Rolov7-based inquiry identification auxiliary method and system

Abstract

The invention relates to a method and a system for assisting inquiry identification based on yolov7. The method comprises the following steps: S1, performing data annotation processing on an original image to obtain an anchor frame, namely annotation information; s2, dividing the marked data into pictures and marking information, training by using a yov7 model, and dividing the pictures into training, testing and evaluation data; s3, in the training process of the yolov7 model, a batch value is set, and the training precision is determined through the number of rounds of training; s4, the trained yolov7 model is tested, and whether the actual accuracy rate meets the requirement or not is determined; s5, the yolov7 model meeting the requirement is converted into an onx model; s6, performing reasoning and recognition by using the converted onx model; s7, setting a video frame extraction interval for frame extraction according to a service scene requirement; and S8, carrying out model reasoning on the frame-extracted picture, determining the number of people in the business scene at the moment, and determining whether to carry out early warning processing or not.

Description

Technical Field

[0001] The present invention relates to a yolov7-based inquiry recognition auxiliary method and system. Background Art

[0002] In the digital age, human body recognition technology is becoming an area of ​​great concern. With the continuous advancement of science and technology and the expansion of application scenarios, we need to develop this technology to meet social, economic and security needs.

[0003] For certain scenario modes, such as negotiations, program reviews, product evaluation meetings, interviews, visits, and interrogations, in order to prevent personal subjective factors, generally no less than two people shall participate in the questioning to ensure the fairness and effectiveness of the questioning process and prevent possible subjective factors, incomplete considerations, or other improper behaviors. The present invention believes that it is urgent to use human body recognition technology to judge the compliance of the questioning process through machines. This can not only timely discover misjudgments, misjudgments or non-compliance in the questioning and communication process, but also save related labor costs.

[0004] Due to the movement of people and camera angle issues during the questioning and communication process, there may be a high degree of overlap of people or people may not appear completely in the camera, which makes the model unable to accurately identify the number of people in the questioning and communication process.

[0005] At the same time, the poor generalization of the YOLOv7 model requires a large number of scenes for training, which places high demands on scene collection and equipment.

[0006] Existing techniques such as Yolov4 and Yolov5 may have limitations in speed and accuracy, especially when dealing with large-scale datasets and complex scenarios. These limitations may be caused by insufficiently optimized model structures, high computational complexity, or inaccurate label assignment strategies. Although MediaPipe is fast on CPU, it may not perform as well as Yolov7 on GPU. Summary of the invention

[0007] The purpose of the present invention is to use predefined anchor frames based on yolov7 to assist training to identify human body conditions in images, thereby providing a yolov7-based inquiry recognition auxiliary method and system.

[0008] To solve the above problems, the technical solution adopted by the present invention is: A yolov7-based inquiry recognition auxiliary method comprises the following steps: S1, perform data annotation processing on the original image to obtain the anchor frame, i.e., annotation information; S2, divide the labeled data into images and annotation information, use yolov7 model training, and divide it into training, testing, and evaluation data; S3, during the training process of the yolov7 model, set the batch value and determine the training accuracy by the number of training rounds; S4, test the trained yolov7 model and check whether the actual accuracy meets the requirements; S5, converting the yolov7 model that meets the requirements into an onnx model; S6, use the converted onnx model for recognition and reasoning; S7, set the video frame extraction interval to extract frames according to the business scenario requirements; S8, perform onnx model inference on the extracted frame image to confirm the number of people in the business scenario at this time and determine whether to perform early warning processing.

[0009] Further, in step S1, the following steps are included: S1.1, image preprocessing, preliminary screening; First, image data is collected in the set business scenarios. All collected images are confidential data, and all image data processing is performed without an external network environment. Then, after the image data is collected, preliminary screening is performed to remove image data with a similarity greater than a set threshold or a fuzziness greater than a set threshold. The training image data needs to include all business scenarios, and each scenario includes different types of image data to meet the set quantity. S1.2, after the preliminary screening, the image data is manually reviewed for a second time to confirm the factors that affect the accuracy of person recognition in the image data, and the factors that affect the accuracy of person recognition are added to the annotation process and trained; Factors that affect the accuracy of person recognition include: objects whose coincidence with the person in the image data is greater than a set threshold, objects whose color similarity with the person's clothing in the image data is greater than a set threshold, and objects that reflect the person in the image data; S1.3, confirm the labels of the annotation process, set the annotation rules, and annotate the image data according to the annotation rules; when annotating, first, determine the label content required for this annotation according to the set requirements; then, annotate according to the annotation rules; secondly, train the annotated data; thirdly, determine whether the annotation content needs to be adjusted. If so, return to the beginning of the annotation and re-annotate; otherwise, export the annotated image and annotation file.

