A method and device for detecting prohibited items in security inspection images

Abstract

The present invention discloses a method and device for detecting prohibited items in security inspection images. The method includes the following steps: obtaining a security inspection image to be detected; inputting the security inspection image to be detected into a pre-trained multi-feature fusion model based on target object edge information aggregation to obtain a feature map with enhanced boundary features for target object prediction of the image. The feature map with enhanced boundary features obtained by using the method provided by the present invention is rich in features and precise in boundary, so as to facilitate the detector to detect prohibited items in luggage and effectively solve the problems of blurred and distorted boundaries of target objects in security inspection images taken by X-rays, monotonous image colors and lack of luster.

Description

Technical Field

[0001] The present invention relates to a method for detecting prohibited items in security inspection images, and also relates to a corresponding prohibited item detection device, belonging to the technical field of security inspection. Background Art

[0002] With the increase in the flow of people in public transportation, security inspection has become increasingly important. In the security inspection scenario, an X-ray scanner is usually used to check whether there are prohibited items in the luggage, and security inspectors need to concentrate on checking for prohibited items for a long time. After viewing complex security inspection images with high concentration for a long time, security inspectors may experience visual fatigue, resulting in a decline in work efficiency.

[0003] In security inspection images taken by high-energy rays such as X-rays, the boundary features of the target objects provide strong discriminative information for accurate target object detection, especially in the case where the targets are mutually occluded and the features are not obvious. However, due to the monotonous color and lack of luster of the images taken by X-rays, the edges of the target objects are slightly distorted and blurred compared with real images. Therefore, it is difficult for existing target object detection methods to extract and utilize the target object boundary features well, which poses a challenge to the prohibited item detection task determined based on the target object boundary features.

[0004] Yang Xiaogang et al. discussed a new method for enhancing X-ray security inspection images in the paper "Research on X-ray Security Inspection Image Enhancement Method" (published in the 4th issue of CT Theory and Application Research in 2012). The idea of this method is as follows: First, eliminate the background area noise of the security inspection image, then enhance the edges of the image using the Laplace transform, then enhance the contrast of the image through the CLAHE (Constrained Local Histogram Equalization) method, and finally reduce the image noise while maintaining the image edges using a bilateral filter. This method well adapts to the characteristics of X-ray security inspection images and obtains a good enhancement effect. Summary of the Invention

[0005] The primary technical problem to be solved by the present invention is to provide a method for detecting prohibited items in security inspection images.

[0006] Another technical problem to be solved by the present invention is to provide a device for detecting prohibited items in security inspection images.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] According to the first aspect of the embodiments of the present invention, a method for detecting prohibited items in security inspection images is provided, including the following steps:

[0009] Step S1, obtain a security inspection image to be detected;

[0010] Step S2: Input the security inspection image to be detected into a pre-trained multi-feature fusion model based on target object edge information aggregation, and obtain a feature map with enhanced boundary features for target object prediction of the image.

[0011] Preferably, the multi-feature fusion model based on target object edge information aggregation is obtained through the following steps:

[0012] Step S21: Establish an image data set containing prohibited items, and divide the training set data and the test set data;

[0013] Step S22: For each batch of images input from the training set data into the bidirectional feature propagation network, after generating multi-layer initial feature maps of each image, extract multi-layer target object edge feature maps of each image based on the multi-layer initial feature maps;

[0014] Step S23: Input the multi-layer target object edge feature maps in each image into the boundary feature aggregation module to obtain multi-layer boundary feature enhanced feature maps of the target object in each image;

[0015] Step S24: Input the multi-layer boundary feature enhanced feature maps of the target object in each image into the bidirectional feature propagation network to achieve fusion with the multi-layer initial feature maps, and obtain boundary feature enhanced feature maps for target object prediction of each image;

[0016] Step S25: Repeat steps S22 to S24 until a preset number of iterations is reached to obtain a multi-feature fusion model based on target object edge information aggregation;

[0017] Step S26: Use the test set data to test the accuracy of the multi-feature fusion model based on target object edge information aggregation.

[0018] Preferably, when establishing an image data set containing prohibited items, collect several security inspection images, label the items in each image, and then select the images containing prohibited items from them to establish the image data set.

[0019] Preferably, in step S22, use the convolutional layer for each batch of input images to generate multi-layer initial feature maps of each image, and input the multi-layer initial feature maps into the deep feature propagation module of the bidirectional feature propagation network, and perform target object edge feature propagation through the top-down path, so that each layer of target object edge feature map contains the target object edge feature information of all subsequent layers.

[0020] Preferably, in step S23, each layer of boundary feature enhanced feature map is obtained through the boundary feature aggregation module by the following steps;

[0021] Step S230: Copy an edge feature map of the target object at one layer to obtain a corresponding number of edge feature maps of the target object at the corresponding layers.

[0022] Step S231: Perform multi-direction pooling operations on the copied edge feature maps of the target object to obtain boundary feature maps of the target object in the corresponding directions.

[0023] Step S232: Fuse the multi-direction boundary feature maps of the target object with the edge feature maps of the target object at the corresponding layers to obtain feature maps with enhanced boundary features at the corresponding layers.

