc., to determine whether there is a risk. If it is determined that there is a risk in this step, it will return to the damage assessment step on the system, and the damage assessor will communicate with the repair shop about the relevant risks. After the communication, the damage assessor will modify the input information and resubmit the damage assessment step until the damage assessor determines that the damage assessment data is normal.

[0007] (5) Claims settlement: After the loss assessment, the loss assessor’s review results will be verified first, and then the insurance terms, accident liability, etc. will be reviewed to determine the amount of the case data.

[0008] (6) Adjustment: After the claims verification phase, the amount determined in the loss verification phase will be paid to the insurer / repair shop.

[0009] In response to the above-mentioned insurance claims process, a car insurance claims risk control detection system has been launched. The car insurance claims risk control detection system intervenes in different links of the insurance claims process, receives case data in the insurance company's claims system through data transmission, performs detection based on the received case data, feeds back the detected risks to the insurance company's claims system, and prompts the operators of the relevant links of the insurance company's claims system, so as to improve and modify the case according to the prompted risks. In the car insurance claims risk control detection system, the main modules are divided into case receiving module, case conversion module, case binding module, rule engine module, case feedback module, case management module, rule management module, and data management module.

[0010] 1) The case receiving module is mainly responsible for data connection with the insurance company's claims system. The developers of both parties will formulate data receiving methods, including but not limited to dedicated lines, the Internet, VPN, etc. for data communication, and data transmission in the form of files, pictures, JSON using HTTP or HTTPS. The data is parsed in the transmission format agreed upon by both parties, and then transferred to the case conversion module.

[0011] 2) Case conversion module: The auto insurance claims risk control detection system is aimed at multiple insurance companies. The data structures of different insurance companies are different. This module creates a comprehensive and standard data structure to convert the data structures of all insurance companies into a standard structure.

[0012] 3) Case binding module: Different insurance companies have different comprehensiveness and standards for information entry. This module will standardize the insurance company's data and supplement the current data based on the data maintained by the data management module.

[0013] 4) The rule engine module is a set of processes for rule execution. It can set up exclusive rule pools according to different insurance companies, insurance company branches, and case links, and execute rules according to the rule execution process pre-set by the rule management module. Each rule has two states: triggered and not triggered. When a rule is triggered, the risk to which the rule belongs will be output at the same time.

[0014] 5) Case feedback module: Based on the feedback from the rule engine module, the module determines whether to feedback the current results to the insurance company, whether to merge the rule risks, optimize the rule prompt statements, and other operations. In addition, in order to adapt to the feedback results of different insurance companies, the feedback structure of different insurance companies can be set separately.

[0015] 6) Case management module is responsible for case storage, case display, rule triggering status display, case label maintenance, data encryption storage, regular data deletion, display of changed damage assessment data and corresponding amounts according to changes before and after different links of the case, as well as data statistical analysis and other operations.

[0016] 7) The rule management module is mainly responsible for the writing, deletion, review, online and offline operations of rules. Industry experts can combine and form rules in this module based on their own experience and by limiting the conditions of the case.

[0017] 8) Data management module, which is mainly responsible for adding, modifying, deleting, and querying the industry's vehicle data, vehicle parts data, repair shop data, price data, blacklist, etc.

[0018] In the existing technologies of damage assessment and loss verification, the traditional approach of image recognition usually relies on a large amount of manually annotated photo data. This process is not only time-consuming and labor-intensive, but also may lead to inconsistencies and deviations in the annotation results due to the subjectivity of manual annotation. Especially when dealing with images with high similarity, such as accessories in the left and right directions, the limitations of traditional methods become more obvious. Due to the high similarity of images in the left and right directions, the models trained based on these annotated data often lack the ability to recognize subtle differences in the left and right directions. This weak recognition ability not only reduces the accuracy of the model in practical applications, but may also lead to misjudgments, thereby affecting the insurance risk control decisions of the subsequent auto insurance claims risk control detection system. Therefore, there is an urgent need for a new image recognition method to improve the recognition accuracy and robustness of the model in the left and right directions.

