Multi-view security check image automatic identification method based on deep learning

Abstract

The invention discloses a multi-view security check image automatic identification method based on deep learning, and the method comprises the steps: enabling bags and suitcases in which various known dangerous articles and non-dangerous articles are mixed to pass through security check equipment equipped with a multi-view probe, obtaining the security check images of each bag and carrying out the marking of the security check images, and obtaining a data set; and training the network model through the training set and deploying the network model in security inspection equipment. During actual detection, collected multi-view security check images are input into the network model, the images are sequentially subjected to preprocessing, feature extraction, view angle transformation and feature fusion processing in the network model, and finally fused features are input into the classifier for classification, so that whether an object to be detected contains dangerous goods or not is obtained, and if the object to be detected contains the dangerous goods, the category of the dangerous goods is judged. According to the invention, a series of problems caused by wrong detection and missing detection due to the existence of a detection blind area in a security check angle and large dependence of manual detection on personnel in the prior art can be solved.

Description

Technical Field

[0001] The present invention relates to an improvement in security inspection image recognition technology, and in particular to a multi-view security inspection image automatic recognition method based on deep learning, which belongs to the field of security inspection technology. Background Art

[0002] With the advancement of detection technology and the improvement of public safety awareness, security inspection technology is widely used in various important places and transportation hubs. Airports, stations, government agencies, large-scale event venues and other public places all need effective security inspection measures to prevent potential threats. The objects of security inspection are usually people and their belongings. After the luggage, parcels and other items carried by people are sent to the conveyor belt and passed through the inspection channel, an X-ray image of the security inspection items will be formed.

[0003] In the prior art, the imaging of security inspection equipment is performed by collecting images through one or two cameras fixed in the inspection channel. In this case, there are still cases where the items are in the camera detection blind spot due to stacking or placement angles, resulting in the misdetection of some dangerous goods, thus affecting public transportation and the safety of life and property of the people. In addition, most of the existing security inspection modes are manual inspections, that is, security inspectors check the X-ray imaging images one by one and manually determine whether there are prohibited items. The manual inspection method has the following shortcomings: 1. It relies heavily on the experience of the inspectors, and the judgment and recognition levels of different inspectors are different; and when manual work is carried out for a long time and at a high intensity, it is easy to cause eye fatigue and lead to wrong judgment. 2. When transportation is busy and the flow of people is large, manual inspections may be too busy, resulting in missed inspections or false inspections, thereby affecting the efficiency and effectiveness of security inspections. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the object of the present invention is to provide a multi-perspective security inspection image automatic recognition method based on deep learning. The present invention can solve the series of problems caused by the false detection and missed detection due to the presence of blind spots in the security inspection angles in the prior art, as well as the heavy reliance on personnel for manual inspection and the resulting consequences.

[0005] The technical solution of the present invention is achieved in this way:

[0006] The multi-view security inspection image automatic recognition method based on deep learning is carried out in the following steps:

[0007] 1) Prepare the data set; determine the various types of dangerous goods to be detected, and pack all the dangerous goods to be detected and ordinary items that are not dangerous goods into bags, so that each bag contains a mixture of ordinary items and at least one dangerous goods; pass all bags through the security inspection equipment equipped with multi-viewing probes, obtain security inspection images of all bags from multiple perspectives and annotate them, thereby obtaining the data set; divide the data set into training set, validation set and test set;

[0008] 2) Preprocess each security inspection image in the data set to make the scale of the security inspection images at each perspective consistent for subsequent processing;

[0009] 3) Inputting the security inspection images of each viewpoint corresponding to the same bag in the training set preprocessed in step 2) into the initialized convolutional neural network, and using the convolutional neural network to extract features from the security inspection images of each viewpoint, the features contain local and global information of the security inspection images, and converting them into high-dimensional feature maps;

[0010] 4) Establish a query matrix whose size includes the entire inspection plane. The query matrix is ​​used to collect the features and coordinate projections of the multi-view security inspection images. The query matrix uses three-dimensional grid points, and the three-dimensional grid points of the query matrix are projected to each two-dimensional view plane using the coordinate change method. The position of the grid points on the view plane is calculated, and the features of the security inspection images of each view corresponding to the same bag are queried in the area with the mapping value.

