A dangerous goods target identification method of millimeter wave security door adaptive to incremental learning
By generating echo data of the target to be identified and performing incremental learning, the adaptability and flexibility of dangerous goods target identification in millimeter-wave security gates under changing environments have been solved, and efficient identification of real-time data has been achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING HUAHANG RADIO MEASUREMENT & RES INST
- Filing Date
- 2024-12-30
- Publication Date
- 2026-06-30
AI Technical Summary
Existing methods for identifying dangerous goods targets using millimeter-wave security gates have poor adaptability and flexibility in changing environments, and the lack of incremental learning sample data leads to increased training time and resource requirements.
By acquiring batches of raw millimeter-wave security gate echo data of dangerous goods and their labels, a basic model is trained, and a matching echo test system is built to generate echo data of the target to be identified. Incremental learning is then performed to form a full model for identifying real-time data.
It enables flexible identification of targets in changing environments, reduces training time and resource requirements, and improves the adaptability and flexibility of dangerous goods target identification by millimeter-wave security gates.
Smart Images
Figure CN122307487A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of dangerous goods target recognition technology for millimeter-wave security gates, and more particularly to a method for dangerous goods target recognition using millimeter-wave security gates that can adapt to incremental learning. Background Technology
[0002] In the process of dangerous goods target identification in millimeter-wave security gates, due to the ever-changing application environment of target identification, it is necessary to ensure that the dangerous goods target identification method of millimeter-wave security gates can adapt to the changing environment, so that the adopted dangerous goods target identification model of millimeter-wave security gates has the ability of incremental learning and can dynamically update the dangerous goods target identification model of millimeter-wave security gates, thereby adapting to the changing application environment.
[0003] At the same time, during incremental learning, there may be a lack of sample data that matches the incremental target, which increases the difficulty of incremental learning.
[0004] Therefore, how to generate sample data that matches the incremental target and form a millimeter-wave security gate dangerous goods target recognition method that can adapt to incremental learning, improve the flexibility of millimeter-wave security gate dangerous goods target recognition, and at the same time reduce training time and resource requirements are the technical problems that urgently need to be solved. Summary of the Invention
[0005] Based on the above analysis, the embodiments of the present invention aim to provide a millimeter-wave security gate dangerous goods target recognition method that can adapt to incremental learning, in order to solve the problems of poor adaptability of millimeter-wave security gate dangerous goods target recognition to changing environments and poor flexibility of millimeter-wave security gate dangerous goods target recognition methods in the prior art.
[0006] This invention discloses a method for identifying hazardous materials targets in millimeter-wave security gates that can adapt to incremental learning. The method includes:
[0007] Acquire multiple sets of batch raw millimeter-wave security gate dangerous goods echo data and their corresponding target category labels, and train a basic millimeter-wave security gate dangerous goods target recognition model;
[0008] Based on the newly added dangerous goods targets to be identified by millimeter-wave security gates, a matching millimeter-wave security gate dangerous goods echo test system is built; under the millimeter-wave security gate dangerous goods echo test system, batch millimeter-wave security gate dangerous goods echo data of the targets to be identified are generated;
[0009] Based on the batch of millimeter-wave security gate dangerous goods echo data of newly added dangerous goods targets to be identified, the basic millimeter-wave security gate dangerous goods target identification model is incrementally learned to obtain the full millimeter-wave security gate dangerous goods target identification model;
[0010] Using a full-volume millimeter-wave security gate hazardous material target recognition model, hazardous material targets are identified from real-time millimeter-wave security gate echo data.
[0011] Based on the above solution, the present invention also makes the following improvements:
[0012] Furthermore, based on the newly added dangerous goods targets to be identified by the millimeter-wave security gate, a matching millimeter-wave security gate dangerous goods echo testing system was built and implemented:
[0013] Configure the millimeter-wave security gate hazardous materials equipment according to the scenario requirements; determine the target scattering characteristic model based on the newly added hazardous materials to be identified by the millimeter-wave security gate, and embed the target scattering characteristic model into the target equipment; construct the scenario of interaction between the millimeter-wave security gate hazardous materials equipment and the target equipment into a millimeter-wave security gate hazardous materials echo test system.