[0019] Some method descriptions of the technical solutions of the prior art:

[0020] (1) Image classification methods based on convolutional neural networks (CNNs): Convolutional neural networks (CNNs) are widely used in existing technologies for image classification tasks. These methods rely on a large amount of labeled data for training, but when processing bilaterally symmetrical or highly similar images, the performance of the model is limited, which can easily lead to recognition errors.

[0021] (2) Data augmentation technology: In order to improve the generalization ability of the model, existing technologies often use data augmentation technology, such as rotation, flipping, and cropping. However, these methods have limited effects on the problem of image similarity in the left and right directions, and it is difficult to effectively improve the recognition ability of the model.

[0022] (3) Traditional feature extraction methods: Many existing technologies rely on traditional feature extraction methods (such as SIFT, HOG, etc.) for image recognition. Although these methods are effective in some scenarios, the accuracy and robustness of feature extraction are insufficient for complex backgrounds and bilaterally symmetrical objects.

[0023] (4) Recognition method based on template matching: Some existing technologies use template matching methods for image recognition. This method is effective when dealing with simple scenes, but when faced with complex left-right symmetrical images, the flexibility and adaptability of template matching are poor, resulting in poor recognition results.

[0024] (5) Transfer learning technology: In the existing technology, transfer learning is widely used in image recognition tasks, especially when data is scarce. However, when processing left-right symmetrical or highly similar images, the effect of transfer learning is still affected by the original model training data and cannot fully solve the problem of insufficient left-right recognition ability.

[0025] The prior art has the following disadvantages:

[0026] Insufficient model generalization ability: Existing image classification methods based on convolutional neural networks (CNN) often have insufficient generalization ability when processing bilaterally symmetrical or highly similar images. This makes the model prone to misjudgment in practical applications, affecting recognition accuracy.

[0027] (2) Subjectivity and consistency issues of manual labeling: Traditional manual labeling methods have subjectivity in labeling results, which may lead to inconsistency between different labelers for the same image. This inconsistency directly affects the quality of training data, thereby reducing the learning effect and recognition performance of the model.

[0028] (3) Data augmentation has limited effect: Although data augmentation technology can improve the generalization ability of the model to a certain extent, for the problem of image similarity in the left and right directions, simple augmentation methods (such as flipping and rotation) do not effectively improve the recognition ability of the model, but may introduce more noise.

[0029] (4) Limitations of feature extraction: Traditional feature extraction methods (such as SIFT, HOG, etc.) have poor accuracy and robustness when dealing with complex backgrounds and bilaterally symmetrical objects. This makes these methods perform poorly when faced with diverse and complex images.

[0030] (5) Insufficient flexibility of template matching: Recognition methods based on template matching may be effective when processing simple images. However, in complex scenes, especially in the recognition of bilaterally symmetrical objects, the flexibility and adaptability of template matching are poor, resulting in inaccurate recognition results.

[0031] (6) Transfer learning relies on original data: Although transfer learning can help improve model performance when data is scarce, its effectiveness is highly dependent on the training data of the original model. If the original data is insufficient in the ability to recognize left and right directions, the effect of transfer learning will also be limited, and it will not be able to fully solve the problem of insufficient ability to recognize left and right directions. Summary of the invention

[0032] The present invention provides a method and system for identifying the left and right directions of vehicle accessories, which solves the following problems:

[0033] (1) Insufficient recognition accuracy of bilaterally symmetrical objects: Existing image recognition technology is prone to misjudgment when processing bilaterally symmetrical objects, resulting in reduced recognition accuracy. The present invention aims to enhance the recognition ability of bilaterally symmetrical objects and improve the accuracy and robustness of the model by improving the algorithm.

[0034] (2) Dependence on data enhancement effects: In the prior art, although data enhancement can improve the performance of the model, the effect is limited when processing bilaterally symmetrical images. The present invention uses a data enhancement strategy combined with contextual information to effectively improve the performance of the model in bilaterally symmetrical object recognition.

