A convolutional-long short-term memory neural network item detection and recognition method

Abstract

The present application relates to the technical field of security inspection, and particularly relates to a convolution-long short-term memory neural network article detection and identification method, comprising the following steps: converting a current millimeter wave image data frame into M parallel memory-type millimeter wave image data frames, wherein M is greater than or equal to 2; outputting the M parallel memory-type millimeter wave image data frames one by one to M spatial attention convolution modules for processing to obtain M first weighted feature maps; outputting the M first weighted feature maps one by one to M channel attention LSTM modules for processing to obtain M second weighted feature maps, which are used for subsequent contraband detection and identification. The present application realizes automatic identification of contraband in a millimeter wave image by fusing an attention mechanism, CNN and LSTM, so as to improve the identification accuracy of contraband and meet the urgent needs of entry and exit site security for efficient, safe and highly privacy-protecting security technologies.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of security inspection, and particularly relates to a convolution-long short-term memory neural network article detection and recognition method. BACKGROUND

[0002] As an advanced security inspection method, millimeter wave human body imaging technology has a wide detection range. It can not only detect various metal articles, but also identify weapons, knives, liquids, powders and fire sources, greatly enhancing the comprehensiveness and accuracy of security inspection. The high-resolution function of millimeter wave imaging can clearly display the position and shape of prohibited articles, which helps security personnel make quick judgments. At the same time, as a safe electromagnetic wave, millimeter wave does not penetrate the human body and has no ionizing radiation, so it is very safe for children, the elderly, pregnant women and other special groups, and therefore has a wide application prospect in the field of security inspection.

[0003] Traditional security inspection methods often rely on manual observation and judgment, which is not only inefficient, but also easily affected by human factors, leading to missed detection or false positives. Therefore, it is particularly important to develop an intelligent detection and recognition method that can automatically, efficiently and accurately identify prohibited articles in millimeter wave images. However, the efficient and accurate identification of prohibited articles in millimeter wave images is still a difficult problem to be solved. In order to solve the problem of efficient identification of prohibited articles in millimeter wave images, deep learning technology is introduced into the field of security inspection. Deep learning is a computational model that simulates the neural network of the human brain. Through learning and training, it can realize the automatic identification and classification of objects. In object recognition, convolutional neural network (CNN) is one of the most commonly used structures. CNN extracts features in images through multiple convolutional layers and pooling layers, and classifies through fully connected layers. This structure enables the neural network to effectively process image data and has high recognition accuracy.

[0004] However, a single CNN model may not be able to fully capture the key information in complex and variable millimeter wave images, resulting in poor recognition results, which needs to be further improved. SUMMARY

[0005] The present application aims to provide a convolution-long short-term memory neural network article detection and recognition method, which introduces long short-term memory network (LSTM) and attention mechanism to improve recognition performance. LSTM is a special recurrent neural network (RNN) that can handle long sequence data and capture temporal dependencies in data. The attention mechanism enables the model to perform dot product operations on the current hidden state and the hidden state output by the encoder at each time step of the decoder, producing attention scores, thereby dynamically focusing on the importance of different parts of the input sequence.

[0006] The application fuses attention mechanism, CNN and LSTM, and proposes a convolutional-long short-term memory neural network article detection and recognition method, realizes automatic recognition of prohibited articles in a millimeter wave image, improves the recognition accuracy of prohibited articles, reduces the false positive rate and the missed detection rate, guarantees the safety of security check, avoids exposure of the millimeter wave scanning image of passengers to others, and protects the personal privacy of passengers. The application mainly originates from deep fusion of millimeter wave imaging technology and deep learning technology, so as to meet the urgent needs of entry and exit on-site security check for efficient, safe and privacy protection security check technology.

[0007] To achieve the purpose of the application, the application adopts the following technical solutions:

[0008] A convolutional-long short-term memory neural network article detection and recognition method comprises the following steps:

[0009] The current millimeter wave image data frame is converted into M parallel memory type millimeter wave image data frames, wherein M is greater than or equal to 2;

[0010] The M parallel memory type millimeter wave image data frames are one-to-one correspondingly output to M spatial attention convolution modules for processing to obtain M first weighted feature maps;

[0011] The M first weighted feature maps are one-to-one correspondingly output to M channel attention LSTM modules for processing to obtain M second weighted feature maps. Figure One The M second weighted feature maps are used for subsequent prohibited article detection and recognition.