[0014] Under the millimeter-wave security gate dangerous goods echo test system, according to the scenario setting requirements and the target scattering characteristic model, batch millimeter-wave security gate dangerous goods echo data of the target to be identified are generated.
[0015] Furthermore, the step of generating batch millimeter-wave security gate dangerous goods echo data of the target to be identified is performed as follows:
[0016] Based on the parameters of the millimeter-wave security gate hazardous materials equipment in the scenario settings, the millimeter-wave security gate hazardous materials equipment generates a millimeter-wave security gate hazardous materials transmission signal.
[0017] The target device obtains the corresponding millimeter-wave security gate dangerous goods target echo data based on the target scattering characteristic model and the millimeter-wave security gate dangerous goods emission signal;
[0018] The combination of several millimeter-wave security gate dangerous goods target echo data is used as the batch millimeter-wave security gate dangerous goods echo data of the target to be identified.
[0019] Furthermore, incremental learning is performed on the basic millimeter-wave security gate dangerous goods target recognition model to obtain the full millimeter-wave security gate dangerous goods target recognition model, which is then executed as follows:
[0020] Add a target category branch to the basic millimeter-wave security gate dangerous goods target recognition model and freeze the weight parameters of the existing target category branches;
[0021] The batch of millimeter-wave security gate dangerous goods echo data of newly added dangerous goods targets to be identified are input into the basic millimeter-wave security gate dangerous goods target recognition model to train the weight parameters of the newly added target category branch, thereby realizing incremental learning of the basic millimeter-wave security gate dangerous goods target recognition model.
[0022] By restoring the weight parameters of the existing target category branches and combining them with the weight parameters of the newly added target category branches obtained through training, a full millimeter-wave security gate dangerous goods target recognition model is obtained.
[0023] Furthermore, the training of the basic millimeter-wave security gate dangerous goods target recognition model performs the following:
[0024] The basic millimeter-wave security gate dangerous goods target recognition model receives a batch of raw millimeter-wave security gate dangerous goods echo data, and performs feature extraction and dimension transformation on the batch of raw millimeter-wave security gate dangerous goods echo data to obtain a feature map of each target category;
[0025] The basic millimeter-wave security gate dangerous goods target recognition model performs vector aggregation on the feature maps of each target category to obtain the total score for each target category; and inputs the total score of all target categories into the Softmax function to calculate the predicted probability of each target category.
[0026] The loss function is calculated based on the predicted probability of each target category output by the Softmax function and the true target category label. The gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm, and the parameters of the basic millimeter-wave security gate dangerous goods target recognition model are updated.
[0027] Determine whether the training termination condition of the basic millimeter-wave security gate dangerous goods target recognition model has been met. If it has been met, stop training and obtain the successfully trained millimeter-wave security gate dangerous goods target recognition model.
[0028] Furthermore, the batch of raw millimeter-wave security gate dangerous goods echo data Where B represents the batch size, C in Indicates the number of input channels; D and R represent the dimensions of the original millimeter-wave security gate hazardous material echo data, where D represents the size of the original millimeter-wave security gate hazardous material echo data in the height dimension, and R represents the size of the original millimeter-wave security gate hazardous material echo data in the width dimension.
[0029] Furthermore, feature extraction is performed on the batch of raw millimeter-wave security gate hazardous material echo data, and the following steps are executed:
[0030] Feature extraction is performed on the batch of raw millimeter-wave security gate dangerous goods echo data using a backbone feature extraction network to obtain a high-level feature map of the batch of raw millimeter-wave security gate dangerous goods echo data;
[0031] The high-level feature map Where C represents the number of feature channels in the high-level feature map, H represents the height of the high-level feature map, and W represents the width of the high-level feature map.