[0035] (3) Misjudgment caused by feature similarity: In mirror images, the feature similarity of the left and right symmetrical parts is high, which may cause the recognition algorithm to make misjudgments when performing feature matching. For example, the algorithm may mistakenly identify the left feature as the right feature, thus affecting the overall recognition accuracy.

[0036] (4) Lack of ability to recognize symmetry features: Traditional image recognition algorithms often focus on extracting local features, but lack the ability to extract the overall symmetry features of mirror images. This makes it impossible for the model to fully utilize the symmetry information when processing bilaterally symmetrical objects, resulting in unsatisfactory recognition results.

[0037] (5) Neglect of contextual information: In many cases, the contextual information of the mirror image is crucial for accurate recognition. Existing algorithms often fail to effectively integrate contextual information, resulting in the loss of important clues in the recognition of mirror images, thus affecting the accuracy of recognition.

[0038] The technical solution is as follows:

[0039] A method for identifying the left and right directions of a vehicle accessory comprises the following steps:

[0040] S1: training model step, including model training for mirror behavior judgment and mirror image judgment of feature objects, as well as model training for vehicle orientation, central axis, and vehicle accessories recognition;

[0041] S2: Model recognition step, extracting the features of the received image to perform mirror image judgment, after judging the orientation and central axis of the vehicle in the image, identifying the vehicle accessories and judging the final left and right orientation of the accessories;

[0042] S3: Summarize all extracted accessory information and determined accessory directions, and feed back the results to the user.

[0043] Furthermore, in step S1, the model training step includes model training of the following models:

[0044] S11: Model training for feature extraction: the feature objects are selected from the license plate number, store plaque, and road sign in the vehicle image. By separately training the feature objects that are different in the mirror image, it can be used to determine whether there is mirroring behavior in the mirror image;

[0045] S12: Model training for judging mirror images: mirror the feature objects separately to generate mirror images, randomly select mirror data on one side of the mirror image, discard the mirror data on the other side, and judge whether the selected mirror data is the original image or the mirror image;

[0046] S13: Vehicle orientation model training: According to the orientation of the vehicle, the orientation direction is divided into forward and backward;

[0047] S14: Model training of the vehicle center axis: used to determine the center position of the vehicle in the image and draw a center axis in the center position of the vehicle;

[0048] S15: Vehicle accessories recognition model: All individual accessories in the vehicle are marked, and the accessories are trained for recognition. The model has the ability to extract accessories from images separately, and trains to obtain the corresponding contour rectangles and accessory classifications.

[0049] Furthermore, in step S15, the outline rectangle is a minimum area rectangle obtained by the accessory using the outline, and the left point and the right point of the rectangle are found according to the current rectangular shape.

[0050] The model identification step includes the following steps:

[0051] S21: receiving pictures;

[0052] S22: extracting features from the image using the first model, so as to determine whether the image has mirroring behavior in a subsequent step;

[0053] S23: calling the second model for judging the mirror image to judge whether the extracted feature object is a mirror image;

[0054] S23: determining the orientation and center axis coordinates of the vehicle in the image, determining whether the vehicle is in the front or rear according to the third model for determining the orientation of the vehicle, and then determining the center axis coordinates of the vehicle according to the fourth model for determining the center axis of the vehicle;

[0055] S24: Identify vehicle parts, determine the coordinates of the vehicle parts, and classify the parts;

[0056] S25: Determine the final left and right position of the accessory.

[0057] Further, in step S22, if the image has a mirroring behavior, if it is a mirrored image, then step S26 is applied after step S25; if it is not a mirrored image, then step S26 is not applied and the process goes to step S23;

[0058] Step S26: Depending on whether the image is a mirror image, the left and right directions of the accessory are reversed to obtain the final direction of the accessory.

