Security check whole-line intelligent identification system and method

Abstract

The invention discloses a security check whole-line intelligent identification system and a security check whole-line intelligent identification method, which are used for providing intelligent security check service for the whole line from the front end to the rear end of a security check process, providing auxiliary graph discrimination prompts for a boot operator and a box opening operator in a first-line security check process, and pushing security check images to a second-line review workstation, so that review is more efficient and accurate. Meanwhile, the performance updating of the intelligent identification model is embedded into the whole security check process, the sample data required by optimization is obtained by comparing and screening the manual conclusion and the intelligent identification result, and the model is upgraded in an incremental training mode, so that the model optimization can be directly carried out on the intelligent identification server on site; therefore, the efficiency of new sample collection and model training is greatly improved, and the data confidentiality requirement can be met.

Description

Technical Field

[0001] The present invention relates to the technical field of security inspection equipment, and in particular to a full-line intelligent identification system and method for security inspection. Background Art

[0002] With the rapid increase in passenger and cargo throughput of civil aviation (hereinafter referred to as civil aviation), security inspection faces the dual challenges of safety and efficiency. The application of intelligent recognition and image judgment system for security inspection is an urgent need for the construction of intelligent informatization of civil aviation.

[0003] Chinese patent application CN118278785A discloses a security inspection work quality management system, method, device, electronic device and computer storage medium. Among them, the security inspection work quality management system includes: a quality control terminal and a quality control server, the quality control terminal is connected to the quality control server in communication; the quality control terminal is used to collect the working status data of the security inspectors and the detection data of the security inspection equipment, and send the working status data and the detection data to the quality control server; the quality control server is used to process the working status data and the detection data according to the data collected by the quality control terminal to obtain the security inspection work quality data. According to the embodiment, the security inspection work quality control indicators can be fully monitored and warned, so as to effectively avoid security inspection errors caused by the working status of the security inspectors and other reasons. However, this solution can only serve the quality control system, and cannot provide auxiliary image judgment prompts to the front-line startup and unpacking personnel. The quality control system can only perform security inspection work quality control monitoring and warning, and does not further utilize the data that may be collected.

[0004] In addition, in the prior art, whenever it is necessary to increase the recognition category of prohibited items and improve the performance of certain categories of prohibited items, the recognition model needs to be retrained together with the data for training the initial model using the newly added sample data collected on site, and then the retrained recognition model needs to be updated to the site before it can be used on site. This process has a long implementation cycle and low efficiency. If the newly added samples are transmitted to the developer of the initial model, the user's need for data confidentiality cannot be met; if the data for training the initial model is stored on site, it will not only increase the storage pressure on the running equipment, but also leak the data and methods used by the developer. There are as many as 100,000 security inspection images on site per day. How to quickly and effectively extract security inspection image data that can improve or supplement the performance of the recognition model as new sample data is an urgent problem to be solved.

[0005] Chinese patent application CN118570587A discloses an online training method and device for a radiation image recognition model, and a radiation image recognition method and device. Preset data is pre-stored in the radiation inspection system, and the preset data includes a first radiation image. The training method includes: collecting a second radiation image at the radiation inspection site; annotating the second radiation image online to form field-collected data with second annotation information, wherein the second annotation information is incomplete annotation information of the second radiation image, and the second annotation information is used to indicate a part of the object of interest in the second radiation image; online synthesis of preset data and field-collected data to form a training data set; and online training of the radiation image recognition model using the training data set. However, this solution requires the use of preset data as a background image for synthesizing training data, and the preset data is also used in the online training stage, resulting in slow model training and increased storage space for storing preset data. Summary of the invention

[0006] In view of the deficiencies in the prior art, the present invention aims to provide a full-line intelligent identification system and method for security inspection.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] A security inspection full-line intelligent identification method comprises the following steps:

[0009] S101, using the security inspection data collection device to collect the security inspection image data of the current passenger's luggage, and send it to the intelligent recognition server;

[0010] S102, the intelligent recognition server uses the recognition model to analyze the received image data and returns the intelligent recognition result to the security inspection data collection device, the intelligent recognition result including whether there are suspected contraband or suspected contraband and the category, quantity and location information of the suspected contraband;

[0011] S103, the security inspection data collection device sends the security inspection image data and intelligent recognition results to the manual start-up site; the manual start-up site first displays the security inspection image data on the screen, and a professional operator interprets the security inspection image data and gives the operator's manual conclusion, which includes whether there are suspected contraband or suspected contraband and the type, quantity and location of the suspected contraband;

