A Deep Learning-Based Automatic Recognition Method for Multi-View Security Inspection Images

Abstract

This invention discloses an automatic multi-view security inspection image recognition method based on deep learning. Bags containing a mixture of known dangerous and non-dangerous goods are passed through a security inspection device equipped with multi-view probes, generating security inspection images of each bag from multiple perspectives. These images are then labeled to create a dataset. A network model is trained using the training set and deployed in the security inspection device. During actual detection, the acquired multi-view security inspection images are input into the network model. The images undergo preprocessing, feature extraction, perspective transformation, and feature fusion within the network model. Finally, the fused features are input into a classifier for classification, determining whether the object to be detected contains dangerous goods. If dangerous goods are found, their category is determined. This invention solves the problems of missed detections due to blind spots in security inspection angles and the heavy reliance on manual inspection, as well as the related issues.

Description

Technical Field

[0001] This invention relates to improvements in security inspection image recognition technology, specifically to an automatic recognition method for multi-view security inspection images based on deep learning, belonging to the field of security inspection technology. Background Technology

[0002] With advancements in detection technology and increased public awareness of safety, security screening technology is widely used in various important locations and transportation hubs. Airports, train stations, government agencies, large event venues, and other public places all require effective security measures to prevent potential threats. The objects of security screening are usually people and their belongings. After bags, packages, and other items carried by people are sent onto a conveyor belt and passed through the inspection channel, X-ray images of the items are formed.

[0003] In existing technologies, security screening equipment captures images using one or two cameras fixed in the inspection channel. However, this method still has limitations; items may be placed in blind spots due to stacking or angled placement, leading to the misdetection or missed detection of some dangerous goods, thus impacting public transportation and the safety of people and property. Furthermore, most current security screening methods rely on manual inspection, where security personnel examine X-ray images one by one to manually determine the presence of contraband. Manual inspection has the following drawbacks: 1. It heavily relies on the experience of the inspectors, whose judgment and identification skills vary; moreover, prolonged high-intensity work can easily lead to eye fatigue and incorrect judgments. 2. During busy transportation periods and with high passenger flow, manual inspection may be overwhelmed, resulting in missed or false detections, thus affecting the efficiency and effectiveness of security checks. Summary of the Invention

[0004] To address the aforementioned shortcomings of existing technologies, the present invention aims to provide an automatic recognition method for multi-view security inspection images based on deep learning. This invention can solve the problems of false positives and false negatives caused by blind spots in security inspection angles and the heavy reliance on personnel for manual inspections, as well as the series of problems arising therefrom.

[0005] The technical solution of this invention is implemented as follows:

[0006] The deep learning-based method for automatic recognition of multi-view security inspection images proceeds as follows:

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

[0008] 2) Preprocess each security inspection image in the dataset to ensure that the security inspection images from each viewpoint are of consistent scale for subsequent processing;

[0009] 3) Input the security inspection images of the same bag from various perspectives in the training set that have been preprocessed in step 2) into the convolutional neural network after initialization. Use the convolutional neural network to extract features from the security inspection images from each perspective. The features contain local and global information of the security inspection images and are converted 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 projection of security inspection images from multiple perspectives. The query matrix uses three-dimensional grid points. The three-dimensional grid points of the query matrix are projected onto each two-dimensional view plane using coordinate transformation methods. The position of the grid points on the view plane is calculated. Within the area with the mapped value, the features of the security inspection images of the same bag corresponding to each perspective are queried.

[0011] 5) Feature fusion; fuse the features of the security inspection images from different perspectives corresponding to the same bag obtained in step 4);

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

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

[0014] 8) During actual testing, the object to be tested is passed through the security inspection equipment. The multi-view probe of the security inspection equipment acquires security inspection images of the object from multiple perspectives and inputs them into the multi-view security inspection image automatic recognition network model.