[0059] Furthermore, in step S25, determining the final left and right position of the accessory includes the following steps:

[0060] S31: judging whether the accessory has a left-right direction according to different categories of accessories, if there is no left-right direction accessory, directly save the accessory information; if there is a left-right direction accessory, go to step S32;

[0061] S32: If there are accessories in the left and right directions, the accessories are identified in the left and right directions. After the origin is set in the picture, the following judgment method can be determined according to the coordinates of the accessories and the coordinates of the central axis:

[0062] Accessory coordinates: (x1, y1) represents the upper left corner, (x2, y2) represents the lower right corner, which is a matrix;

[0063] Central axis coordinates: (a1, b1) represents the starting point, (a2, b2) represents the end point, which is a line;

[0064] if

[0065] a1 < (x2-x1) / (y2-y1) * b1;

[0066] and

[0067] a2 < (x2-x1) / (y2-y1) * b2;

[0068] If the fitting is located on the left side of the central axis (PL), then the fitting is located on the right side of the central axis (PR).

[0069] S33: determining the front and rear direction of the accessory according to the vehicle orientation;

[0070] If the vehicle is facing forward and the accessory is PL, the accessory belongs to the right side;

[0071] If the vehicle is facing forward and the accessory is PR, the accessory belongs to the left side;

[0072] If the vehicle is facing backwards and the accessory is PL, the accessory belongs to the left side;

[0073] If the vehicle is facing rearward and the accessory is PR, the accessory is on the right side.

[0074] Further, in step S31, the accessories that do not exist in the left and right directions are accessories that are used only by the car, including a front bumper cover, a rear bumper cover, and a center grille.

[0075] A system for identifying the left and right directions of a vehicle accessory, comprising:

[0076] The first model is used to extract features, such as license plate numbers, store signs, and road signs in vehicle images;

[0077] The second model is used for mirror image judgment. It performs mirror processing on the feature object separately to generate a mirror image. It randomly selects the mirror image data on one side of the mirror image and discards the mirror image data on the other side to determine whether the selected mirror image data is the original image or the mirror image.

[0078] The third model is used to judge the vehicle's direction. The vehicle's direction is divided into forward and backward.

[0079] The fourth model: used to determine the center axis of the vehicle, draw a center axis in the middle of the vehicle;

[0080] The fifth model is used for accessory recognition. It can extract accessories from pictures separately and train them to obtain the corresponding contour rectangles and accessory classifications.

[0081] The method and system for identifying the left and right directions of vehicle accessories have the following advantages:

[0082] (1) Improve recognition accuracy: By optimizing the feature extraction algorithm and introducing the recognition mechanism of symmetry features, the present invention can significantly improve the recognition accuracy of mirror images. When the model faces left-right symmetrical objects, it can better judge the left and right features and reduce the occurrence of misjudgment.

[0083] (2) Enhanced model robustness: This paper introduces a data enhancement strategy to effectively improve the model’s adaptability to complex scenes and mirror images. This makes the model more stable in different environments and can maintain efficient recognition performance in a variety of application scenarios.

[0084] (3) Effective integration of contextual information: The present invention uses an algorithm to more effectively integrate contextual information in mirror images, thereby enhancing the model's ability to understand complex scenes and improving recognition accuracy.

[0085] (4) Targeted data enhancement effect: By adopting an enhancement strategy specifically for mirror images, the model can better learn the characteristics of bilaterally symmetrical objects during training, thereby demonstrating stronger recognition capabilities in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 It is a flow chart of the method for identifying the left and right directions of vehicle accessories. DETAILED DESCRIPTION

[0087] like Figure 1 As shown, the method for identifying the left and right directions of vehicle accessories comprises the following steps:

[0088] S1: training model step, including model training for mirror behavior judgment and mirror image judgment of feature objects, as well as model training for vehicle orientation, central axis, and vehicle accessories recognition;

[0089] The training model step includes model training of the following models:

[0090] S11: Model training for feature object extraction: the feature objects include license plate numbers, store signs, road signs, etc. in vehicle images. These feature objects that are obviously different from mirror images can be used to determine whether mirror images have mirroring behavior by training them separately. Therefore, the feature objects need to be extracted from the images, thereby training a first model for feature object extraction.