[0032] Furthermore, the batch of raw millimeter-wave security gate hazardous material echo data undergoes dimensional transformation, and the following steps are performed:
[0033] Spatial dimension feature expansion is performed on the high-level feature map of the batch of raw millimeter-wave security gate dangerous goods echo data to obtain the corresponding two-dimensional feature map;
[0034] A nonlinear transformation of the feature dimension is performed on the two-dimensional feature map, and the corresponding target category dimension is added during the nonlinear transformation process;
[0035] The feature map with the corresponding target category dimension is reshaped into a target category branch, resulting in a feature map for each target category.
[0036] Furthermore, vector aggregation is performed on the feature maps of each target category to obtain the total score for each target category, and then the following steps are executed:
[0037] The spatial dimension HW is reduced for the feature map of each target category to obtain the score vector of each target category;
[0038] The score vectors of each target category across all feature channels are summed to obtain the total score for that target category.
[0039] Furthermore, the predicted probability P of target category s s Represented as:
[0040]
[0041] The exp operation typically refers to exponentiation with the base e of the natural logarithm; score s The total score for the target category s;
[0042] Total score of target category s s Represented as:
[0043]
[0044] Among them, v s,j The score vector v represents the target category s. s The score vector of the j-th feature channel in the feature channel dimension.
[0045] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:
[0046] The present invention provides an adaptive incremental learning method for identifying dangerous goods targets in millimeter-wave security gates. This method can build a matching millimeter-wave security gate dangerous goods echo testing system based on newly added dangerous goods targets to be identified. Under this system, batch millimeter-wave security gate dangerous goods echo data of the targets to be identified is generated. Based on the basic millimeter-wave security gate dangerous goods target identification model, incremental learning is performed on the model using the batch of newly added dangerous goods echo data to obtain a full-scale millimeter-wave security gate dangerous goods target identification model. This full-scale model can then be used to identify dangerous goods targets in real-time millimeter-wave security gates based on the collected echo data.
[0047] In summary, this invention, by constructing a millimeter-wave security gate hazardous materials echo testing system, generates batches of millimeter-wave security gate hazardous materials echo data for identification targets. This data is then used to incrementally train the millimeter-wave security gate hazardous materials target identification model. Consequently, the resulting full-scale millimeter-wave security gate hazardous materials target identification model can identify echo data including the categories of hazardous materials to be identified. This effectively solves the problems of poor adaptability of existing millimeter-wave security gate hazardous materials target identification to changing environments and poor flexibility of existing identification methods.
[0048] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description
[0049] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.
[0050] Figure 1 A flowchart illustrating a millimeter-wave security gate hazardous material target identification method with adaptive incremental learning provided in an embodiment of the present invention. Detailed Implementation
[0051] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.
[0052] A specific embodiment 1 of the present invention discloses a method for identifying dangerous goods targets in millimeter-wave security gates that can adapt to incremental learning. The flowchart of the method is as follows. Figure 1 As shown, the specific implementation process of the method in this embodiment is described below.
[0053] Step S1: Obtain multiple sets of batch raw millimeter-wave security gate dangerous goods echo data and their corresponding target category labels, and train a basic millimeter-wave security gate dangerous goods target recognition model.
[0054] Step S2: Based on the newly added dangerous goods targets to be identified by the millimeter-wave security gate, build a matching millimeter-wave security gate dangerous goods echo test system; under the millimeter-wave security gate dangerous goods echo test system, generate batch millimeter-wave security gate dangerous goods echo data of the targets to be identified.
[0055] Step S3: Based on the batch echo data of the newly added dangerous goods targets to be identified by the millimeter-wave security gate, perform incremental learning on the basic millimeter-wave security gate dangerous goods target identification model to obtain the full millimeter-wave security gate dangerous goods target identification model.
[0056] Step S4: Using the full-volume millimeter-wave security gate dangerous goods target recognition model, perform millimeter-wave security gate dangerous goods target recognition on the collected real-time millimeter-wave security gate dangerous goods echo data.
[0057] The training process and subsequent incremental learning process of the basic millimeter-wave security gate dangerous goods target recognition model will be described in detail later. Below, we will first provide a detailed explanation of the specific implementation process of step S2.
[0058] Step S21: Based on the scenario setting requirements of the newly added millimeter-wave security gate dangerous goods target to be identified, select a matching target scattering characteristic model and build a millimeter-wave security gate dangerous goods echo test system.