[0010] Further, in step S2, the labeled data is classified, and the image data is divided into a training set, a test set, and a validation set according to a set ratio; The training set is used for model training, the test set is used to test the overall accuracy of the model from the training set, and the validation set is used to verify the model; Specific steps are as follows: S2.1, input the annotated image and annotation file in step S1.3; S2.2, divide the annotated images and annotated files into a training set, a test set, and a validation set, which are used for training, testing, and evaluation respectively; S2.3, according to the training requirements, set the yaml file in the model and confirm the data path and training parameter settings; S2.4, perform model training; S2.5, first, evaluate the trained model; then, determine whether the model meets the accuracy requirements; if it does, export the model; if it does not, execute S2.6; S2.6, first, test the model and determine the need for improvement based on the tested images; then, determine whether to directly adjust the parameters for direct training; if so, return to S2.4; otherwise, execute S2.7; S2.7, determine whether the annotation needs to be adjusted; if so, modify the annotation content of the image according to the adjustment; otherwise, increase the data volume of the image and annotate it.

[0011] Further, in step S7, the camera is connected, a set frame image is extracted within a set time, and frame extraction processing is performed using a frame extraction technology. The extracted static image is saved in a fixed position to become the original image of this process and is not subjected to secondary processing. All processed images are copies of the original image. Specifically, the following steps are performed; S7.1, create environment; S7.2, convert the training model into onnx format; S7.3, configure the model startup code according to the onnx model; S7.4, determine whether the model is started normally; if so, the model is deployed successfully; otherwise, execute S7.5; S7.5, determine whether the environment creation does not meet the requirements; if so, return to step S7.1; otherwise, execute S7.6; S7.6, determine whether the model conversion is incorrect; if so, return to step S7.2; otherwise, return to step S7.3.

[0012] Further, in step S8, the scene is set in a set room; a close-up camera and a panoramic camera are equipped in the scene; During questioning and communication, first, the close-up camera is identified and processed. The close-up camera is used to shoot the person being questioned during the questioning process; secondly, when the close-up camera recognizes a person in the scene, it is determined that the room has entered the questioning state. At this time, the panoramic camera pictures extracted from the same time period are identified. If the number of people appearing in the panoramic camera at this time is greater than the set number N, then the questioning and conversation process is judged to be compliant. If the number of people identified is less than the set number N, then the questioning and conversation process is judged to be in violation of the regulations and is defined as a failure.

[0013] Furthermore, in step S8, when a person enters or leaves the room in the scene, S frame extraction judgments are performed within the specified time according to the actual inquiry and conversation requirements; when the number of violations reaches M times, an early warning process is performed, and it is clear that a violation has occurred in this inquiry and conversation.

[0014] Further, in step S8, the following steps are specifically performed: S8.1, first, read the image and perform detection; then, determine whether the detection result shows that a set number of people exist; S8.2, if not, output the result; otherwise, execute step S8.3; S8.3, determine whether it is higher than the threshold, if yes, output the result; otherwise, execute step S8.4; S8.4, delete the data and output the results.

[0015] Furthermore, when training the YOLOv7 model in step 2, first, based on the E-ELAN network, planned re-parameterized convolution and label assignment method, according to the lead guided assigner method, the lead head is used to control the label assignment of lead and aux, the aux head learns the coarse label, and the lead head learns the finelabel method; Then, yolov7 uses a box allocation method, each real label is assigned to a grid of no more than 3, and the iou value is calculated for all the real boxes after allocation and each predicted box; secondly, the first m = min (10, num (predict)) iou values ​​are screened; secondly, the dynamic k value is calculated based on the first m iou values, that is, Dynamic_k = ; Since smiOTA considers the classification score and regression score, the cost of smiOTA is cost = Loss cls + 3Losss iou ; Among them, Loss cls is the classification loss, Losss iou is the regression loss; Get the index of the Dynamic_k predicted labels with the lowest cost according to the cost; index refers to the predicted label that matches the true label; Finally, when the predicted value matches multiple true values, the one with the lowest cost is taken as the match and the others are removed.