[0012] S104. When at least one of the operator's manual conclusion and the intelligent recognition result determines that there are suspected contraband in the current security inspection image data, the manual operation site will send the current security inspection image data, the operator's manual conclusion and the intelligent recognition result to the manual unpacking site, which will display them on its screen. At the same time, the luggage corresponding to the current security inspection image data will be sorted to the unpacking workbench corresponding to the manual unpacking site, waiting for the unpacking operator to unpack and inspect. The unpacking operator obtains the manual unpacking conclusion of the corresponding luggage based on the unpacking inspection result, and the manual unpacking conclusion includes whether the luggage is safe or whether there are contraband in the luggage and the type, quantity and location information of the contraband;

[0013] S105. The unpacker enters the manual unpacking conclusion obtained in step S104 at the manual unpacking station. For luggage judged to be safe, the unpacker's manual conclusion and intelligent recognition result are deleted. For luggage judged to contain contraband, the unpacker's manual conclusion or intelligent recognition result is confirmed or modified.

[0014] Furthermore, in step S102, when any one of the delayed prompt conditions is triggered, the manual start-up site will delay displaying the intelligent recognition result on the screen to provide auxiliary prompts to the operator; the delayed prompt conditions include: (1) the current security inspection image data is about to leave the screen of the manual start-up site; (2) the current security inspection image data starts timing from entering the screen of the manual start-up site, and after a predefined period of time, the security inspector has not given the operator's manual conclusion; (3) the security inspector starts the transmission device of the security inspection data acquisition equipment again after changing shifts.

[0015] Furthermore, as one of the implementation schemes, the above method also includes step S106: the manual startup site further performs a difference calculation on the manual conclusion of the operator and the intelligent recognition result. When the calculation result shows that the manual conclusion of the operator and the intelligent recognition result are inconsistent, that is, only one of the manual conclusion of the operator and the intelligent recognition result determines that there are suspected contraband in the current security inspection image data, or the manual conclusion of the operator and the intelligent recognition result both determine that there are suspected contraband in the current security inspection image data but there are differences in the judgment on the quantity, category and / or location of the suspected contraband, the manual startup site further pushes the current security inspection image data to the manual review site, which is reviewed by senior security personnel and a manual review conclusion is given; if the manual review conclusion determines that the intelligent recognition result is an underreporting or false alarm, the manual review site pushes the current security inspection image data to the manual labeling workstation, which is collected as new sample data after manual labeling for retraining the recognition model. As another implementation scheme, another implementation process of step S106 is: the manual start-up site pushes all security inspection image data to the manual review site for review, and gives a manual review conclusion; if the manual review conclusion determines that the intelligent recognition result is underreported or falsely reported, the manual review site pushes the current security inspection image data to the manual labeling workstation, and after manual labeling, it is collected as new sample data for retraining the recognition model; if the manual review conclusion is that both the operator's manual conclusion and the intelligent recognition result are underreported, the corresponding luggage will be immediately retrieved for unpacking inspection, and the final security inspection conclusion will be given.

[0016] Furthermore, the above method also includes step S107: if the manual unpacking site finds that the intelligent recognition result is underreported or falsely reported based on the manual unpacking conclusion, the corresponding security inspection image data is transmitted to the manual labeling site, and after manual labeling, it is collected as new sample data for retraining the recognition model.

[0017] Furthermore, the above method further includes step S108:

[0018] The recognition model obtained by the most recent update is used as the initial model, and the initial model is used as the teacher model. The parameters of the teacher model are used to initialize the student model whose network structure is exactly the same as the teacher model, and the student model is trained with new sample data; the distillation loss term L based on the knowledge distillation technology is added to the loss function used when training the original recognition model deployed for the first time. distill , which constitutes the loss function of student model training, L distill The calculation formula is as follows:

[0019]

[0020] cls new and cls oldRepresents the category probability of the new model and the old model on the candidate area, box new and box old Represent the bounding box distribution of the new model and the old model on the candidate area, respectively, and N is the number of data points used to calculate the average value:

[0021] The final loss function expression of the student model training is:

[0022] L=L original +λL distill

[0023] Among them, L original is the loss function used when training the original recognition model that was first deployed.