[0011] 5) Feature fusion: Fusion of features of security inspection images of different viewing angles corresponding to the same bag obtained in step 4);

[0012] 6) Input the fused features of each bag into the classifier for classification to determine whether it contains dangerous goods. If so, determine the category of dangerous goods;

[0013] 7) Repeat steps 3)-6) for other data in the training set preprocessed in step 2) until the set conditions are met to obtain a multi-view security inspection image automatic recognition network model; and deploy the multi-view security inspection image automatic recognition network model to the security inspection equipment;

[0014] 8) During the actual inspection, the object to be inspected passes through the security inspection equipment, and the multi-viewing probe of the security inspection equipment collects security inspection images of the object to be inspected from multiple perspectives, and inputs the multi-viewing security inspection image automatic recognition network model;

[0015] 9) The multi-view security inspection image automatic recognition network model pre-processes each security inspection image according to step 2), and then performs steps 3)-6) to obtain whether the object to be inspected contains dangerous goods. If so, the category of the dangerous goods is determined.

[0016] Furthermore, the parameters of the initialization setting of the convolutional neural network in step 3) are obtained through pre-training, and the pre-training is performed using the ViT model; during pre-training, the loss function threshold and the maximum number of training times are set, and the training parameters are initialized; the initialization parameters of the convolutional neural network are obtained through pre-training.

[0017] Furthermore, the feature fusion step in step 5) is:

[0018] 5.1) Establish a key matrix and a value matrix. The key matrix represents the importance of other perspectives to the current perspective, and the value matrix represents the eigenvalues ​​of each perspective.

[0019] 5.2) Calculate the dot product of the query matrix and the key matrix to obtain the attention score representing the correlation between different views;

[0020] 5.3) Normalize the attention scores to obtain the attention weights;

[0021] 5.4) Multiply the attention weights and the value matrix to obtain the weighted feature representation.

[0022] Furthermore, in step 6), the classifier consists of a fully connected layer and a Softmax layer. The fully connected layer consists of Z x =(W x ) T h x Indicates that Z x represents the output of the fully connected layer, i.e., the regression value of each detection category, h x represents the eigenvector obtained after the output matrix of step 5) is flattened into a one-dimensional array, W x represents the weight matrix of the fully connected layer; the Softmax layer is represented by Where K is the total number of detection categories, z i is the fully connected layer output matrix Z x The value of the i-th category in i is the probability that the detection sample vector z belongs to the i-th category.

[0023] Furthermore, in step 3), a convolutional neural network model combining ResNet and FPN is adopted; the input data includes batch size, number of consecutive frames, number of images in each frame, number of image channels, image length, and image width; ResNet includes an initial convolution layer, a residual module, and a global average pooling layer; the residual module includes multiple convolution layers, and there is an identical mapping relationship between layers, that is, F(x)=H(x)+x, where x represents input, F(x) represents output of the residual module, and H(x) represents output of the convolution layer; FPN includes an upsampler, which extracts features from conv2, conv3, conv4, and conv5 layers of ResNet and outputs them through the convolution layer.

[0024] Furthermore, in step 4), when establishing the query matrix, the query matrix size is set to (H, W), the resolution is s, and the points of the query matrix are recorded as (x, y). While establishing the (x, y) coordinate position, the z axis is sampled so that the grid point has depth information; in the 3D space, the sampling point is represented as (x′, y′, z′x)

[0025] The expressions of sampling points x′ and y′ are:

[0026] x′=(xW / 2)×s

[0027] y′=(yW / 2)×s

[0028] Map the sample points in 3D space back to the 2D image plane and query features in the area with mapped values.

[0029] Furthermore, in step 1), the multi-viewing probe includes a main-viewing probe, a downward-viewing probe and a side-viewing probe, and the corresponding security inspection images include a main-viewing security inspection image, a downward-viewing security inspection image and a side-viewing security inspection image.

[0030] Furthermore, in step 2), the preprocessing includes cropping, rotation, scaling and normalization.

[0031] Furthermore, in step 7), achieving the set condition means that mAP@50 meets the set requirement; mAP@50 refers to the average precision obtained when the intersection-over-union ratio threshold is 50%.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] 1. The present invention realizes the automatic identification of dangerous goods and contraband during security inspection. The objects to be inspected are automatically identified by the security inspection equipment with multi-view security inspection image automatic recognition model, and the category of dangerous goods is determined if there are dangerous goods. Compared with manual inspection, there is no misjudgment caused by long-term high-intensity work, and continuous reliability can be maintained.