[0016] A yolov7-based inquiry recognition auxiliary system for executing the above method; the system includes an independent room, in which a close-up camera and a panoramic camera are arranged; The close-up camera and the panoramic camera are electrically connected to the processor.

[0017] Compared with the prior art, the present invention is beneficial in that the model provided by the present invention can more accurately identify the human body, and can better distinguish objects that are close to the human body, and can better identify two human bodies with an overlap of less than 70%.

[0018] The specific advantages of the present invention are: a. It improves the detection speed and accuracy in real-time applications; b. The optimized model structure reduces the computational complexity and improves the generalization ability of the model; c. More image processing information is provided through the training of the auxiliary head, which is helpful to learn useful feature representations; d. Composite model scaling enables the model to adapt to different application requirements; e. The trainable BoF method improves model performance without increasing the training cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a data annotation flow chart implemented by the present invention.

[0020] Figure 2 It is a flow chart of model training implemented by the present invention.

[0021] Figure 3 It is a model deployment flow chart implemented by the present invention.

[0022] Figure 4 It is a flow chart of image recognition implemented by the present invention.

[0023] Figure 5 It is a business process diagram implemented by the present invention. DETAILED DESCRIPTION

[0024] The technical solution adopted by the present invention is: S1, perform data annotation processing on the original image to obtain the anchor frame, i.e., annotation information; S2, divide the labeled data into two contents: pictures and annotation information, use yolov7 model training, and divide it into training, testing, and evaluation data; S3, during the training process, you can set the batch value and determine the training accuracy by the number of training rounds; S4, test the trained model to see if the actual accuracy meets the requirements; S5, converting the model that meets the requirements into an onnx model; S6, use the converted onnx model for recognition and reasoning; S7, set the video frame extraction interval to extract frames according to the business scenario requirements; S8, perform model inference on the extracted frame image, confirm the number of people in the business scene at this time, and determine whether to perform early warning processing.

[0025] For image annotation, in step S1, there is also a preceding step S1.1, image preprocessing, for preliminary screening; First, because the business scenarios used in this invention are relatively special, it is necessary to collect images in the set business scenarios. All collected images are confidential data and cannot be disseminated. All data processing is performed without an external network environment. Then, after data collection, it is necessary to perform preliminary screening and processing on images with high similarity and fuzziness to avoid affecting the accuracy of the overall training model. In addition, to ensure the generalization of the training model, the training data needs to include all business scenarios, and each scenario needs to have different types of data to reach about 100-200 optimally.

[0026] The core steps of image annotation are as follows: S1.2, after preliminary screening, manually review the images to confirm the factors that affect the accuracy of person recognition in the images, add these factors to the annotation process at the same time, and train them to distinguish them from people. Factors that affect the accuracy of person recognition include: objects whose coincidence with the human body in the image data is greater than the set threshold, such as chairs, sofas, computers, mice, keyboard lights, objects whose color similarity with the human body in the image data is greater than the set threshold, such as clothes, hangers, etc., and objects that reflect the human body in the image data, such as televisions, mirrors, etc.; S1.3, according to the above situation, confirm the labels of the annotation process, and set the annotation rules to prevent different annotation situations from occurring during the annotation process, which will affect the recognition accuracy. According to the annotation rules, the relevant image data should be annotated. For the amount of data, a total of about 5,000 data pieces are required for different scenes. Of course, the more data there is, the higher the accuracy will be. When annotating, first, determine the label content required for this annotation according to the set requirements; then, annotate according to the annotation rules; secondly, train the annotated data; thirdly, determine whether the annotation content needs to be adjusted. If necessary, return to the beginning of the annotation and re-annotate; otherwise, export the annotated images and annotation files. In computer science, labels is a general concept used to describe the meaning of a piece of code or data.

[0027] For model training, in step S2, the labeled data is classified. In the present invention, the image data is divided into a training set, a test set, and a validation set according to a ratio of 8:1:1.