[0024] The present invention also provides a system for realizing the above-mentioned full-line intelligent identification system for security inspection, including an intelligent identification server and a plurality of manual start-up sites, manual unpacking sites, manual review site one and manual review site two, wherein the manual start-up sites, manual unpacking sites, manual review site one and manual review site two are all communicatively connected with the intelligent identification server; the manual review site one and manual review site two are respectively used for comparative review and full review.

[0025] The beneficial effects of the present invention are as follows: the present invention provides intelligent security inspection services for the entire security inspection process from the front end to the back end, provides auxiliary image judgment prompts to the operator and the unpacker in the first-line security inspection process, and pushes the security inspection images to the second-line review workstation to help the review be more efficient and accurate. At the same time, the present invention embeds the performance update of the intelligent recognition model into the entire security inspection process, uses the comparison and screening of manual conclusions and intelligent recognition results to obtain the sample data required for optimization, and uses incremental training to upgrade the model, so that the model can be directly optimized on the on-site intelligent recognition server, thereby greatly improving the efficiency of new sample collection and model training, and meeting data confidentiality requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is an overall flow chart of the method of Example 1 of the present invention;

[0027] Figure 2 This is a flowchart of model retraining in the method of Example 1 of the present invention;

[0028] Figure 3 Schematic diagram of the system architecture in Example 2 of the present invention. DETAILED DESCRIPTION

[0029] The present invention will be further described below in conjunction with the accompanying drawings. It should be noted that this embodiment is based on the technical solution and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to this embodiment.

[0030] Example 1

[0031] This embodiment provides a security inspection full-line intelligent identification method, such as Figure 1 As shown, the following steps are included:

[0032] S101, collecting security inspection data. Specifically, using security inspection data collection equipment such as X-ray security inspection equipment or CT security inspection equipment to collect security inspection image data of the current passenger's luggage, and send it to the intelligent recognition server;

[0033] S102, the intelligent recognition server uses the recognition model to analyze the received image data and returns the intelligent recognition result (i.e. Figure 1 The intelligent recognition results include whether there are no suspected contraband or there are suspected contraband, as well as the category, quantity and location information of the suspected contraband.

[0034] S103, the security inspection data collection equipment sends the security inspection image data and intelligent recognition results to the manual start-up site; the manual start-up site first displays the security inspection image data on the screen, and a professional operator interprets the security inspection image data and gives the operator a manual conclusion. The operator's manual conclusion includes the absence of suspected contraband or the presence of suspected contraband and the category, quantity and location of the suspected contraband. The manual start-up site will display the intelligent recognition results on the screen with a delay to provide auxiliary prompts to the operator. The intelligent recognition results are given in a delayed manner, the purpose of which is not to interfere with the operator's interpretation of the image before he gives a conclusion.

[0035] In this embodiment, when any of the delayed prompt conditions is triggered, the manual start-up site will display the intelligent recognition result on the screen with a delay. The delayed prompt conditions include: (1) the current security inspection image data is about to leave the screen of the manual start-up site; (2) the current security inspection image data starts timing from entering the screen of the manual start-up site, and after a predefined period of time, the security inspector has not given a manual conclusion; (3) the security inspector starts the transmission device of the security inspection data collection equipment again after changing shifts.

[0036] S104. When at least one of the operator's manual conclusion and the intelligent recognition result determines that there are suspected contraband in the current security inspection image data, the manual operation site will send the current security inspection image data, the operator's manual conclusion and the intelligent recognition result to the manual unpacking site, which will display them on its screen. At the same time, the luggage corresponding to the current security inspection image data will be sorted to the unpacking workbench corresponding to the manual unpacking site, waiting for the unpacking operator to unpack and inspect. The unpacking operator obtains the manual unpacking conclusion of the corresponding luggage based on the unpacking inspection result, and the manual unpacking conclusion includes whether the luggage is safe or whether there are contraband in the luggage and the type, quantity and location information of the contraband;

[0037] S105. The unpacker enters the manual unpacking conclusion obtained in step S104 at the manual unpacking station. For luggage judged to be safe, the unpacker's manual conclusion and intelligent recognition result are deleted. For luggage judged to contain contraband, the unpacker's manual conclusion or intelligent recognition result is confirmed or modified.

[0038] The above is the first-line security inspection process. The above steps S101-S105 provide a humanized auxiliary image judgment service for baggage security inspection, which can effectively reduce missed inspections caused by a large number of bags and subjective problems of security inspectors.