[0012] Further improvement lies in that each spatial attention convolution module comprises a first convolution layer, a first maximum pooling layer, a spatial attention mechanism module and a first weighted fusion module.

[0013] Further improvement lies in that the specific method for one-to-one corresponding output of the M parallel memory type millimeter wave image data frames to the M spatial attention convolution modules for processing to obtain the M first weighted feature maps comprises:

[0014] The input memory type millimeter wave image data frame is preliminarily feature-extracted through the first convolution layer to generate a first original feature map;

[0015] The first original feature map is processed through the first maximum pooling layer to reduce the calculation amount while retaining important features;

[0016] The first original feature map processed through the first maximum pooling layer is processed through the spatial attention mechanism module to generate a spatial attention map;

[0017] The spatial attention map and the first original feature map are multiplied element by element through the first weighted fusion module, effective fusion of features is realized, and a first weighted feature map is generated.

[0018] Further improvement is that each of the channel attention LSTM modules comprises a first LSTM unit, a second max-pooling layer, a channel attention mechanism module and a second weighted fusion module.

[0019] Further improvement is that the M first weighted feature maps are input into the M channel attention LSTM modules for processing to obtain M second weighted feature maps. Figure One A specific method for outputting the M second weighted feature maps to the M channel attention LSTM modules for processing includes:

[0020] The first LSTM unit and the second max-pooling layer are used to process the input first weighted feature maps to further extract deep feature information, thereby obtaining a second original feature map.

[0021] The channel attention mechanism module is used to process the second original feature map to generate a channel attention map.

[0022] The second weighted fusion module is used to multiply the channel attention map and the second original feature map output after the second max-pooling layer processing, thereby realizing effective fusion of the features and generating a second weighted feature map.

[0023] Further improvement is that the article detection and recognition method further comprises the following steps:

[0024] The M second weighted feature maps obtained by processing the M channel attention LSTM modules are cascaded by a cascading layer to form a weighted feature panoramic map covering M parallel memory-type millimeter wave image data frames.

[0025] The weighted feature panoramic map is input into a second LSTM unit and a third max-pooling layer for deep feature information extraction of the prohibited articles, wherein the second LSTM unit is used to capture the time dependence in the feature sequence, and the third max-pooling layer is used to reduce the dimension of the data and retain the most important feature information.

[0026] Further improvement is that the article detection and recognition method further comprises the following steps:

[0027] The extracted prohibited article feature information is input into a fully connected layer for feature fusion and classification decision.

[0028] Further improvement is that the fully connected layer comprises at least two fully connected networks, and the fully connected layer maps the feature information to a classification space through linear transformation and activation function, thereby realizing detection and recognition of the prohibited articles, and finally outputting the feature classification result to determine whether the prohibited articles are contained and the type of the prohibited articles.

[0029] Further improvement lies in that the article detection and identification method further comprises the steps of marking the detected prohibited articles and visualizing display of location and classification information.

[0030] Further improvement lies in that converting the current millimeter wave image data frame into M parallel memory type millimeter wave image data frames is achieved by using a time delay neural network.

[0031] The present application has the following advantages:

[0032] The CNN-LSTM neural network article detection and identification method proposed by the present application has significant technical advantages compared with existing entry and exit on-site security millimeter wave imaging processing technology.

[0033] Firstly, in the data preprocessing stage, the present application converts the current millimeter wave image data frame into M parallel memory type millimeter wave image data frames by using a time delay neural network. This innovative conversion method effectively improves the data processing speed and parallel processing capability, laying a solid foundation for subsequent efficient feature extraction.

[0034] Secondly, in the feature extraction aspect, the present application combines spatial attention mechanism and channel attention mechanism, and processes the continuous M frames of millimeter wave image data frames in parallel through M spatial attention convolution modules and M channel attention LSTM modules. The use of this double attention mechanism enables the system to focus more accurately on the key information in the image, effectively suppressing irrelevant information interference, thereby improving the accuracy and robustness of prohibited article identification.

[0035] In addition, the present application further extracts deep-level feature information in the millimeter wave image through LSTM layers and maximum pooling layers, which not only enhances the recognition ability of the system, but also improves the identification accuracy of prohibited articles in complex scenes.