[0028] The training set is used for model training, the test set is used to test the overall accuracy of the model from the training set, and the validation set is used to verify the model. Execute the following steps: S2.1, input the annotated image and annotation file in step S1.3; S2.2, divide the annotated images and annotated files into a training set, a test set, and a validation set, which are used for training, testing, and evaluation respectively; S2.3, according to the training requirements, set the yaml file in the model and confirm the data path and training parameter settings; the yaml file is a highly readable format used to express data serialization S2.4, perform model training; S2.5, first, evaluate the trained model; then, determine whether the model meets the accuracy requirements; if it does, export the model; if it does not, execute S2.6; S2.6, first, test the model and determine the need for improvement based on the tested images; then, determine whether to directly adjust the parameters for direct training; if so, return to S2.4; otherwise, execute S2.7; S2.7, determine whether the annotation needs to be adjusted; if so, modify the annotation content of the image according to the adjustment; otherwise, increase the data volume of the image and annotate it; For video stream frame extraction, in step S7, the camera is connected and 5 frames of pictures are extracted in 1 minute. The present invention uses Java to perform frame extraction technology processing. The extracted static pictures will be saved in a fixed position to become the original images of this process. The original images will not be processed twice during the entire research and development process. All processed pictures are copies of the original pictures to prevent the original pictures from being modified and affecting the subsequent processes. Specifically perform the following steps; S7.1, create environment; S7.2, convert the training model pt format into onnx format; S7.3, configure the model startup code according to the onnx model; S7.4, determine whether the model is started normally; if so, the model is deployed successfully; otherwise, execute S7.5; S7.5, determine whether the environment creation does not meet the requirements; if so, return to step S7.1; otherwise, execute S7.6; S7.6, determine whether the model conversion is incorrect; if so, return to step S7.2; otherwise, return to step S7.3; With regard to early warning processing, in step S8, the main auxiliary scenarios of the present invention are interview processes, audit processes, negotiation processes, or multi-person questioning processes of a single person, equipped with close-up cameras and panoramic cameras; during questioning and communication, the two cameras need to be processed. First, the close-up camera is identified and processed, and the close-up camera is used to shoot the person being questioned during the questioning process; secondly, when the close-up camera identifies a person in the scene, it is determined that the room has entered the questioning state. At this time, the panoramic camera pictures extracted from the same time period are identified. If the number of people appearing in the panoramic camera at this time is greater than the set number N, for example N=3, then the questioning and conversation process is judged to be compliant. If the number of people identified is less than the set number N, then the questioning and conversation process is judged to be in violation of regulations and is defined as a failure.

[0029] However, in view of the fact that some special situations may occur during the entire inquiry and conversation process, such as personnel needing to temporarily go out to sign or go to the toilet, the present invention has carried out further processing. Then, according to the actual inquiry and conversation requirements, S frame extraction judgments are carried out within the specified time, for example, S=5; when the number of violations reaches M times, early warning processing is carried out, and it is clear that a violation has occurred in this inquiry and conversation.

[0030] Specific steps are as follows: S8.1, first, read the image and perform detection; then, determine whether there is a person in the detection result; S8.2, if not, output the result; otherwise, execute step S8.3; S8.3, determine whether it is higher than the threshold, if yes, output the result; otherwise, execute step S8.4; S8.4, delete the data and output the results.

[0031] In step 2, when training the yolov7 model, first, based on the E-ELAN network, planned re-parameterized convolution and label assignment method; yolov7 proposed a lead guided assigner method, which controls the label assignment of lead and aux through the lead head, the aux head learns the coarse label, and the lead head learns the fine label; In the past, auxiliary loss heads were generally decoupled to assign labels separately, but it was difficult to align the loss of the aux head with that of the lead head. Therefore, a lead head guided label assigner was proposed. On this basis, a method was proposed in which the aux head learns coarse labels and the lead head learns fine labels, so that the aux head learns more labels and has better recall, while the lead head learns labels and focuses more on precision.

[0032] Then, yolov7 uses a box allocation method, and each true label is assigned to a grid of no more than 3. The iou value is calculated for all the real boxes after allocation and each predicted box; secondly, the first m = min (10, num (predict)) iou values ​​are screened. Secondly, the dynamic k value is calculated based on the first m iou values, that is, Dynamic_k = ; Since smiOTA considers the classification score and regression score, the cost of smiOTA is cost = Loss cls + 3Losss iou ; Among them, Loss cls is the classification loss, Losss iou is the regression loss; Get the index of the Dynamic_k predicted labels with the lowest cost according to the cost. The index here refers to the predicted label that matches the true label.

[0033] Finally, we need to deal with the situation where the predicted value matches multiple true values. Here, the one with the lowest cost is taken as the match, and the others are removed.

[0034] Notes on technical terms: Yolov7: The seventh version of the YOLO algorithm, which improves on the original one and increases detection speed and accuracy.