[0039] Generally, civil aviation security inspection has strict baggage passing rate requirements to prevent passengers and baggage congestion. Therefore, the time for operators and unpackers to judge images or check baggage is very limited. In order to improve the safety margin of security inspection, large airports have set up review posts and equipped them with senior professional security inspectors to review security inspection image data in real time or quasi-real time in the background.

[0040] As an implementation mode, the method of this embodiment further includes step S106, such as Figure 1As shown: the manual start-up site further calculates the difference between the manual conclusion of the operator and the intelligent recognition result. When the calculation result shows that the manual conclusion of the operator and the intelligent recognition result are inconsistent, that is, only one of the manual conclusion of the operator and the intelligent recognition result determines that there are suspected contraband in the current security inspection image data, or the manual conclusion of the operator and the intelligent recognition result both determine that there are suspected contraband in the current security inspection image data, but there are differences in the judgment of the number, category and / or location of the suspected contraband, the manual start-up site further pushes the current security inspection image data to the manual review site, which is reviewed by senior security inspectors and gives a manual review conclusion. This method is called comparative review. If the manual review conclusion determines that the intelligent recognition result is an underreporting (no suspected contraband is identified) or a false alarm (the suspected contraband is identified but the quantity, category and / or location are incorrectly identified), the manual review site pushes the current security inspection image data to the manual annotation workstation, and after manual annotation, it is collected as new sample data for retraining the recognition model to achieve the update of the recognition model. Furthermore, the manual start-up conclusions and intelligent recognition results may not cover all prohibited items. During the manual unpacking process, the items are checked according to the approximate locations indicated by the manual start-up conclusions and intelligent recognition results. Therefore, there are cases where other possible prohibited items are not checked. When comparing and reviewing, the reviewer not only looks at the manual start-up conclusions and intelligent recognition results, but also performs overall image judgment. If any manual omissions are found, the luggage needs to be retrieved for unpacking inspection and the unpacking conclusions need to be supplemented.

[0041] Another implementation of step S106 may be that the difference between the operator's manual conclusion and the intelligent recognition result is not calculated, and the manual startup site pushes all security inspection image data to the manual review site for review and gives a manual review conclusion. This method is called a full review. Similarly, if the manual review conclusion determines that the intelligent recognition result is an underreporting (no suspected contraband is identified) or a false alarm (the suspected contraband is identified but the quantity, category and / or location are incorrectly identified), the manual review site pushes the current security inspection image data to the manual annotation workstation, and after manual annotation, it is collected as new sample data for retraining the recognition model to update the recognition model. In addition, if the manual review conclusion is that both the operator's manual conclusion and the intelligent recognition result are underreported (no suspected contraband is identified), the corresponding luggage will be immediately retrieved for unpacking inspection, and the final security inspection conclusion will be given.

[0042] Another implementation method is that the comparative review and the full review are performed simultaneously.

[0043] Furthermore, in this embodiment, step S107 is also included: if the manual unpacking site finds that the intelligent recognition result is underreported or falsely reported based on the manual unpacking conclusion, the corresponding security inspection image data is transmitted to the manual labeling site, and after manual labeling, it is collected as new sample data for retraining the recognition model.

[0044] Compared with manual image judgment, intelligent recognition will not be affected by fatigue and subjective state. Because it imitates human image judgment rules and learns image features, the performance level of intelligent recognition in the initial stage is usually not as good as that of manual image judgment. Based on this situation, in the method of this embodiment, when it is found that the intelligent recognition results are missed or falsely reported, the corresponding security inspection image data will be collected as new sample data, and the new sample data will be directly used to retrain the recognition model to quickly improve the performance of the recognition model. The method of this embodiment compares the manual judgment results (manual unpacking conclusions and / or manual review conclusions) with the intelligent recognition results to obtain sample data for retraining, and combines the work of collecting sample data with the existing work of security inspectors, which not only utilizes the professional ability of security inspectors in image judgment, but also does not increase their workload.

[0045] The three types of prohibited items that are of most concern in security checks are fire starters, knives, and lithium batteries. At present, there are already abundant samples of common fire starters such as lighters, household knives, and lithium batteries. However, due to the scarcity and sensitivity of samples, developers usually cannot easily obtain uncommon fire starters such as magnesium rods and uncommon controlled knives. For these samples, the processes of step S106 and step S107 can be used to collect them. You can also consider preparing samples on your own and using the existing technology of hazardous material injection (TIP) to expand the samples. However, it should be noted that in the process of expanding the samples, real data collected on the front line of security inspection needs to be used as the background for injection to prevent the background from being too single, resulting in overfitting of the model learning and poor generalization.