[0036] Finally, the present application realizes intelligent prohibited article detection and discrimination by fusing and classifying the features of prohibited articles through a fully connected layer. This intelligent detection method not only improves the security efficiency, but also greatly reduces the work burden of security personnel, providing strong support for improving security efficiency and accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 The flowchart of the CNN-LSTM neural network article detection and identification method of the present application;

[0038] Figure 2 The structure and function block diagram of the CNN-LSTM neural network article detection and identification method of the present application;

[0039] Figure 3 Memory type millimeter wave image data frame conversion schematic diagram in input pre-processing layer;

[0040] Figure 4 Structural function block diagram of spatial attention convolution module;

[0041] Figure 5 Structural function block diagram of channel attention LSTM module. DETAILED DESCRIPTION

[0042] In order for those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the scope of protection of the present application.

[0043] It should be noted that the terms "first", "second" and the like in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units does not necessarily limit to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0044] The following is an explanation of the technical terms involved in the present application:

[0045] Millimeter wave security imaging: This is a technology that uses millimeter waves (a type of electromagnetic wave) to scan the human body or objects, which can generate X-ray-like photos for checking whether they carry prohibited items.

[0046] Convolutional neural network (CNN): A special deep learning model that is particularly good at processing image data, which automatically learns features from images by simulating the way the human brain works, thereby performing classification or recognition tasks.

[0047] Long short-term memory network LSTM: A special recurrent neural network (RNN) that can handle long sequence data and capture temporal dependencies in data.

[0048] Attention mechanism: This is a technique that mimics human vision, allowing the model to "focus" on certain specific areas when processing images, while ignoring other unimportant parts.

[0049] In the entry and exit site, the millimeter wave security imaging system plays a vital role, but the traditional method often faces challenges in efficiency and accuracy when dealing with complex scenes and identifying diverse prohibited items. To solve this problem, the embodiment of the present application proposes a convolution-long short-term memory (CNN-LSTM) neural network item detection and identification method.

[0050] Please refer to the attached Figure 1 - attached Figure 5 The embodiment of the present application proposes a convolution-long short-term memory (CNN-LSTM) neural network item detection and identification method, which can be applied to millimeter wave security imaging systems in public places such as entry and exit, airports, and stations, to achieve efficient and accurate detection and identification of prohibited items. As shown in Figure 1 The item detection and identification method includes the following steps:

[0051] Step S1: Convert the current millimeter wave image data frame x(n) into M parallel memory-type millimeter wave image data frames through the input preprocessing layer, where M≥2.

[0052] Specifically, in step S1, the current millimeter wave image data frame is converted into M parallel memory-type millimeter wave image data frames using a time-delay neural network (TDNN). TDNN is used to convert a single data frame into a parallel data frame with M outputs. Each output data frame is a different representation of the original data, enhancing the parallelism and memory of the data and providing more perspectives for subsequent feature extraction. In simple terms, an original millimeter wave image is converted into M different image versions, each of which can provide different perspectives or angles of information about the original image, providing more diverse input data for subsequent analysis steps.

[0053] It can be understood that since the millimeter wave image may be blurred or complex, especially when multiple objects appear in the image at the same time, it is difficult to directly identify specific prohibited items from the image. Therefore, the purpose of step S1 of the present application is to increase the parallelism and memory of the data, providing a more abundant information base for subsequent feature extraction and identification to solve the above problems.

[0054] Step S2: Output the M parallel memory-type millimeter wave image data frames one by one to the M spatial attention convolution modules for processing to obtain M first weighted feature maps.

[0055] It can be understood that by respectively assigning the M parallel memory type millimeter wave image data frames to the M spatial attention convolution modules, each spatial attention convolution module is responsible for processing one frame of millimeter wave image data, that is, the spatial attention convolution module 1, the spatial attention convolution module 2, …, and the spatial attention convolution module M are respectively used to perform convolution processing on the M parallel memory type millimeter wave image data frames to extract the spatial feature information of the millimeter wave image in parallel, which can more efficiently and accurately process the image data.