[0035] Deep Learning: A subfield of machine learning that is based on learning with artificial neural networks, specifically layers with multiple nonlinear transformations (i.e. deep neural networks).

[0036] Person Identification: Identifying individuals in images or videos through computer vision technology, usually involving technologies such as facial recognition and gait analysis.

[0037] Data Preprocessing: A series of processing steps performed on data before model training, including normalization, scaling, enhancement, etc., to improve model performance.

[0038] Image Scaling: Change the size of the image to adapt it to the requirements of the model input.

[0039] Normalization: Scale the data so that it falls into a small specific interval, usually [0, 1] or [-1, 1], to speed up training and improve model stability.

[0040] Data Augmentation: Artificially increase the size and diversity of the data set through rotation, scaling, cropping, etc. to improve the generalization ability of the model.

[0041] Annotation: Labeling images in a dataset to indicate the target category and location contained in the image is a necessary step in training models in supervised learning.

[0042] Model Training: Use labeled datasets to train deep learning models so that they can learn patterns and features in the data.

[0043] Bounding Box: A rectangular box used to represent the location of a target in an image, usually defined by coordinates (x, y) (upper left corner) and width and height.

[0044] Class Probability: The confidence that the model predicts that the target belongs to a certain category, usually output by the softmax function.

[0045] Generalization Ability: The ability of a model to make predictions about unseen data is an important indicator for measuring model performance.

[0046] Graphical User Interface (GUI): An interface that allows users to interact with electronic devices, usually including elements such as windows, icons, and buttons.

[0047] Data Interface: A mechanism that allows data to be exchanged between different software or systems, which can be an API, database connection, etc.

[0048] Frame Extraction: The process of extracting image frames at a specific time point from a video stream.

[0049] The present invention is fully described for a clearer disclosure, and the prior art is not listed one by one.

[0050] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some of the technical features therein with equivalents; it is obvious for those skilled in the art to combine multiple technical solutions of the present invention. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention. The technical contents not described in detail in the present invention are all known technologies.

Claims

1. A yolov7-based inquiry recognition auxiliary method, characterized in that: The steps include: S1, perform data annotation processing on the original image to obtain the anchor frame, i.e., annotation information; S2, divide the labeled data into images and annotation information, use yolov7 model training, and divide it into training, testing, and evaluation data; S3, during the training process of the yolov7 model, set the batch value and determine the training accuracy by the number of training rounds; S4, test the trained yolov7 model and check whether the actual accuracy meets the requirements; S5, converting the yolov7 model that meets the requirements into an onnx model; S6, use the converted onnx model for reasoning and recognition; S7, set the video frame extraction interval to extract frames according to the business scenario requirements; S8, perform onnx model inference on the extracted frame image to confirm the number of people in the business scenario at this time and determine whether to perform early warning processing.

2. The inquiry identification auxiliary method based on yolov7 according to claim 1 is characterized in that: In step S1, the following steps are included: S1.1, image preprocessing, preliminary screening; First, image data is collected in the set business scenario. All collected images are confidential data, and all image data processing is performed without an external network environment. Then, after the image data is collected, the image data with similarity greater than a set threshold or blur greater than a set threshold is initially screened and removed; the training image data needs to include all business scenarios, and each scenario includes different types of image data to meet the set number; S1.2, after the preliminary screening, the image data is manually reviewed for a second time to confirm the factors that affect the accuracy of person recognition in the image data, and the factors that affect the accuracy of person recognition are added to the annotation process and trained; Factors that affect the accuracy of person recognition include: objects whose coincidence with the person in the image data is greater than a set threshold, objects whose color similarity with the person's clothing in the image data is greater than a set threshold, and objects that reflect the person in the image data; S1.3, confirm the labels of the annotation process, set the annotation rules, and annotate the image data according to the annotation rules; when annotating, first, determine the label content required for this annotation according to the set requirements; then, annotate according to the annotation rules; secondly, train the annotated data; thirdly, determine whether the annotation content needs to be adjusted. If so, return to the beginning of the annotation and re-annotate; otherwise, export the annotated image and annotation file.