[0046] like Figure 2 As shown, the sample data collected through the process of step S106 and step S107 are divided into four types: false alarms of intelligent recognition results determined by manual unpacking conclusions (manual unpacking conclusion AI false alarms), missed alarms of intelligent recognition results determined by manual unpacking conclusions (manual unpacking conclusion AI missed alarms), missed alarms of intelligent recognition results determined by manual review conclusions (manual review conclusion AI missed alarms), and false alarms of intelligent recognition results determined by manual review conclusions (manual review conclusion AI false alarms). After being organized into incremental training data, the recognition model can be retrained.

[0047] The method of this embodiment also includes step S108: using the recognition model obtained from the most recent training as the initial model, using the initial model as the teacher model, and directly using the distillation method on the intelligent recognition server to use the collected new sample data to train the student model, that is, completing the retraining at the user's site, which can quickly meet the user's demand for upgrading the recognition model performance while meeting the data confidentiality requirements.

[0048] Specifically, the recognition model obtained by the most recent update is used as the initial model, and the initial model is used as the teacher model (if the current update is the first time, the initial model is the original recognition model deployed for the first time). The parameters of the teacher model are used to initialize the student model whose network structure is exactly the same as the teacher model. The loss function of the student model training is designed so that the student model does not reduce the recognition performance of the initial model while learning the characteristics of the new sample data. In this embodiment, a distillation loss term L based on the knowledge distillation technology is added to the loss function used when training the original recognition model. distill , which constitutes the loss function of student model training, L distill The calculation formula is as follows:

[0049]

[0050] The distillation loss measures the difference between the old model and the new model in the response of the old task. In essence, it uses the classification and bounding box learned by the old model on the old training data as the reference value for the new training. new and cls old Represents the category probability of the new model and the old model on the candidate area, box new and box old They represent the bounding box distribution of the new model and the old model on the candidate area respectively, and N is the number of data points used to calculate the average value.

[0051] Specifically, the output of the new model is used as prediction data, and the output of the old model is used as reference data. By calculating the mean of the sum of squared errors of the corresponding data points, i.e., the mean square error (MSE) loss, the output distribution of the new model is forced to converge to the old model, realizing knowledge transfer, thereby obtaining an incremental model and finally obtaining an updated recognition model.

[0052] The final loss function expression of the student model training is:

[0053] L=L original +λL distill

[0054] Among them, L original It is the loss function used when training the original recognition model deployed for the first time. It mainly consists of three parts of loss through weighted summation:

[0055] L original =aL box +bL cls +cL dfl

[0056] Example 2

[0057] This embodiment provides a security inspection full-line intelligent identification system for implementing the method of embodiment 1, such as Figure 3 As shown, it includes an intelligent identification server 10 and several manual startup sites 12, manual unpacking sites 11, manual review site one 13, and manual review site two 14; the manual startup sites, manual unpacking sites, manual review site one and manual review site two are all communicatively connected with the intelligent identification server, and the manual review site one and manual review site two are used for comparative review and full review respectively.

[0058] For those skilled in the art, various corresponding changes and modifications can be made according to the above technical solutions and concepts, and all of these changes and modifications should be included in the protection scope of the claims of the present invention.

Claims

1. A security inspection full-line intelligent identification method, characterized in that: The steps include: S101, using the security inspection data collection device to collect the security inspection image data of the current passenger's luggage, and send it to the intelligent recognition server; S102, the intelligent recognition server uses the recognition model to analyze the received image data and returns the intelligent recognition result to the security inspection data collection device, the intelligent recognition result including whether there are suspected contraband or suspected contraband and the category, quantity and location information of the suspected contraband; S103, the security inspection data collection device sends the security inspection image data and intelligent recognition results to the manual start-up site; the manual start-up site first displays the security inspection image data on the screen, and a professional operator interprets the security inspection image data and gives the operator's manual conclusion, which includes whether there are suspected contraband or suspected contraband and the type, quantity and location of the suspected contraband; S104. When at least one of the operator's manual conclusion and the intelligent recognition result determines that there are suspected contraband in the current security inspection image data, the manual operation site will send the current security inspection image data, the operator's manual conclusion and the intelligent recognition result to the manual unpacking site, which will display them on its screen. At the same time, the luggage corresponding to the current security inspection image data will be sorted to the unpacking workbench corresponding to the manual unpacking site, waiting for the unpacking operator to unpack and inspect. The unpacking operator obtains the manual unpacking conclusion of the corresponding luggage based on the unpacking inspection result, and the manual unpacking conclusion includes whether the luggage is safe or whether there are contraband in the luggage and the type, quantity and location information of the contraband; S105. The unpacker enters the manual unpacking conclusion obtained in step S104 at the manual unpacking site; for luggage judged to be safe, the unpacker's manual conclusion and intelligent recognition result are deleted; for luggage judged to contain contraband, the unpacker's manual conclusion or intelligent recognition result is confirmed or modified.