[0056] Step S3: outputting the M first weighted feature maps to the M channel attention LSTM modules for processing to obtain M second weighted feature maps for subsequent detection and identification of prohibited articles. Figure One

[0057] It can be understood that by respectively assigning the M first weighted feature maps to the M channel attention LSTM modules, each channel attention LSTM module is responsible for processing one first weighted feature map, that is, the channel attention LSTM module 1, the channel attention LSTM module 2, …, and the channel attention LSTM module M are respectively used to process the M first weighted feature maps to further extract the spatial feature information of the millimeter wave image.

[0058] In one preferred embodiment of the present application, each spatial attention convolution module includes a first convolution layer, a first maximum pooling layer, a spatial attention mechanism module, and a first weighted fusion module.

[0059] In step S2, the specific method of one-to-one corresponding outputting the M parallel memory type millimeter wave image data frames to the M spatial attention convolution modules for processing to obtain the M first weighted feature maps includes:

[0060] Step S21: performing preliminary feature extraction on the input memory type millimeter wave image data frame by the first convolution layer to generate a first original feature map.

[0061] It can be understood that each spatial attention convolution module first performs preliminary feature extraction on the input millimeter wave image data frame by the first convolution layer, which generates a group of original feature maps. These feature maps contain local patterns and structural information in the input image, such as edges, textures, etc.

[0062] Step S22: processing the first original feature map by the first maximum pooling layer to reduce the amount of calculation while retaining important features.

[0063] ​It can be understood that the image size is reduced by the first max-pooling layer, the calculation amount is reduced while the important features are retained, this step does not change the content of the original feature map, only reduces the resolution, the first max-pooling layer is added to avoid overfitting and increase the generalization ability of the model.

[0064] Step S23: processing the first original feature map processed by the first max-pooling layer through the spatial attention mechanism module to generate a spatial attention map.

[0065] It can be understood that the spatial attention map reflects the importance of different regions in the image. The model can automatically learn and focus on more important regions in the image, thereby improving the accuracy of feature extraction.

[0066] Step S24: multiplying the spatial attention map and the first original feature map by the first weighted fusion module element by element to realize effective fusion of the features and generate a first weighted feature map.

[0067] It can be understood that by multiplying the spatial attention map and the original feature map element by element, the first weighted feature map is generated, which enhances the feature representation of the key region, so that the model can focus more on the part that may contain prohibited items.

[0068] In summary, the purpose of steps S21-S24 in the present application is to extract spatial features in the image, and to enhance important regions and weaken irrelevant parts through the spatial attention mechanism. Improve the quality of feature representation, so that subsequent processing focuses more on key information. Specifically, through the spatial attention mechanism, the model can automatically learn and focus on more important regions in the image, thereby improving the accuracy of feature extraction. Suppress irrelevant information, weaken irrelevant or interfering background information, so that the model focuses more on the target object. Improve the quality of feature representation, extract more representative spatial features through convolution and pooling operations, and provide high-quality input for subsequent processing.

[0069] In one preferred embodiment of the present application, each channel attention LSTM module includes a first LSTM unit, a second max-pooling layer, a channel attention mechanism module, and a second weighted fusion module.

[0070] Specifically, in step S3, M first weighted feature maps Figure One The specific method of outputting to M channel attention LSTM modules to obtain M second weighted feature maps includes:

[0071] Step S31: processing the input first weighted feature map through the first LSTM unit and the second max-pooling layer to further extract deep feature information and obtain a second original feature map.

[0072] It can be understood that the first LSTM unit can effectively handle the long-term dependency problem in sequence data, and the LSTM is used to capture the feature changes between different time points, which is particularly important for object detection in dynamic scenes, especially when objects move or deform in millimeter wave images. The second max-pooling layer reduces the size of the feature map, reducing the computational burden of subsequent processing while maintaining key features.

[0073] Step S32: processing the second original feature map through a channel attention mechanism module to generate a channel attention map.

[0074] It can be understood that the channel attention mechanism module calculates the average and maximum values of each channel in the entire feature map, and predicts the weight of each channel through a multi-layer perceptron (MLP). Then, these weights are applied to the second original feature map to generate a channel attention map, reflecting the importance of different channels for the current task.