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

[0016] Furthermore, the parameters for initializing the convolutional neural network in step 3) are obtained through pre-training, which utilizes the ViT model. During pre-training, the loss function threshold and the maximum number of training iterations 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 as follows:

[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 perspectives;

[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, with the fully connected layer consisting of Z... x =(W x ) T h x It indicates that Z x h represents the output of the fully connected layer, i.e., the regression value for each detected category. x W represents the eigenvectors obtained after flattening the output matrix from step 5) into a one-dimensional array. x The weight matrix of the fully connected layer is represented as follows; the Softmax layer is represented as... Where K is the total number of detected categories, z i Z is the output matrix of the fully connected layer. x The value of the i-th category, p i It represents the probability that the detected sample vector z belongs to the i-th category.

[0023] Further, 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 convolutional layer, a residual module, and a global average pooling layer; the residual module contains multiple convolutional layers, and there is an identity mapping relationship between the layers, i.e., 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 upsampling unit that extracts features from the conv2, conv3, conv4, and conv5 layers of ResNet and outputs them through the convolutional layers.

[0024] Furthermore, in step 4), when establishing the query matrix, let the size of the query matrix be (H, W), the resolution be s, and the points of the query matrix be denoted as (x, y). While establishing the (x, y) coordinate positions, the z-axis is sampled to give the grid points depth information; in 3D space, the sampled points are represented as (x′, y′, z′x).

[0025] The expressions for sampling points x′ and y′ are as follows:

[0026] x′=(xW / 2)×s

[0027] y′=(yW / 2)×s

[0028] The sampling points in 3D space are mapped back to the 2D image plane, and features are queried in the regions with mapped values.

[0029] Furthermore, in step 1), the multi-view probes include a front-view probe, a top-view probe, and a side-view probe, and the corresponding security inspection images include a front-view security inspection image, a top-view security inspection image, and a side-view security inspection image.

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

[0031] Further, in step 7), achieving the set condition means that mAP@50 meets the set requirements; mAP@50 refers to the average accuracy obtained when the cross-union threshold is 50%.

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

[0033] 1. This invention enables the automatic identification of dangerous and prohibited items during security checks. The security inspection equipment, equipped with a multi-view security image automatic recognition model, automatically identifies whether the object to be inspected contains dangerous items, and if so, determines the category of the dangerous items. Compared to manual inspection, it eliminates the error judgments that can occur during prolonged high-intensity work, maintaining continuous reliability.

[0034] 2. This invention employs multi-view probes to acquire security inspection images. Each probe (camera) captures an image containing its unique perspective and spatial information, avoiding the impact of item stacking or placement angles on subsequent image recognition. After multi-angle acquisition, the items undergo feature extraction, perspective transformation, and feature fusion processing, and are finally recognized by a deep learning model, demonstrating high accuracy in identifying various hazardous materials. The average precision (P) and recall (R) for detecting all hazardous materials reach 89.2% and 83.6%, respectively, with mAP@50 and mAP@50:95 scores reaching 91.8% and 71.5%, respectively. Attached Figure Description

[0035] Figure 1 This is a flowchart illustrating the overall process of the automatic recognition method for multi-view security inspection images based on deep learning, as described in this invention.

[0036] Figure 2 This is a schematic diagram of the network structure of the automatic recognition method for multi-view security inspection images based on deep learning according to the present invention;

[0037] Figure 3 This is a schematic diagram illustrating the process of establishing a three-dimensional security inspection image dataset according to an embodiment of the present invention. Detailed Implementation

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

[0039] This invention provides an automatic recognition method for multi-view security inspection images based on deep learning, which reduces the errors and omissions caused by stacking and blind spots due to defects in the security inspection angles of existing technologies, as well as a series of shortcomings of manual security inspections. The overall security inspection process of this invention is as follows: Figure 1 As shown. The present invention is carried out according to the following steps (see also...). Figure 2 and Figure 3 ),

[0040] 1) Prepare the dataset; determine the types of hazardous materials to be detected, and pack all hazardous materials and non-hazardous ordinary items into separate bags. Each bag should contain a mixture of ordinary items and at least one hazardous material. Pass all bags through a security inspection device equipped with multi-view probes to obtain security inspection images (X-ray images) from multiple perspectives for each bag, and label them with hazardous material categories. This will form the dataset. Divide the dataset into a training set, a validation set, and a test set, with proportions of 70%, 15%, and 15%, respectively. Hazardous material categories mainly include hazardous chemicals (strong acids and alkalis, etc.), flammable liquids (alcohol, gasoline, etc.), firearms and ammunition, and controlled knives. Of course, this can be expanded according to actual detection needs.