[0059] Preferably, in this embodiment, the scene setting requirements include: (1) the target type of the newly added millimeter-wave security gate hazardous material target to be identified; (2) the millimeter-wave security gate hazardous material equipment parameters, including: the millimeter-wave security gate hazardous material band and polarization state; (3) the scene environment parameters, including: environmental information, the altitude of the millimeter-wave security gate hazardous material equipment, and the millimeter-wave security gate hazardous material-target distance. Based on the newly added millimeter-wave security gate hazardous material target to be identified, the target scattering characteristic model is determined. Specifically, the target scattering characteristic model consists of the range image position of several stable scattering points on the four polarization channels HH, HV, VH and VV, the scattering point phase and the scattering point amplitude.
[0060] Configure a millimeter-wave security gate hazardous materials equipment according to the scenario requirements. Integrate the target scattering characteristic model into the target equipment. Preferably, the millimeter-wave security gate hazardous materials equipment includes a millimeter-wave security gate hazardous materials transmitting module, a millimeter-wave security gate hazardous materials transmitting antenna, and a millimeter-wave security gate hazardous materials receiving antenna; the target equipment also includes a polarized antenna. Construct a scenario of interaction between the millimeter-wave security gate hazardous materials equipment and the target equipment as a millimeter-wave security gate hazardous materials echo testing system.
[0061] Step S22: Under the millimeter-wave security gate dangerous goods echo test system, generate batch millimeter-wave security gate dangerous goods echo data of the target to be identified, according to the scene setting requirements and the target scattering characteristic model.
[0062] Step S221: Based on the parameters of the millimeter-wave security gate hazardous materials equipment in the scenario settings, the millimeter-wave security gate hazardous materials equipment generates a millimeter-wave security gate hazardous materials transmission signal.
[0063] In the specific implementation process, the millimeter-wave security gate dangerous goods transmission module generates a millimeter-wave security gate dangerous goods transmission signal, and transmits the millimeter-wave security gate dangerous goods transmission signal to the target device via the millimeter-wave security gate dangerous goods transmission antenna.
[0064] Step S222: The target device obtains the corresponding millimeter-wave security gate dangerous goods target echo data based on the target scattering characteristic model and the millimeter-wave security gate dangerous goods emission signal.
[0065] The target device transmits millimeter-wave security gate dangerous goods echo signals through a polarized antenna, and the millimeter-wave security gate dangerous goods receiving antenna receives the millimeter-wave security gate dangerous goods echo data transmitted by the polarized antenna.
[0066] Step S223: Repeat steps S221-S223 to generate batch millimeter-wave security gate dangerous goods echo data of the target to be identified.
[0067] The training process of the basic millimeter-wave security gate dangerous goods target recognition model is explained below.
[0068] Step S11: The basic millimeter-wave security gate dangerous goods target recognition model receives a batch of raw millimeter-wave security gate dangerous goods echo data, performs feature extraction and dimension transformation on the batch of raw millimeter-wave security gate dangerous goods echo data, and obtains the feature map of each target category.
[0069] In this embodiment, batch raw millimeter-wave security gate dangerous goods echo data Where B represents the batch size, that is, the number of data samples that are input into the basic millimeter-wave security gate dangerous goods target recognition model (implemented based on a deep learning model architecture) at one time during the training process; C in The input channel number (i.e., the number of characteristic channels of the original millimeter-wave security gate hazardous materials echo data, exemplarily including: the number of polarization channels and / or the number of characteristic channels after spectral decomposition); D and R represent the size of the original millimeter-wave security gate hazardous materials echo data, where D represents the size of the original millimeter-wave security gate hazardous materials echo data in the height dimension and R represents the size of the original millimeter-wave security gate hazardous materials echo data in the width dimension.
[0070] Step S111: Extract features from the batch of raw millimeter-wave security gate dangerous goods echo data using a backbone feature extraction network to obtain a high-level feature map of the batch of raw millimeter-wave security gate dangerous goods echo data.