[0075] The introduction of the channel attention mechanism has the following benefits: it enhances the model's understanding of the relationship between feature channels, thereby improving the accuracy of feature extraction. Specifically, it enhances important channels: the channel attention map can highlight those channels that are more important for the current task, thereby improving the relevance and effectiveness of feature representation. It suppresses irrelevant channels: it weakens channels that are less important for the current task, reducing the impact of noise.

[0076] Step S33: multiplying the channel attention map and the second original feature map output after the second max-pooling layer through a second weighted fusion module to achieve effective fusion of features and generate a second weighted feature map.

[0077] It can be understood that the purpose of step S33 is to adjust the intensity of each feature map according to the importance of the channel, thereby achieving effective feature fusion, and the final generated second weighted feature map contains not only the spatial information of the original feature map, but also the important channel information emphasized by the channel attention map. By combining the attention mechanisms in both spatial and channel dimensions, the generated second weighted feature map focuses more on the key features of the target object, improving the accuracy of detection and recognition. Since the attention mechanism can automatically adapt to different inputs, the model can still maintain high performance even in the presence of interference or noise.

[0078] Overall, the processing of steps S31-S33 not only further extracts deep feature information, but also optimizes the quality of the feature map, making the final generated second weighted feature map more suitable for the detection and identification of contraband items, effectively improving the accuracy and robustness of the system, providing strong support for practical applications.

[0079] In one preferred embodiment of the present embodiment, the article detection and recognition method further comprises the following steps:

[0080] Step S4: The M second weighted feature maps obtained by processing the M channel attention LSTM modules are cascaded by a cascading layer to form a weighted feature panorama covering M parallel memory-type millimeter wave image data frames.

[0081] It can be understood that the cascading layer cascades the M second weighted feature maps in a specific order (for example, according to the time order of the millimeter wave image data frames corresponding to them). The key of this step is to integrate the information of multiple feature maps into a larger feature map to form a weighted feature panorama covering M parallel memory-type millimeter wave image data frames. Cascading can be achieved by simple concatenation, or by more complex fusion strategies, such as learning a weight matrix to dynamically adjust the contribution of each feature map.

[0082] Step S5: The weighted feature panorama is input into a second LSTM unit and a third max-pooling layer for deep-level prohibited article feature information extraction, wherein the second LSTM unit is used to capture the time dependence in the feature sequence, and the third max-pooling layer is used to reduce the dimension of the data and retain the most important feature information.

[0083] It can be understood that the deep-level prohibited article feature information is extracted from the weighted feature panorama, which provides strong support for the final detection and recognition. After the information integration of the weighted feature panorama, the long-term dependence between these information can be better understood through the second LSTM unit. The weighted feature panorama processed by the second LSTM unit is again down-sampled by the third max-pooling layer, which helps to reduce the data volume while retaining the most significant features, so that the model can focus on the most important information, thereby improving the computational efficiency and model generalization ability.

[0084] In one preferred embodiment of the present embodiment, the article detection and recognition method further comprises the following steps:

[0085] Step S6: The extracted prohibited article feature information is input into a fully connected layer for feature fusion and classification decision.

[0086] Specifically, the fully connected layer includes at least two fully connected networks (a first fully connected layer and a second fully connected layer). The fully connected layer maps the feature information to a classification space through linear transformation and activation function, thereby realizing prohibited article detection and recognition, and finally outputs the feature classification result to determine whether there is a prohibited article and the type of the prohibited article.

[0087] It can be understood that the full connection layer can capture the complex relationship between the input features, effectively fuse the features from different sources or different levels through multi-layer linear transformation and nonlinear activation.

[0088] In one preferred embodiment of the present embodiment, the article detection and identification method further comprises the steps of marking the detected prohibited articles, visualizing the positioning and classification information, so as to facilitate the security personnel to make a quick judgment and response.

[0089] Specifically, the security personnel can be intuitively presented through the user interface to assist them to make a quick and accurate judgment.

[0090] In addition, the network structure parameters such as convolution layer, maximum pooling layer and full connection layer in the present application can be optimized and adjusted according to specific application scenarios to improve the detection accuracy and identification efficiency.

[0091] The CNN-LSTM neural network article detection and identification method proposed in the present application has significant technical advantages compared with the existing entry and exit site security millimeter wave imaging processing technology.