[0091] S12: Model training for mirror image judgment: Mirror the feature object separately to generate a mirror image, randomly select the mirror data on one side of the mirror image, discard the mirror data on the other side, and judge whether the selected mirror data is the original image or the mirror image, thereby training a second model for mirror image judgment, which can judge whether the feature object has mirror behavior. Because the feature objects have absolute mirror differences, the accuracy of the second model for mirror image judgment is extremely high.

[0092] S13: vehicle orientation model training: according to the orientation of the vehicle, the orientation direction is divided into front and rear, thereby training a third model for vehicle orientation judgment;

[0093] S14: Model training of the vehicle center axis: used to determine the center position of the vehicle in the picture, draw a center axis at the center position of the vehicle, and thus train a fourth model for determining the vehicle center axis;

[0094] S15: Vehicle accessories recognition model: All individual accessories in the vehicle are marked, and the accessories are trained for recognition, and the accessories are extracted from the picture separately, and the corresponding contour rectangles and accessory classifications are obtained through training, so as to obtain the fifth model for accessory recognition;

[0095] The outline rectangle is the minimum area rectangle obtained by the accessory using the outline. The left and right points of the rectangle can be found by observing and studying the properties of the rectangle and giving suitable x, y judgment formulas according to the current shape.

[0096] S2: Model recognition step, extracting the features of the received image to perform mirror image judgment, after judging the orientation and central axis of the vehicle in the image, identifying the vehicle accessories and judging the final left and right positions of the accessories.

[0097] The model identification step includes the following steps:

[0098] S21: receiving pictures;

[0099] S22: extracting features from the image using the first model to facilitate subsequent steps in determining whether the image has mirroring behavior;

[0100] S23: calling the second model for judging the mirror image to judge whether the extracted feature object is a mirror image. If it is a mirror image, step S26 is applied after step S25. If it is not a mirror image, step S26 is not applied.

[0101] S23: Determine the orientation and center axis coordinates of the vehicle in the image;

[0102] The third model for judging the vehicle's orientation is used to judge whether the vehicle is in the front or rear direction, and the fourth model for judging the vehicle's central axis is used to judge the coordinates of the vehicle's central axis.

[0103] S24: Identify vehicle parts, determine the coordinates of the vehicle parts, and classify the parts;

[0104] Extracting accessory coordinates and accessory classification information of all accessories present in the vehicle according to the fifth model of accessory recognition;

[0105] S25: Determine the final left and right position of the accessory. Specifically, the following steps are included:

[0106] S31: judging whether the accessory has a left-right direction according to different categories of accessories, if there is no left-right direction accessory, directly save the accessory information; if there is a left-right direction accessory, go to step S32;

[0107] Some accessories do not have left and right directions, such as front bumper skin, rear bumper skin, center grille, etc., and the accessory information can be saved directly;

[0108] S32: If there is an accessory in the left and right directions, the left and right directions of the accessory are identified;

[0109] Some accessories have left and right directions, such as left and right headlights, left and right front doors, etc. After setting the origin in the picture, the following judgment methods can be determined based on the coordinates of the accessories and the coordinates of the center axis:

[0110] Accessory coordinates: (x1, y1) represents the upper left corner, (x2, y2) represents the lower right corner, which is a matrix;

[0111] Central axis coordinates: (a1, b1) represents the starting point, (a2, b2) represents the end point, which is a line;

[0112] if

[0113] a1 < (x2-x1) / (y2-y1) * b1;

[0114] and

[0115] a2 < (x2-x1) / (y2-y1) * b2;

[0116] If the fitting is located on the left side of the central axis (PL), then the fitting is located on the right side of the central axis (PR).

[0117] S33: determining the front and rear direction of the accessory according to the vehicle orientation;

[0118] If the vehicle is facing forward and the accessory is PL, the accessory belongs to the right side;

[0119] If the vehicle is facing forward and the accessory is PR, the accessory belongs to the left side;

[0120] If the vehicle is facing backwards and the accessory is PL, the accessory belongs to the left side;

[0121] If the vehicle is facing backwards and the accessory is PR, the accessory belongs to the right side;

[0122] S26: Finally, according to whether the image is a mirror image, the left and right directions of the accessory are reversed to obtain the final direction of the accessory.