[0034] 2. The present invention uses multi-view probes to collect security inspection images. The images captured by each probe (camera) contain its unique view and spatial information, which prevents the stacking or placement angle of objects from affecting the later recognition of the collected images. After the objects are collected from multiple angles, they are processed through feature extraction, view transformation, feature fusion, etc., and finally recognized by a deep learning model, which has a high recognition accuracy for various types of dangerous goods. The average precision (P) and recall rate (R) for all dangerous goods detection can reach 89.2% and 83.6% respectively, and the mAP@50 and mAP@50:95 scores can reach 91.8% and 71.5%. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 The figure is a general flow chart of the automatic recognition method of multi-view security inspection images based on deep learning of the present invention;

[0036] Figure 2 A schematic diagram of the network structure of the automatic recognition method of multi-view security inspection images based on deep learning of the present invention;

[0037] Figure 3 A schematic diagram of a process for establishing a three-dimensional security inspection image data set according to an embodiment of the present invention. DETAILED DESCRIPTION

[0038] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0039] The present invention provides an automatic recognition method for multi-view security inspection images based on deep learning, which is used to reduce the series of deficiencies in the existing technology such as stacking and visual blind spots caused by defects in security inspection angles, as well as the shortcomings of manual security inspection. The overall security inspection process of the present invention is as follows: Figure 1 The present invention is specifically carried out according to the following steps (see also Figure 2 and Figure 3 ),

[0040] 1) Prepare the data set; determine the various types of dangerous goods to be tested, and pack all the dangerous goods to be tested and ordinary items that are not dangerous goods into bags, each of which contains a mixture of ordinary items and at least one dangerous goods; pass all bags through security inspection equipment equipped with multi-view probes, obtain security inspection images (X-ray images) of all bags from multiple viewpoints and label them with dangerous goods categories, thereby obtaining the data set; divide the data set into training set, validation set and test set, with the proportions of 70%, 15% and 15% respectively. Dangerous goods categories mainly include dangerous chemicals (strong acids and alkalis, etc.), flammable liquids (alcohol, gasoline, etc.), firearms and ammunition, controlled knives, etc. Of course, it can be expanded according to actual detection needs.

[0041] In order to avoid the stacking or placement angle of objects affecting the later recognition of the collected images, compared with the original one or two cameras, the present invention adopts multi-view probe recognition, which includes a main view probe, a downward view probe and a side view probe. The main view probe faces the direction of the object to collect the front information of the object (in view of the continuity of security inspection and the need for the passage of objects, the main view probe is not facing the object from the front, but is at a higher height than the front, and faces the direction of the object from an angle downward), the downward view probe collects the information above the object from top to bottom, and the side view probe collects the side information of the object from the left or right side. The corresponding security inspection images include the main view security inspection image, the downward view security inspection image and the side view security inspection image. The image captured by each probe (camera) contains its unique perspective and spatial information.

[0042] 2) Preprocess each security inspection image in the data set to make the scale of the security inspection images of each perspective consistent and convenient for subsequent processing. During preprocessing, the multi-view security inspection images are cropped to a suitable size, and the image data is enhanced by methods such as rotation, scaling, and normalization.

[0043] 3) The security inspection images of each viewpoint corresponding to the same bag containing dangerous goods in the preprocessed training set in step 2) are input into the initialized convolutional neural network (CNN), and the convolutional neural network is used to extract features from the security inspection images of each viewpoint, the features containing local and global information of the security inspection images, and convert them into high-dimensional feature maps.

[0044] The convolutional neural network model can be obtained by training and verifying a specific image data set using a selected convolutional neural network. As a preferred embodiment of the present invention, a convolutional neural network model combining ResNet and FPN (Feature Pyramid Network) is used in step 3); ResNet-50 is used as a convolutional neural network to extract multi-scale image features, and FPN is used to enhance the multi-scale feature extraction capability. The input image includes the main view, top view and side view of the object to be detected, and the input data includes the batch size bs (BatchSize), the number of consecutive frames (Quene), the number of images in each frame (corresponding to the number of viewing angles, 3 in the embodiment), the number of image channels n, the image length h, and the image width w. ResNet includes an initial convolutional layer, a residual module, and a global average pooling layer; the residual module includes multiple convolutional layers, and there is an identical mapping relationship between the layers, that is, F(x) = H(x) + x, where x represents the input, F(x) represents the output of the residual module, and H(x) represents the output of the convolutional layer. FPN includes an upsampler that extracts features from the conv2, conv3, conv4, and conv5 layers of ResNet and outputs them through the convolutional layer.