[0092] Firstly, in the data preprocessing stage, the present application converts the current millimeter wave image data frame into M parallel memory type millimeter wave image data frames through the time delay neural network, which effectively improves the data processing speed and parallel processing capability, and lays a solid foundation for subsequent efficient feature extraction.

[0093] Secondly, in the aspect of feature extraction, the present application combines the spatial attention mechanism and the channel attention mechanism, and through M spatial attention convolution modules and M channel attention LSTM modules, respectively processes the continuous M frame millimeter wave image data frames in parallel. The use of this double attention mechanism enables the system to focus more accurately on the key information in the image, effectively suppresses the interference of irrelevant information, and thus improves the accuracy and robustness of prohibited article identification.

[0094] In addition, the present application further extracts deep features through the LSTM layer and the maximum pooling layer, further excavates the deep feature information in the millimeter wave image, and this deep feature extraction method not only enhances the recognition ability of the system, but also improves the identification accuracy of prohibited articles in complex scenes.

[0095] Finally, the present application fuses and classifies the features of prohibited articles through the full connection layer, realizes intelligent prohibited article detection and discrimination, and this intelligent detection method not only improves the security efficiency, but also greatly reduces the work burden of security personnel, and provides strong support for improving the security efficiency and accuracy.

[0096] The above-described embodiments are only used to illustrate the technical solutions of the present application, but not limit them; although the present application is described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1.A method for item detection and recognition using a convolutional-long short-term memory neural network, the method comprising: The method comprises the following steps: Converting a current millimeter wave image data frame into M parallel memory millimeter wave image data frames, wherein M≥2; The M parallel memory millimeter wave image data frames are one-to-one output to M spatial attention convolution modules for processing to obtain M first weighted feature maps; each spatial attention convolution module comprises a first convolution layer, a first maximum pooling layer, a spatial attention mechanism module and a first weighted fusion module; The M first weighted feature maps are one-to-one output to M channel attention LSTM modules for processing to obtain M second weighted feature maps, which are used for subsequent contraband detection and identification; each channel attention LSTM module comprises a first LSTM unit, a second maximum pooling layer, a channel attention mechanism module and a second weighted fusion module; The specific method for processing the M parallel memory millimeter wave image data frames to obtain the M first weighted feature maps comprises: performing preliminary feature extraction on the input memory millimeter wave image data frame through the first convolution layer to generate a first original feature map; processing the first original feature map through the first maximum pooling layer; processing the first original feature map processed by the first maximum pooling layer through the spatial attention mechanism module to generate a spatial attention map; multiplying the spatial attention map and the first original feature map element by element through the first weighted fusion module to realize feature fusion and generate a first weighted feature map; The specific method for processing the M first weighted feature maps to obtain the M second weighted feature maps comprises: processing the input first weighted feature map through the first LSTM unit and the second maximum pooling layer to further extract deep feature information to obtain a second original feature map; processing the second original feature map through the channel attention mechanism module to generate a channel attention map; multiplying the channel attention map and the second original feature map output after being processed by the second maximum pooling layer element by element through the second weighted fusion module to realize feature fusion and generate a second weighted feature map; The M second weighted feature maps obtained by processing the M channel attention LSTM modules are cascaded through a cascaded layer to form a weighted feature panoramic map covering the M parallel memory millimeter wave image data frames; The weighted feature panoramic map is input into a second LSTM unit and a third maximum pooling layer for deep contraband feature information extraction, wherein the second LSTM unit is used to capture the time dependence in the feature sequence, and the third maximum pooling layer is used to reduce the dimension of the data and retain the most important feature information. 2.The method of claim 1, wherein, The article detection and identification method further comprises the following steps: inputting the extracted contraband feature information into a fully connected layer for feature fusion and classification decision. 3.The method of claim 2, wherein, The fully connected layer comprises at least two fully connected networks; the fully connected layer maps the feature information to a classification space through linear transformation and the action of an activation function, and finally outputs the feature classification result to determine whether there is a contraband and the type of the contraband. 4.The method of claim 3, wherein, The article detection and identification method further comprises the following steps: marking the detected prohibited articles, and visualizing display of positioning and classification information. 5.The method of claim 1, wherein, The conversion of the current millimeter wave image data frame into M parallel memory millimeter wave image data frames is implemented by using a time delay neural network.

Images