[0041] To avoid the impact of stacked or angled items on subsequent image recognition, this invention employs a multi-view probe system, compared to the previous method using one or two cameras. The multi-view probes include a front-view probe, a top-view probe, and a side-view probe. The front-view probe faces the direction from which the item is coming to collect information from its front (but due to the continuity of security checks and the need for item passage, the front-view probe is not directly facing the item, but rather at a higher height, angled downwards towards the direction from which the item is coming). The top-view probe collects information from above the item, and the side-view probe collects information from the left or right side of the item. The corresponding security images include a front-view security image, a top-view security image, and a side-view security image. Each probe (camera) captures an image containing its unique perspective and spatial information.

[0042] 2) Preprocess each security inspection image in the dataset to ensure that the security inspection images from each viewpoint are of consistent scale and facilitate subsequent processing. During preprocessing, the multi-view security inspection images are cropped to an appropriate size, and the image data is enhanced through methods such as rotation, scaling, and normalization.

[0043] 3) Input the security inspection images of the same bag containing dangerous goods from various perspectives in the training set that has been preprocessed in step 2) into the initialized Convolutional Neural Network (CNN). Use the CNN to extract features from the security inspection images from each perspective. The features contain local and global information of the security inspection images and are converted into high-dimensional feature maps.

[0044] A convolutional neural network (CNN) model can be obtained by training and validating a selected CNN on a specific image dataset. In a preferred embodiment of this invention, step 3) uses a CNN model combining ResNet and FPN (Feature Pyramid Network); ResNet-50 is used as the CNN to extract multi-scale image features, and FPN is used to enhance the multi-scale feature extraction capability. The input image includes the front view, top view, and side view of the object to be detected. The input data includes the batch size (bs), the number of consecutive frames (Quene), the number of images in each frame (corresponding to the number of viewpoints, taken as 3 in this 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 contains multiple convolutional layers, and there is an identity mapping relationship between the layers, i.e., 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 convolutional layers.

[0045] 4) Viewpoint Transformation: Establish a query matrix whose size encompasses the entire inspection plane. The query matrix is ​​used to collect features and coordinate projections of multi-view security inspection images. The query matrix uses three-dimensional grid points. The three-dimensional grid points of the query matrix are projected onto each two-dimensional view plane using coordinate transformation methods. The positions of the grid points on the view plane are calculated. Within the regions with mapped values, the features of the security inspection images corresponding to the same bag at various viewpoints are queried.

[0046] When establishing the query matrix, let the size of the query matrix be (H, W) and the resolution be s. Let the points of the query matrix be denoted as (x, y). While establishing the (x, y) coordinate positions, the z-axis is sampled to give the grid points depth information. In 3D space, the sampled points are represented as (x′, y′, z′). i ).

[0047] The expressions for sampling points x′ and y′ are as follows:

[0048] x′=(xW / 2)×s

[0049] y′=(yW / 2)×s

[0050] The sampling points in 3D space are mapped back to the 2D image plane, and features are queried in the regions with mapped values.

[0051] 5) Feature fusion: Design a self-attention mechanism to fuse the features of security inspection images from various perspectives corresponding to the same bag obtained in step 4) by weighted summation, highlighting the features of important perspectives.

[0052] In a preferred embodiment of the present invention, the specific steps of feature fusion are as follows:

[0053] 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.

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

[0055] 5.3) Normalize the attention scores to obtain attention weights; in this example, the attention scores are converted into attention weights using the Softmax function.