[0045] 4) Perspective transformation: Establish a query matrix (Query) whose size covers the entire inspection plane. The query matrix is ​​used to collect the features and coordinate projections of multi-view security inspection images. The query matrix uses three-dimensional grid points, and the three-dimensional grid points of the query matrix are projected to each two-dimensional view plane using the coordinate transformation method. The position of the grid points on the view plane is calculated, and the features of the security inspection images of each perspective corresponding to the same bag are queried in the area with the mapping value.

[0046] When establishing the query matrix, the query matrix size is (H, W), the resolution is s, and the query matrix points are recorded as (x, y). While establishing the (x, y) coordinate position, the z axis is sampled so that the grid points have depth information; in 3D space, the sampling points are represented as (x′, y′, z′ i ).

[0047] The expressions of sampling points x′ and y′ are:

[0048] x′=(xW / 2)×s

[0049] y′=(yW / 2)×s

[0050] Map the sample points in 3D space back to the 2D image plane and query features in the area with mapped values.

[0051] 5) Feature fusion: Design a self-attention mechanism to fuse the features of the security inspection images of the same bag from different perspectives obtained in step 4) through the weighted summation method to highlight the features of important perspectives.

[0052] As a preferred implementation of the present invention, the specific steps of feature fusion are:

[0053] 5.1) Establish a key matrix (Key) and a value matrix (Value). The key matrix represents the importance of other perspectives to the current perspective, and the value matrix represents the eigenvalues ​​of each perspective;

[0054] 5.2) Calculate the dot product of the query matrix and the key matrix to obtain the attention scores that represent the correlation between different perspectives;

[0055] 5.3) The attention scores are normalized to obtain attention weights (AttentionWeights); in the embodiment, the attention scores are converted into attention weights through the Softmax function;

[0056] 5.4) Multiply the attention weights and the value matrix to obtain the weighted feature representation.

[0057] 6) The fused features of each bag are input into the classifier for classification to determine whether it contains dangerous goods. If there are dangerous goods, the category of the dangerous goods is determined and output to the interactive interface for display.

[0058] In step 6), the classifier consists of a fully connected layer and a Softmax layer. The fully connected layer consists of Z x =(W x ) T h x Indicates that Z x represents the output of the fully connected layer, i.e., the regression value of each detection category, h x represents the eigenvector obtained after the output matrix of step 5) is flattened into a one-dimensional array, W x represents the weight matrix of the fully connected layer; the Softmax layer is represented by Where K is the total number of detection categories, z i is the fully connected layer output matrix Z x The value of the i-th category in i is the probability that the detection sample vector z belongs to the i-th category.

[0059] 7) Repeat steps 3)-6) for other data in the training set preprocessed in step 2) until the set conditions are met to obtain a multi-view security inspection image automatic recognition network model. The multi-view security inspection image automatic recognition network model is deployed to the security inspection equipment to provide multi-view X-ray image automatic analysis services.

[0060] 8) During actual inspection, the object to be inspected passes through the security inspection equipment, and the multi-viewing probe of the security inspection equipment collects multiple-viewing security inspection images of the object to be inspected (main-viewing security inspection image, overhead-viewing security inspection image, and measuring-viewing security inspection image), and inputs them into the multi-viewing security inspection image automatic recognition network model.

[0061] 9) The multi-view security inspection image automatic recognition network model pre-processes each security inspection image according to step 2), and then performs steps 3)-6) to obtain whether the object to be inspected contains dangerous goods. If so, the category of the dangerous goods is determined.

[0062] In step 3), the parameters for initializing the convolutional neural network are obtained through pre-training, and the pre-training is performed using the ViT model; during pre-training, the loss function threshold and the maximum number of training times are set, and the training parameters are initialized; the convolutional neural network initialization parameters are obtained through pre-training.

[0063] In the specific implementation, the model evaluation uses a confusion matrix to present the classification results, and defines the confusion matrix as a matrix of size S*S, where S is the number of target detection categories. Each row of the confusion matrix represents the correct category of the detection object including the background, and each column represents the category predicted by the model including the background. For the output results of the target detection algorithm, four expressions are defined: (1) TP: The true category of the target and the result predicted by the model are both True. (2) TN: The true category of the target and the result predicted by the model are both Negative. (3) FP: The true category of the target is Negative, and the result predicted by the model is True. (4) FN: The true category of the target is True, and the result predicted by the model is Negative. The present invention evaluates the effect of model prediction through precision P, recall rate R, mean average precision mAP@50 and mAP@50:95.