2. The method according to claim 1, characterized in that In step S102, when any one of the delayed prompt conditions is triggered, the manual startup site will display the intelligent recognition result on the screen with a delay to provide auxiliary prompts to the startup operator; Delay prompt conditions include: (1) the current security inspection image data is about to leave the screen of the manual start-up station; (2) The current security inspection image data starts timing from the screen entering the manual start-up site, and after a predefined period of time, the security inspector does not give a manual conclusion of the operator; (3) After the security inspector changes shifts, the transmission device of the security inspection data collection equipment is started again.

3. The method according to claim 1, characterized in that It also includes step S106: the manual startup site further calculates the difference between the manual conclusion of the operator and the intelligent recognition result. When the calculation result shows that the manual conclusion of the operator is inconsistent with the intelligent recognition result, that is, only one of the manual conclusion of the operator and the intelligent recognition result determines that there are suspected contraband in the current security inspection image data, or both the manual conclusion of the operator and the intelligent recognition result determine that there are suspected contraband in the current security inspection image data but there are differences in the judgment on the quantity, category and / or location of the suspected contraband, the manual startup site further pushes the current security inspection image data to the manual review site, which is reviewed by senior security personnel and a manual review conclusion is given; if the manual review conclusion determines that the intelligent recognition result is an underreporting or false alarm, the manual review site pushes the current security inspection image data to the manual labeling workstation, which is collected as new sample data after manual labeling for retraining the recognition model.

4. The method according to claim 1, characterized in that: The method further includes step S106: the manual start-up site pushes all security inspection image data to the manual review site for review, and gives a manual review conclusion; if the manual review conclusion determines that the intelligent recognition result is an underreporting or a false alarm, the manual review site pushes the current security inspection image data to the manual annotation workstation, and after manual annotation, it is collected as new sample data for retraining the recognition model; If the manual review concludes that both the operator's manual conclusion and the intelligent recognition result are underreported, the corresponding luggage will be immediately retrieved for unpacking inspection and the final security inspection conclusion will be given.

5. The method according to claim 1, characterized in that It also includes step S107: if the manual unpacking site finds that the intelligent recognition result is missed or falsely reported based on the manual unpacking conclusion, the corresponding security inspection image data is transmitted to the manual labeling site, and after manual labeling, it is collected as new sample data for retraining the recognition model.

6. The method according to any one of claims 3 to 5, characterized in that: Step S108: The recognition model obtained by the most recent update is used as the initial model, and the initial model is used as the teacher model. The parameters of the teacher model are used to initialize the student model whose network structure is exactly the same as the teacher model, and the student model is trained with new sample data; the distillation loss term L based on the knowledge distillation technology is added to the loss function used when training the original recognition model deployed for the first time. distill , which constitutes the loss function of student model training, L distill The calculation formula is as follows: cls new and cls old Represents the category probability of the new model and the old model on the candidate area, box new and box old Represent the bounding box distribution of the new model and the old model on the candidate area, respectively, and N is the number of data points used to calculate the average value: The final loss function expression of the student model training is: L=L original +λL distill Among them, L original is the loss function used when training the original recognition model that was first deployed.

7. A security inspection full-line intelligent identification system for implementing any of the methods of claims 1-6, characterized in that: It includes an intelligent identification server and several manual startup sites, manual unpacking sites, manual review site one and manual review site two. The manual startup sites, manual unpacking sites, manual review site one and manual review site two are all communicatively connected to the intelligent identification server; manual review site one and manual review site two are used for comparative review and full review respectively.

Images