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

[0057] 6) Input the features of each bag after fusion into the classifier for classification to determine whether it contains dangerous goods. If it contains dangerous goods, determine the category of dangerous goods and output it 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 is composed of Z... x =(W x ) T h x It indicates that Z x h represents the output of the fully connected layer, i.e., the regression value for each detected category. x W represents the eigenvectors obtained after flattening the output matrix from step 5) into a one-dimensional array. x The weight matrix of the fully connected layer is represented as follows; the Softmax layer is represented as... Where K is the total number of detected categories, z i Z is the output matrix of the fully connected layer. x The value of the i-th category, p i It represents the probability that the detected sample vector z belongs to the i-th category.

[0059] 7) Repeat steps 3)-6) on the other data in the training set after preprocessing in step 2) until the set conditions are met to obtain the multi-view security inspection image automatic recognition network model. Then deploy the multi-view security inspection image automatic recognition network model into security inspection equipment to provide multi-view X-ray image automatic analysis services.

[0060] 8) During actual testing, the object to be tested is passed through the security inspection equipment. The multi-view probe of the security inspection equipment collects security inspection images of the object from multiple perspectives (main view security inspection image, top view security inspection image, and side view security inspection image), and inputs them into the multi-view security inspection image automatic recognition network model.

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

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

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

[0064] The expression for precision is:

[0065]

[0066] The expression for recall rate is:

[0067]

[0068] The average precision (AP) of the predicted categories is expressed as:

[0069]

[0070] To measure the average precision across all classes, mAP is typically used to calculate the mean AP for each class, expressed as:

[0071]

[0072] mAP@50 refers to the average accuracy obtained when the cross-union ratio (CUN) threshold is 50%, while mAP@50:95 refers to the average accuracy calculated over a CUN threshold range of 50% to 95% with a step size of 5%.

[0073] Considering the large number of hazardous materials categories, the different requirements for the same indicator for each type of hazardous material, and the multidimensional nature of model evaluation indicators, this invention actually uses mAP@50 as the model deployment setting condition in step 7). In actual implementation, a comprehensive balance can be struck between detection accuracy and the efficiency of security inspection items; for example, mAP@50 can be set to no less than 82%.

[0074] Typically, as the number of training epochs increases, the loss for bounding box prediction, object confidence prediction, and category prediction all tend to decrease, while precision and recall tend to increase. Furthermore, in the latter half of training, the decrease in loss and the increase in precision and recall gradually slow down until they fluctuate within a certain range, indicating that the model is nearing the end of its fit. If the loss increases while precision and recall decrease, it indicates that the model is overfitting to some extent, requiring an appropriate reduction in the number of training epochs or network optimization.

[0075] Experiments show that the average precision (P) and average recall (R) for all hazardous materials detection in this embodiment reach 89.2% and 83.6%, respectively, with mAP@50 and mAP@50:95 scores reaching 91.8% and 71.5%, respectively. Specifically, the mAP@50 score exceeds 97% for firearms, 91.3% for pressure vessels, 93.7% for plastic bottles, 86.4% for rifle and pistol ammunition, and even 82.4% for knives, which have a lower detection rate. This demonstrates that the multi-view security inspection image automatic recognition method based on deep learning proposed in this invention achieves automatic recognition, significantly reducing the burden of manual inspection, and also boasts high accuracy, making it fully applicable to practical security inspections.