3. The inquiry identification auxiliary method based on yolov7 according to claim 2 is characterized in that: In step S2, the labeled data is classified into a training set, a test set, and a validation set according to a set ratio; The training set is used for model training, the test set is used to test the overall accuracy of the model from the training set, and the validation set is used to verify the model; Specific steps are as follows: S2.1, input the annotated image and annotation file in step S1.3; S2.2, divide the annotated images and annotated files into a training set, a test set, and a validation set, which are used for training, testing, and evaluation respectively; S2.3, according to the training requirements, set the yaml file in the model and confirm the data path and training parameter settings; S2.4, perform model training; S2.5, first, evaluate the trained model; then, determine whether the model meets the accuracy requirements; if it does, export the model; if it does not, execute S2.6; S2.6, first, test the model and determine the need for improvement based on the tested images; then, determine whether to directly adjust the parameters for direct training; if so, return to S2.4; otherwise, execute S2.7; S2.7, determine whether the annotation needs to be adjusted; if so, modify the annotation content of the image according to the adjustment; otherwise, increase the data volume of the image and annotate it.

4. The inquiry identification auxiliary method based on yolov7 according to claim 3 is characterized in that: In step S7, the camera is connected, a set frame image is extracted within a set time, and frame extraction processing is performed using frame extraction technology. The extracted static image is saved in a fixed position to become the original image of this process and is not subjected to secondary processing. All processed images are copies of the original image. Specifically, the following steps are performed; S7.1, create environment; S7.2, convert the training model into onnx format; S7.3, configure the model startup code according to the onnx model; S7.4, judging whether the model is started normally; If yes, the model is deployed successfully; Otherwise, execute S7.5; S7.5, determine whether the environment creation does not meet the requirements; if so, return to step S7.1; Otherwise, execute S7.6; S7.6, determine whether the model conversion is incorrect; if so, return to step S7.2; otherwise, return to step S7.

3.

5. The inquiry identification auxiliary method based on yolov7 according to claim 4 is characterized in that: In step S8, the scene is set in a set room; a close-up camera and a panoramic camera are equipped in the scene; During questioning and communication, first, the close-up camera is identified and processed. The close-up camera is used to shoot the person being questioned during the questioning process; secondly, when the close-up camera recognizes a person in the scene, it is determined that the room has entered the questioning state. At this time, the panoramic camera pictures extracted from the same time period are identified. If the number of people appearing in the panoramic camera at this time is greater than the set number N, then the questioning and conversation process is judged to be compliant. If the number of people identified is less than the set number N, then the questioning and conversation process is judged to be in violation of the regulations and is defined as a failure.

6. The inquiry identification auxiliary method based on yolov7 according to claim 5 is characterized in that: In step S8, when a person enters or leaves the room in the scene, S frame extraction judgments are performed within the specified time according to the actual inquiry and conversation requirements; when the number of violations reaches M times, an early warning process is performed, and it is clear that a violation has occurred in this inquiry and conversation.

7. The inquiry identification auxiliary method based on yolov7 according to claim 1 is characterized in that: In step S8, the following steps are specifically performed: S8.1, first, read the image and perform detection; then, determine whether the detection result shows that a set number of people exist; S8.2, if not, output the result; Otherwise, execute step S8.3; S8.3, determine whether it is higher than a threshold, and if so, output the result; Otherwise, execute step S8.4; S8.4, delete the data and output the results.

8. The inquiry identification auxiliary method based on yolov7 according to claim 7 is characterized in that: When training the YOLOv7 model in step 2, first, based on the E-ELAN network, planned re-parameterized convolution and label assignment method, according to the lead guided assigner method, the lead head is used to control the label assignment of lead and aux, the aux head learns the coarse label, and the lead head learns the fine label. Then, yolov7 uses a box allocation method, each real label is assigned to a grid of no more than 3, and the iou value is calculated for all the real boxes after allocation and each predicted box; secondly, the first m = min (10, num (predict)) iou values ​​are screened; secondly, the dynamic k value is calculated based on the first m iou values, that is, Dynamic _ k = ; Since smiOTA considers the classification score and regression score, the cost of smiOTA is cost = Loss cls + 3Losss iou ; Among them, Loss cls is the classification loss, Losss iou is the regression loss; Get the index of the Dynamic_k predicted labels with the lowest cost according to the cost; index refers to the predicted label that matches the true label; Finally, when the predicted value matches multiple true values, the one with the lowest cost is taken as the match and the others are removed.

9. A yolov7-based inquiry and identification auxiliary system, characterized in that: Used to perform the method described in any one of claims 1 to 8; the system includes an independent room, in which a close-up camera and a panoramic camera are arranged; The close-up camera and the panoramic camera are electrically connected to the processor.

Images