[0071] In this embodiment, the backbone feature extraction network is preferably ResNet or DarkNet, which extracts high-level feature maps from the raw millimeter-wave security gate hazardous material echo data. Preferably, high-level feature maps are extracted from batches of raw millimeter-wave security gate hazardous material echo data. Where C represents the number of feature channels in the high-level feature map, H represents the height of the high-level feature map, and W represents the width of the high-level feature map.
[0072] In practice, the backbone feature extraction network can perform multiple convolution calculations on batches of raw millimeter-wave security gate dangerous goods echo data (using convolutional layers of different scales), increasing the number of channels from C... in The convolutional layers are expanded to C, and D and R are pruned multiple times according to their size to obtain H and W.
[0073] Step S112: Perform spatial dimension HW feature expansion on the high-level feature map of the batch of raw millimeter-wave security gate dangerous goods echo data to obtain the corresponding two-dimensional feature map.
[0074] Specifically, in this embodiment, the high-level feature map of the batch of raw millimeter-wave security gate dangerous goods echo data can be expanded in spatial dimension HW according to formula (1) to obtain the corresponding two-dimensional feature map:
[0075] F flat =reshape(F backbone (B,C,HW)) (1)
[0076] Where HW = H × W; reshape represents dimensional transformation; reshape(F backbone (B,C,HW)) represents the condition for Fbackbone Perform dimensional transformation according to (B,C,HW) to realize the feature expansion of spatial dimension HW.
[0077] Step S113: Perform a nonlinear transformation on the feature dimension of the two-dimensional feature map, and add the corresponding target category dimension during the nonlinear transformation process.
[0078] Specifically, in this embodiment, the two-dimensional feature map is subjected to a nonlinear transformation of the feature dimension according to formula (2), and the corresponding target category dimension S is added:
[0079]
[0080] Wherein, Linear represents a nonlinear transformation; Indicates F flat A nonlinear transformation is performed, and the target category dimension S is added during the nonlinear transformation process to obtain a feature map with dimensions B×SC×HW.
[0081] In the specific implementation process, it can be done in F flat Based on this, a corresponding target category dimension S is added, and a 2D CNN is used to process the feature map F of dimension B×C×HW. flat Convolution calculations are performed to adjust the feature dimensions to B×SC×HW to achieve a non-linear transformation, resulting in F. linear Alternatively, the feature map F with dimensions B×C×HW can also be used. flat The feature map is adjusted to BHW×C, then further adjusted to BHW×SC through multiple fully connected layers to achieve non-linear transformation. Finally, a feature map F with dimensions of B×SC×HW is obtained through dimensionality transformation. linear The purpose of different implementation methods is to adjust the dimensions of the B×C×HW feature map to the dimensions of B×SC×HW.
[0082] Step S114: Reshape the feature map with the corresponding target category dimension into a target category branch to obtain the feature map of each target category.
[0083] Specifically, according to formula (3), the feature map with the corresponding target category dimension is reshaped into a target category branch to obtain the feature map of each target category.
[0084] The feature map after nonlinear transformation is divided into branches for each target category (i.e., branches specific to each target category), resulting in:
[0085] F class =reshape(F linear (B,S,C,HW)) (3)
[0086] At this point, the feature map for each target category s∈[1,S] is represented as follows: reshape(F linear ,(B,S,C,HW)), represents the condition for F linear Perform a dimensional transformation according to (B, S, C, HW) to obtain F. class .
[0087] Step S12: The feature map of each target category is vector-aggregated by the basic millimeter-wave security gate dangerous goods target recognition model to obtain the total score of each target category; and the total score of all target categories is input into the Softmax function to calculate the predicted probability of each target category.
[0088] Step S121: Reduce the spatial dimension HW of the feature map for each target category to obtain the score vector for each target category.
[0089] Preferably, in this embodiment, the feature map of each target category is summed along the spatial dimension HW, as shown in formula (4), to reduce the dimensionality and form the score vector of the corresponding target category.