[0123] S3: Summarize all extracted accessory information and determined accessory directions, and feed back the results to the user.

[0124] A system for identifying the left and right directions of a vehicle accessory, comprising:

[0125] The first model is used for feature extraction, where the feature objects include license plate numbers, store signs, road signs, etc. in vehicle images;

[0126] The second model is used for mirror image judgment. It performs mirror processing on the feature object separately to generate a mirror image. It randomly selects the mirror image data on one side of the mirror image and discards the mirror image data on the other side to determine whether the selected mirror image data is the original image or the mirror image.

[0127] The third model is used to judge the vehicle's direction. The vehicle's direction is divided into forward and backward.

[0128] The fourth model: used to determine the center axis of the vehicle, draw a center axis in the middle of the vehicle;

[0129] The fifth model is used for accessory recognition. It can extract accessories from pictures separately and train them to obtain the corresponding contour rectangles and accessory classifications.

[0130] The model training for mirror image judgment in the present invention uses the following algorithms:

[0131] (1) Symmetry feature extraction algorithm: The present invention optimizes the left and right features of the mirror image through the symmetry feature extraction algorithm, which can effectively distinguish the features of left-right symmetrical objects and reduce misjudgment.

[0132] (2) Data augmentation strategy: A specific data augmentation method for mirror images is designed to improve the training effect and robustness of the model by generating diverse mirror samples.

[0133] (3) Contextual information integration mechanism: The contextual information integration mechanism is introduced to enable the model to make full use of the information of the surrounding environment when recognizing mirror images, thereby improving the overall recognition accuracy.

[0134] The method and system for identifying the left and right directions of vehicle accessories have the following advantages:

[0135] (1) Improve recognition accuracy: By optimizing the feature extraction algorithm and introducing the recognition mechanism of symmetry features, the present invention can significantly improve the recognition accuracy of mirror images. When the model faces left-right symmetrical objects, it can better judge the left and right features and reduce the occurrence of misjudgment.

[0136] (2) Enhanced model robustness: This paper uses data enhancement strategies to effectively improve the model's adaptability to complex scenes and mirror images. This makes the model more stable in different environments and can maintain efficient recognition performance in a variety of application scenarios.

[0137] (3) Effective integration of contextual information: The present invention uses an algorithm to more effectively integrate contextual information in mirror images, thereby enhancing the model's ability to understand complex scenes and improving recognition accuracy.

[0138] (4) Targeted data enhancement effect: By adopting an enhancement strategy specifically for mirror images, the model can better learn the characteristics of bilaterally symmetrical objects during training, thereby demonstrating stronger recognition capabilities in practical applications.

Claims

1. A method for identifying the left and right directions of a vehicle accessory, characterized in that: The following steps are involved: S1: training model step, including model training for mirror behavior judgment and mirror image judgment of feature objects, as well as model training for vehicle orientation, central axis, and vehicle accessories recognition; S2: Model recognition step, extracting the features of the received image to perform mirror image judgment, after judging the orientation and central axis of the vehicle in the image, identifying the vehicle accessories and judging the final left and right orientation of the accessories; It includes the following steps: S21: receiving pictures; S22: extracting features from the image using the first model, so as to determine whether the image has mirroring behavior in a subsequent step; S23: calling the second model for judging the mirror image to judge whether the extracted feature object is a mirror image; if the image has a mirror image behavior and is a mirror image, then step S26 is applied after step S25; if it is not a mirror image, then step S26 is not applied and the process goes to step S23; S23: determining the orientation and center axis coordinates of the vehicle in the image, determining whether the vehicle is in the front or rear according to the third model for determining the orientation of the vehicle, and determining the center axis coordinates of the vehicle according to the fourth model for determining the center axis of the vehicle; S24: Identify vehicle parts, determine the coordinates of the vehicle parts, and classify the parts; S25: determining the final left and right position of the accessory; Step S26: According to whether the image is a mirror image, the left and right directions of the accessory are reversed to obtain the final direction of the accessory; S3: Summarize all extracted accessory information and determined accessory directions, and feed back the results to the user.