[0064] The expression for accuracy is:

[0065]

[0066] The expression of recall is:

[0067]

[0068] The average precision AP of the predicted category is expressed as:

[0069]

[0070] To measure the average accuracy of all categories, mAP is generally used to average the AP of each category, expressed as:

[0071]

[0072] mAP@50 refers to the average precision obtained when the intersection-over-union ratio threshold is 50%, and mAP@50:95 refers to the average precision calculated within the intersection-over-union ratio threshold range from 50% to 95% with a step size of 5%.

[0073] Considering that there are many types of dangerous goods, the requirements for the same indicator for each type of dangerous goods are different, and the model evaluation indicators are multi-dimensional, based on this, the present invention actually uses mAP@50 as the setting condition for model deployment in step 7). In actual implementation, a comprehensive balance can be made based on the detection accuracy and the efficiency of security inspection items passing through, for example, mAP@50 can be set to no less than 82%.

[0074] Usually, as the number of training rounds increases, the loss of annotation box prediction, object confidence prediction loss, and category prediction loss all show a downward trend, while precision and recall rates all show an upward trend. When the training reaches the second half, the decrease in loss and the increase in precision and recall rates gradually slow down until they fluctuate within a range, which means that the model fitting is approaching completion. If the loss increases and the precision and recall rates decrease, it means that the model has a certain degree of overfitting, and the number of training rounds needs to be appropriately reduced or the network needs to be optimized.

[0075] Experiments show that the average precision (P) and average recall (R) of all dangerous goods detection in this embodiment can reach 89.2% and 83.6% respectively, and the mAP@50 and mAP@50:95 scores can reach 91.8% and 71.5%. Among them, the mAP@50 score of gun detection exceeds 97%, the mAP@50 score of pressure vessels is 91.3%, the mAP@50 score of plastic bottles is 93.7%, the detection score of rifles and pistol bullets is 86.4%, and the knives with lower detection effect also have a score of 82.4%. This shows that the automatic recognition method of multi-view security inspection images based on deep learning proposed by the present invention, on the one hand, realizes automatic recognition, greatly reducing the pressure of manual detection, and on the other hand, the accuracy of automatic recognition is high, which can be fully used for actual security inspection.

[0076] Finally, it should be noted that the above examples of the present invention are merely examples for illustrating the present invention, and are not intended to limit the embodiments of the present invention. Although the applicant has described the present invention in detail with reference to the preferred embodiments, for those of ordinary skill in the art, other different forms of changes and modifications can be made based on the above description. It is impossible to list all the embodiments here. Any obvious changes or modifications derived from the technical solution of the present invention are still within the scope of protection of the present invention.

Claims

1. A multi-view security inspection image automatic recognition method based on deep learning, characterized by: Follow the steps below, 1) Prepare a data set; determine the various types of dangerous goods to be tested, and pack all the dangerous goods to be tested and ordinary items that are not dangerous goods into bags, so that each bag contains a mixture of ordinary items and at least one dangerous goods; pass all bags through security inspection equipment equipped with multi-viewing probes, obtain security inspection images of all bags from multiple viewpoints and annotate them, thereby obtaining a data set; Divide the dataset into training, validation and test sets; 2) Preprocess each security inspection image in the data set to make the scale of the security inspection images at each perspective consistent for subsequent processing; 3) Inputting the security inspection images of each viewpoint corresponding to the same bag in the training set preprocessed in step 2) into the initialized convolutional neural network, and using the convolutional neural network to extract features from the security inspection images of each viewpoint, the features contain local and global information of the security inspection images, and converting them into high-dimensional feature maps; 4) Establish a query matrix whose size covers the entire inspection plane. The query matrix is ​​used to collect the features and coordinate projections of multi-view security inspection images; The query matrix uses three-dimensional grid points, which are projected onto each two-dimensional view plane using the coordinate transformation method. The positions of the grid points on the view plane are calculated, and the features of the security inspection images of each view angle corresponding to the same bag are queried in the area with the mapping value. 5) Feature fusion: Fusion of features of security inspection images of different viewing angles corresponding to the same bag obtained in step 4); 6) Input the fused features of each bag into the classifier for classification to determine whether it contains dangerous goods. If so, determine the category of dangerous goods; 7) Repeat steps 3)-6) for other data in the training set preprocessed in step 2) until the set conditions are met to obtain a multi-view security inspection image automatic recognition network model; and deploy the multi-view security inspection image automatic recognition network model to the security inspection equipment; 8) During the actual inspection, the object to be inspected passes through the security inspection equipment, and the multi-viewing probe of the security inspection equipment collects security inspection images of the object to be inspected from multiple perspectives, and inputs the multi-viewing security inspection image automatic recognition network model; 9) The multi-view security inspection image automatic recognition network model pre-processes each security inspection image according to step 2), and then performs steps 3)-6) to obtain whether the object to be inspected contains dangerous goods. If so, the category of the dangerous goods is determined.