[0090] The score vector v of the target category s s Represented as:
[0091]
[0092] In addition, MaxPooling or AvgPooling can be performed on the feature map of each target category along the spatial dimension HW to reduce dimensionality and form the score vector of the corresponding target category.
[0093] Step S122: Sum the score vectors of each target category across all feature channels to obtain the total score for the corresponding target category.
[0094] Specifically, the total score of the target category s s Represented as:
[0095]
[0096] Among them, v s,j The score vector v represents the target category s. s The score vector of the j-th feature channel in the feature channel dimension.
[0097] Step S123: Input the total score of all target categories into the Softmax function to calculate the predicted probability of each target category.
[0098] Specifically, the predicted probability P of the target category s s Represented as:
[0099]
[0100] The exp operation usually refers to the exponential operation with the base e of the natural logarithm.
[0101] Step S13: Calculate the loss function based on the predicted probability of each target category output by the Softmax function and the true target category label, and use the backpropagation algorithm to calculate the gradient of the loss function with respect to the model parameters, and update the parameters of the millimeter wave security gate dangerous goods target prediction model.
[0102] In this embodiment, for example, the cross-entropy loss function can be used as the model's loss function. For multi-class classification problems, the cross-entropy loss function can measure the difference between the model's predicted probabilities and the true labels.
[0103] Step S14: Determine whether the training termination condition of the millimeter-wave security gate dangerous goods target prediction model has been met. If it has been met, stop training and obtain the millimeter-wave security gate dangerous goods target prediction model that has passed training.
[0104] Preferably, in this embodiment, the training termination condition may be: a predetermined number of iterations, or the gradient of the loss function with respect to the model parameters being less than a set threshold.
[0105] In practical implementation, the basic millimeter-wave security gate dangerous goods target recognition model can be implemented using a deep learning model. During the training of the basic millimeter-wave security gate dangerous goods target recognition model, a loss function is defined, and the gradient of the loss function with respect to the deep learning model parameters is calculated using the backpropagation algorithm, calculating the gradient layer by layer from the output layer to the input layer. Gradient descent or its variants (such as Adam, RMSprop, etc.) are used to update the neural network model parameters based on the calculated gradients. This process is repeated until the performance of the neural network model no longer improves, or until a predetermined number of iterations is reached, thus obtaining a successfully trained basic millimeter-wave security gate dangerous goods target recognition model.
[0106] Once the basic millimeter-wave security gate hazardous material target recognition model has been successfully trained, it can be used for hazardous material target recognition within the millimeter-wave security gate system. Specifically, the collected real-time millimeter-wave security gate hazardous material echo data is directly input into the trained model, which then predicts and outputs the corresponding hazardous material target recognition result. It should be noted that during the training process, to accelerate the training speed, batches of raw millimeter-wave security gate hazardous material echo data are used to train the model. However, when using the trained model for actual target recognition and prediction, the collected real-time millimeter-wave security gate hazardous material echo data is a single, individual data point. That is, B=1. Then, referring to the processing method of the basic millimeter-wave security gate dangerous goods target recognition model in this embodiment, the real-time millimeter-wave security gate dangerous goods echo data is processed, and then the basic millimeter-wave security gate dangerous goods target recognition model can predict and output the corresponding millimeter-wave security gate dangerous goods target recognition result.
[0107] When a new hazardous material target is added to the millimeter-wave security gate for identification, the operation in step S2 is executed to generate batch millimeter-wave security gate hazardous material echo data for the target. Then, based on the batch millimeter-wave security gate hazardous material echo data of the newly added targets, incremental learning is performed on the basic millimeter-wave security gate hazardous material target identification model to obtain the full millimeter-wave security gate hazardous material target identification model. This is the operation performed in step S3. The implementation process of step S3 is explained in detail below.
[0108] Step S31: Add a target category branch to the basic millimeter-wave security gate dangerous goods target recognition model, and freeze the weight parameters of the existing target category branches.
[0109] Step S32: Input the batch millimeter-wave security gate dangerous goods echo data of the newly added millimeter-wave security gate dangerous goods target to be identified into the basic millimeter-wave security gate dangerous goods target recognition model to train the weight parameters of the newly added target category branch, so as to realize incremental learning of the basic millimeter-wave security gate dangerous goods target recognition model.