2. The method for identifying the left and right directions of vehicle accessories according to claim 1, characterized in that: In step S1, the model training step includes model training of the following models: S11: Model training for feature extraction: the feature objects are selected from the license plate number, store plaque, and road sign in the vehicle image. By separately training the feature objects that are different in the mirror image, it can be used to determine whether there is mirroring behavior in the mirror image; S12: Model training for judging mirror images: mirror the feature objects separately to generate mirror images, randomly select mirror data on one side of the mirror image, discard the mirror data on the other side, and judge whether the selected mirror data is the original image or the mirror image; S13: Vehicle orientation model training: According to the orientation of the vehicle, the orientation direction is divided into forward and backward; S14: Model training of the vehicle center axis: used to determine the center position of the vehicle in the image and draw a center axis in the center position of the vehicle; S15: Vehicle accessories recognition model: All individual accessories in the vehicle are marked, and the accessories are trained for recognition. The model has the ability to extract accessories from images separately, and trains to obtain the corresponding contour rectangles and accessory classifications.

3. The method for identifying the left and right directions of vehicle accessories according to claim 2, characterized in that: In step S15, the outline rectangle is the minimum area rectangle obtained by the accessory using the outline, and the left point and the right point of the rectangle are found according to the current rectangular shape.

4. The method for identifying the left and right directions of a vehicle accessory according to claim 1, characterized in that: In step S25, the final left and right position of the accessory is determined, including the following steps: S31: judging whether the accessory has a left-right direction according to different categories of accessories, if there is no left-right direction accessory, directly save the accessory information; if there is a left-right direction accessory, go to step S32; S32: If there are accessories in the left and right directions, the accessories are identified in the left and right directions. After the origin is set in the picture, the following judgment method can be determined according to the coordinates of the accessories and the coordinates of the central axis: Accessory coordinates: (x1, y1) represents the upper left corner, (x2, y2) represents the lower right corner, which is a matrix; Central axis coordinates: (a1, b1) represents the starting point, (a2, b2) represents the end point, which is a line; if a1 < (x2-x1) / (y2-y1) * b1; and a2 < (x2-x1) / (y2-y1) * b2; If the fitting is located on the left side of the central axis (PL), then the fitting is located on the right side of the central axis (PR). S33: determining the front and rear direction of the accessory according to the vehicle orientation; If the vehicle is facing forward and the accessory is PL, the accessory belongs to the right side; If the vehicle is facing forward and the accessory is PR, the accessory belongs to the left side; If the vehicle is facing backwards and the accessory is PL, the accessory belongs to the left side; If the vehicle is facing rearward and the accessory is PR, the accessory is on the right side.

5. The method for identifying the left and right directions of a vehicle accessory according to claim 1, characterized in that: In step S31, the accessories that do not exist in the left and right directions are accessories that are used only by the car, including a front bumper cover, a rear bumper cover, and a center grille.

6. A system for identifying the left and right directions of vehicle accessories, characterized in that: include: The first model is used to extract features, such as license plate numbers, store signs, and road signs in vehicle images; The second model is used for mirror image judgment. It performs mirror processing on the feature object separately to generate a mirror image. It randomly selects the mirror image data on one side of the mirror image and discards the mirror image data on the other side to determine whether the selected mirror image data is the original image or the mirror image. The third model is used to judge the vehicle's direction. The vehicle's direction is divided into forward and backward. The fourth model: used to determine the center axis of the vehicle, draw a center axis in the middle of the vehicle; The fifth model is used for accessory recognition. It can extract accessories from pictures separately and train them to obtain the corresponding contour rectangles and accessory classifications.

Images