[0076] Finally, it should be noted that the above examples of the present invention are merely illustrative and not intended to limit the implementation of the invention. Although the applicant has described the present invention in detail with reference to preferred embodiments, those skilled in the art can make other variations and modifications based on the above description. It is impossible to exhaustively list all possible implementations here. All obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for automatic recognition of multi-view security inspection images based on deep learning, characterized in that: Follow these steps: 1) Prepare the dataset; determine the types of dangerous goods to be detected, pack all the dangerous goods to be detected and ordinary items that are not dangerous goods into bags, and each bag contains ordinary items and at least one dangerous item; pass all bags through security inspection equipment equipped with multi-view probes to obtain security inspection images of all bags from multiple perspectives and label them to obtain the dataset. The dataset is divided into a training set, a validation set, and a test set; 2) Preprocess each security inspection image in the dataset to ensure that the security inspection images from each viewpoint are of consistent scale for subsequent processing; 3) Input the security inspection images of the same bag from various perspectives in the training set that have been preprocessed in step 2) into the convolutional neural network after initialization. Use the convolutional neural network to extract features from the security inspection images from each perspective. The features contain local and global information of the security inspection images and are converted into high-dimensional feature maps. 4) Establish a query matrix whose size encompasses the entire inspection plane. The query matrix is ​​used to collect features and coordinate projections of multi-view security inspection images. The query matrix uses three-dimensional grid points. The three-dimensional grid points of the query matrix are projected onto each two-dimensional view plane using a coordinate transformation method. The position of the grid points on the view plane is calculated. Within the area with mapped values, the features of the security inspection images of the same bag from various perspectives are queried. 5) Feature fusion; fuse the features of the security inspection images from different perspectives corresponding to the same bag obtained in step 4); 6) Input the merged features of each bag into the classifier for classification to determine whether it contains dangerous goods. If it does, determine the category of dangerous goods. 7) Repeat steps 3)-6) on the other data in the training set after preprocessing in step 2) until the set conditions are met to obtain the multi-view security inspection image automatic recognition network model; and deploy the multi-view security inspection image automatic recognition network model into the security inspection equipment. 8) During actual testing, the object to be tested is passed through the security inspection equipment. The multi-view probe of the security inspection equipment acquires security inspection images of the object from multiple perspectives and inputs them into the multi-view security inspection image automatic recognition network model. 9) The multi-view security inspection image automatic recognition network model preprocesses each security inspection image according to step 2), and then performs steps 3)-6), that is, to obtain whether the object to be detected contains dangerous goods. If there are dangerous goods, the category of dangerous goods is determined.

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

3. The multi-view security inspection image automatic recognition method based on deep learning according to claim 1, characterized in that: The feature fusion step in step 5) is as follows: 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 perspectives; 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 method for automatic recognition of multi-view security inspection images based on deep learning according to claim 1, characterized in that: In step 6), the classifier consists of a fully connected layer and a Softmax layer. The fully connected layer is composed of Z... x =(W x ) T h x It indicates that Z x h represents the output of the fully connected layer, i.e., the regression value for each detected category. x W represents the eigenvectors obtained after flattening the output matrix from step 5) into a one-dimensional array. x The weight matrix of the fully connected layer is represented as follows; the Softmax layer is represented as... Where K is the total number of detected categories, z i Z is the output matrix of the fully connected layer. x The value of the i-th category, p i It represents the probability that the detected sample vector z belongs to the i-th category.

5. The method for automatic recognition of multi-view security inspection images based on deep learning according to claim 1, characterized in that: Step 3) employs a convolutional neural network model combining ResNet and FPN. 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 convolutional layer, a residual module, and a global average pooling layer. The residual module contains multiple convolutional layers, and there is an identity mapping relationship between the layers, i.e., 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 upsampling unit that extracts features from the conv2, conv3, conv4, and conv5 layers of ResNet and outputs them through the convolutional layers.

6. The method for automatic recognition of multi-view security inspection images based on deep learning according to claim 1, characterized in that: In step 4), when establishing the query matrix, let the size of the query matrix be (H, W) and the resolution be s. Let the points of the query matrix be (x, y). While establishing the (x, y) coordinate positions, the z-axis is sampled to give the grid points depth information. In 3D space, the sampled points are represented as (x′, y′, z′). i ) The expressions for sampling points x′ and y′ are as follows: x′=(xW / 2)×s y′=(yW / 2)×s The sampling points in 3D space are mapped back to the 2D image plane, and features are queried in the regions with mapped values.

7. The method for automatic recognition of multi-view security inspection images based on deep learning according to claim 1, characterized in that: In step 1), the multi-view probes include a front-view probe, a top-view probe, and a side-view probe, and the corresponding security inspection images include a front-view security inspection image, a top-view security inspection image, and a side-view security inspection image.

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

9. The method for automatic recognition of multi-view security inspection images based on deep learning according to claim 1, characterized in that: In step 7), meeting the set conditions means that mAP@50 meets the set requirements; mAP@50 refers to the average accuracy obtained when the cross-union threshold is 50%.

Images