[0110] Step S33: Restore the weight parameters of the existing target category branches, and combine them with the weight parameters of the newly added target category branches obtained from training to obtain the full millimeter-wave security gate dangerous goods target recognition model.
[0111] That is, the model updated through incremental learning is integrated with the original basic model to form a full-scale millimeter-wave security gate dangerous goods target recognition model.
[0112] In step S4, the full-volume millimeter-wave security gate hazardous materials target recognition model is used to identify hazardous materials targets from the collected real-time millimeter-wave security gate echo data. It should be noted that the target identification results from the real-time millimeter-wave security gate echo data at this point may include previously existing target categories or newly added target categories. This demonstrates the full-volume millimeter-wave security gate hazardous materials target recognition model's ability to identify newly added target categories, effectively solving the problems of poor adaptability to changing environments and limited flexibility in existing millimeter-wave security gate hazardous materials target recognition methods.
[0113] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.
[0114] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for identifying dangerous goods targets in millimeter-wave security gates that can adapt to incremental learning, characterized in that, The method includes: Acquire multiple sets of batch raw millimeter-wave security gate dangerous goods echo data and their corresponding target category labels, and train a basic millimeter-wave security gate dangerous goods target recognition model; Based on the newly added dangerous goods targets to be identified by millimeter-wave security gates, a matching millimeter-wave security gate dangerous goods echo test system is built; under the millimeter-wave security gate dangerous goods echo test system, batch millimeter-wave security gate dangerous goods echo data of the targets to be identified are generated; Based on the batch of millimeter-wave security gate dangerous goods echo data of newly added dangerous goods targets to be identified, the basic millimeter-wave security gate dangerous goods target identification model is incrementally learned to obtain the full millimeter-wave security gate dangerous goods target identification model; Using a full-volume millimeter-wave security gate hazardous material target recognition model, hazardous material targets are identified from real-time millimeter-wave security gate echo data.
2. The method for identifying dangerous goods targets in millimeter-wave security gates with adaptive incremental learning according to claim 1, characterized in that, Based on the newly added dangerous goods targets to be identified by the millimeter-wave security gate, a matching millimeter-wave security gate dangerous goods echo testing system was built and implemented: Configure the millimeter-wave security gate hazardous materials equipment according to the scenario requirements; determine the target scattering characteristic model based on the newly added hazardous materials target to be identified by the millimeter-wave security gate, and embed the target scattering characteristic model into the target equipment; The scenario of interaction between the millimeter-wave security gate's hazardous materials equipment and the target equipment is constructed into a millimeter-wave security gate hazardous materials echo testing system; Under the millimeter-wave security gate dangerous goods echo test system, according to the scenario setting requirements and the target scattering characteristic model, batch millimeter-wave security gate dangerous goods echo data of the target to be identified are generated.
3. The millimeter-wave security gate dangerous goods target identification method with adaptive incremental learning according to claim 2, characterized in that, The process of generating batch millimeter-wave security gate dangerous goods echo data for the target to be identified is performed as follows: Based on the parameters of the millimeter-wave security gate hazardous materials equipment in the scenario settings, the millimeter-wave security gate hazardous materials equipment generates a millimeter-wave security gate hazardous materials transmission signal. The target device obtains the corresponding millimeter-wave security gate dangerous goods target echo data based on the target scattering characteristic model and the millimeter-wave security gate dangerous goods emission signal; The combination of several millimeter-wave security gate dangerous goods target echo data is used as the batch millimeter-wave security gate dangerous goods echo data of the target to be identified.