2. The multi-view security inspection image automatic recognition method based on deep learning according to claim 1 is characterized by: In step 3), the parameters for initializing the convolutional neural network are obtained through pre-training, and the pre-training is performed using the ViT model; during pre-training, the loss function threshold and the maximum number of training times are set, and the training parameters are initialized; the convolutional neural network initialization parameters are obtained through pre-training.

3. The multi-view security inspection image automatic recognition method based on deep learning according to claim 1 is characterized by: The steps of feature fusion in step 5) are: 5.1) Establish a key matrix and a value matrix. The key matrix represents the importance of other perspectives to the current perspective, and the value matrix represents the eigenvalues ​​of each perspective. 5.2) Calculate the dot product of the query matrix and the key matrix to obtain the attention score representing the correlation between different views; 5.3) Normalize the attention scores to obtain the attention weights; 5.4) Multiply the attention weights and the value matrix to obtain the weighted feature representation.

4. The multi-view security inspection image automatic recognition method based on deep learning according to claim 1 is characterized by: In step 6), the classifier consists of a fully connected layer and a Softmax layer. The fully connected layer consists of Z x =(W x ) T h x Indicates that Z x represents the output of the fully connected layer, i.e., the regression value of each detection category, h x represents the eigenvector obtained after the output matrix of step 5) is flattened into a one-dimensional array, W x represents the weight matrix of the fully connected layer; the Softmax layer is represented by Where K is the total number of detection categories, z i is the fully connected layer output matrix Z x The value of the i-th category in i is the probability that the detection sample vector z belongs to the i-th category.

5. The multi-view security inspection image automatic recognition method based on deep learning according to claim 1 is characterized by: In step 3), a convolutional neural network model combining ResNet and FPN is used; the input data includes batch size, number of consecutive frames, number of images in each frame, number of image channels, image length, and image width; ResNet includes an initial convolution layer, a residual module, and a global average pooling layer; the residual module contains multiple convolution layers, and there is an identical mapping relationship between layers, that is, F(x)=H(x)+x, where x represents input, F(x) represents output of the residual module, and H(x) represents output of the convolution layer; FPN includes an upsampler that extracts features from conv2, conv3, conv4, and conv5 layers of ResNet and outputs them through the convolution layer.

6. The multi-view security inspection image automatic recognition method based on deep learning according to claim 1 is characterized by: In step 4), when establishing the query matrix, the query matrix size is set to (H, W), the resolution is s, and the query matrix points are recorded as (x, y). While establishing the (x, y) coordinate position, the z axis is sampled so that the grid points have depth information; in 3D space, the sampling points are represented as (x′, y′, z′ i ) The expressions of sampling points x′ and y′ are: x′=(xW / 2)×s y′=(yW / 2)×s Map the sample points in 3D space back to the 2D image plane and query features in the area with mapped values.

7. The multi-view security inspection image automatic recognition method based on deep learning according to claim 1 is characterized by: In step 1), the multi-viewing probes include a main-viewing probe, a downward-viewing probe and a side-viewing probe, and the corresponding security inspection images include a main-viewing security inspection image, a downward-viewing security inspection image and a side-viewing security inspection image.

8. The multi-view security inspection image automatic recognition method based on deep learning according to claim 1 is characterized by: In step 2), the preprocessing includes cropping, rotation, scaling and normalization.

9. The multi-view security inspection image automatic recognition method based on deep learning according to claim 1 is characterized by: In step 7), achieving the set condition means that mAP@50 meets the set requirement; mAP@50 refers to the average precision obtained when the intersection-over-union ratio threshold is 50%.

Images