4. The millimeter-wave security gate dangerous goods target identification method with adaptive incremental learning according to any one of claims 1-3, characterized in that, Incremental learning is performed on the basic millimeter-wave security gate dangerous goods target recognition model to obtain the full millimeter-wave security gate dangerous goods target recognition model, and then the following is executed: Add a target category branch to the basic millimeter-wave security gate dangerous goods target recognition model and freeze the weight parameters of the existing target category branches; The batch of millimeter-wave security gate dangerous goods echo data of newly added dangerous goods targets to be identified are input into the basic millimeter-wave security gate dangerous goods target recognition model to train the weight parameters of the newly added target category branch, thereby realizing incremental learning of the basic millimeter-wave security gate dangerous goods target recognition model. By restoring the weight parameters of the existing target category branches and combining them with the weight parameters of the newly added target category branches obtained through training, a full millimeter-wave security gate dangerous goods target recognition model is obtained.
5. The millimeter-wave security gate dangerous goods target identification method with adaptive incremental learning according to claim 4, characterized in that, The trained basic millimeter-wave security gate dangerous goods target recognition model performs the following: The basic millimeter-wave security gate dangerous goods target recognition model receives a batch of raw millimeter-wave security gate dangerous goods echo data, and performs feature extraction and dimension transformation on the batch of raw millimeter-wave security gate dangerous goods echo data to obtain a feature map of each target category; The feature maps of each target category are aggregated using the basic millimeter-wave security gate dangerous goods target recognition model to obtain the total score for each target category. The total score of all target categories is then input into the Softmax function to calculate the predicted probability of each target category. The loss function is calculated based on the predicted probability of each target category output by the Softmax function and the true target category label. The gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm, and the parameters of the basic millimeter-wave security gate dangerous goods target recognition model are updated. Determine whether the training termination condition of the basic millimeter-wave security gate dangerous goods target recognition model has been met. If it has been met, stop training and obtain the successfully trained millimeter-wave security gate dangerous goods target recognition model.
6. The millimeter-wave security gate dangerous goods target identification method with adaptive incremental learning according to claim 5, characterized in that, The batch of raw millimeter-wave security gate dangerous goods echo data Where B represents the batch size, C in This indicates the number of input channels; D and R represent the dimensions of the original millimeter-wave security gate hazardous material echo data, with D representing the size of the original millimeter-wave security gate hazardous material echo data in the height dimension and R representing the size of the original millimeter-wave security gate hazardous material echo data in the width dimension.
7. The method for identifying dangerous goods targets in millimeter-wave security gates with adaptive incremental learning according to claim 6, characterized in that, Feature extraction is performed on the batch of raw millimeter-wave security gate hazardous material echo data, and the following steps are executed: Feature extraction is performed on the batch of raw millimeter-wave security gate dangerous goods echo data using a backbone feature extraction network to obtain a high-level feature map of the batch of raw millimeter-wave security gate dangerous goods echo data; The high-level feature map Where C represents the number of feature channels in the high-level feature map, H represents the height of the high-level feature map, and W represents the width of the high-level feature map.
8. The method for identifying dangerous goods targets in millimeter-wave security gates with adaptive incremental learning according to claim 7, characterized in that, The batch of raw millimeter-wave security gate hazardous material echo data is subjected to dimensional transformation, and the following steps are performed: Spatial dimension feature expansion is performed on the high-level feature map of the batch of raw millimeter-wave security gate dangerous goods echo data to obtain the corresponding two-dimensional feature map; A nonlinear transformation of the feature dimension is performed on the two-dimensional feature map, and the corresponding target category dimension is added during the nonlinear transformation process; The feature map with the corresponding target category dimension is reshaped into a target category branch, resulting in a feature map for each target category.
9. The millimeter-wave security gate dangerous goods target identification method with adaptive incremental learning according to claim 8, characterized in that, Perform vector aggregation on the feature maps of each target category to obtain the total score for each target category, and then execute: The spatial dimension HW is reduced for the feature map of each target category to obtain the score vector of each target category; The score vectors of each target category across all feature channels are summed to obtain the total score for that target category.
10. The method for identifying dangerous goods targets in millimeter-wave security gates with adaptive incremental learning according to claim 9, characterized in that, Predicted probability P of target category s s Represented as: The exp operation typically refers to exponentiation with the base e of the natural logarithm; score s The total score for the target category s; Total score of target category s s Represented as: Among them, v s,j The score vector v represents the target category s. s The score vector of the j-th feature channel in the feature channel dimension.