[0024] Preferably, in step S231, when performing multi-direction pooling operations on the copied edge feature maps of the target object, it is implemented according to the following formula;

[0025]

[0026] where C m，n (i, j) represents the pixel value at the i-th row and j-th column of the feature map after boundary aggregation; represents the maximum value among all the pixel points in the column where the pixel point at the i-th row and j-th column is located during the downward pooling operation in the n-th copied edge feature map of the target object at a certain layer of the edge feature map of the target object; represents the maximum value among all the pixel points in the column where the pixel point at the i-th row and j-th column is located during the upward pooling operation in the n-th copied edge feature map of the target object at a certain layer of the edge feature map of the target object; represents the maximum value among all the pixel points in the row where the pixel point at the i-th row and j-th column is located during the leftward pooling operation in the n-th copied edge feature map of the target object at a certain layer of the edge feature map of the target object; represents the maximum value among all the pixel points in the row where the pixel point at the i-th row and j-th column is located during the rightward pooling operation in the n-th copied edge feature map of the target object at a certain layer of the edge feature map of the target object.

[0027] Preferably, in step S232, add the pixel values at the corresponding positions in the multi-direction boundary feature maps of the target object and the edge feature maps of the target object at the corresponding layers to obtain feature maps with enhanced boundary features at the corresponding layers.

[0028] Preferably, in step S24, the shallow feature propagation module in the bidirectional feature propagation network generates feature maps after shallow information propagation through a bottom-up path for the multi-layer boundary feature enhanced feature maps of the target object in each image, and adds the pixel values at the corresponding positions in the multi-layer initial feature maps of the corresponding images to obtain boundary feature enhanced feature maps for target object prediction in each image.

[0029] According to the second aspect of the embodiments of the present invention, there is provided a device for detecting prohibited items in security inspection images, which is characterized by including a processor and a memory. The processor reads computer programs or instructions in the memory and is used to perform the following operations:

[0030] Obtain a security inspection image to be detected;

[0031] Input the security inspection image to be detected into a pre-trained multi-feature fusion model based on the aggregation of target object edge information, and obtain a feature map with enhanced boundary features for target object prediction of the image.

[0032] According to the third aspect of the embodiments of the present invention, there is provided a computer-readable storage medium. Instructions are stored on the readable storage medium. When it runs on a computer, the computer is caused to perform the following operations:

[0033] Obtain a security inspection image to be detected;

[0034] Input the security inspection image to be detected into a pre-trained multi-feature fusion model based on the aggregation of target object edge information, and obtain a feature map with enhanced boundary features for target object prediction of the image.

[0035] Compared with the prior art, the method and device for detecting prohibited items in security inspection images provided by the present invention obtain a security inspection image to be detected and input it into a pre-trained multi-feature fusion model based on the aggregation of target object edge information, and obtain a feature map with enhanced boundary features for target object prediction of the security inspection image. The features of this feature map are rich and the boundaries are precise, so as to facilitate the detector to detect prohibited items in luggage and effectively solve the problems of blurred and distorted boundaries of target objects in security inspection images taken by X-rays, and monotonous image colors and lack of luster. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a flowchart of the method for detecting prohibited items in security inspection images provided by the embodiments of the present invention;

[0037] Figure 2 It is a training flowchart of the multi-feature fusion model based on the aggregation of target object edge information in the method for detecting prohibited items in security inspection images provided by the embodiments of the present invention;

[0038] Figure 3 It is a flowchart of the boundary feature aggregation module in the method for detecting prohibited items in security inspection images provided by the embodiments of the present invention;

[0039] Figure 4 It is a schematic diagram of the boundary feature aggregation module in the method for detecting prohibited items in security inspection images provided by the embodiments of the present invention;

[0040] Figure 5 This is a schematic structural diagram of a prohibited item detection device in a security inspection image provided by an embodiment of the present invention. Specific implementation manners

[0041] The following further elaborates on the technical content of the present invention in detail in conjunction with the accompanying drawings and specific embodiments.

[0042] When using the existing target detection method to detect small target objects in security inspection images, it is difficult to extract and utilize the boundary features of the target objects well. For example Figure 1 As shown, an embodiment of the present invention first provides a method for detecting prohibited items in security inspection images, which at least includes the following steps:

[0043] Step S1: Obtain the security inspection image to be detected.

[0044] In a place where security inspection is required, use security inspection equipment to scan luggage to collect security inspection images of the luggage (referred to as luggage images). Among them, in the present invention, luggage is a general term for luggage, parcels, bags, boxes, backpacks, handbags, etc.

[0045] Step S2: Input the security inspection image to be detected into a pre-trained multi-feature fusion model based on the aggregation of target object edge information to obtain a feature map with enhanced boundary features for target object prediction of the image.

[0046] The feature map with enhanced boundary features for target object prediction obtained through the trained multi-feature fusion model based on the aggregation of target object edge information is used as the input of the detector to extract image features for the prediction of prohibited items.

[0047] Among them, the multi-feature fusion model based on the aggregation of target object edge information is obtained through the following steps:

[0048] Step S21: Establish an image data set containing prohibited items, and divide the training set data and the test set data.

[0049] From the real scene, use security inspection equipment to collect several security inspection images, store them in JPG format, and the resolution size is uncertain; label the items in each image, and the label details record the image name, the types of objects contained in the image, and the positions of the objects (the area positions of the pixels occupied in the image). The label information of each image forms a label file. When an image contains multiple objects, each object has a corresponding label information in the label file. Among them, security inspection items include multiple types, such as coreless power banks, cored power banks, cosmetics, water cups, computers, mobile phones, tablet computers, non-metallic lighters, knives, etc.

[0050] Select images containing prohibited items from the labeled images to establish an image dataset, and divide the images in this dataset into training set data and test set data according to a preset ratio (such as 4:1).

[0051] Step S22: For each batch of images input from the training set data into the bidirectional feature propagation network, first generate multi-layer initial feature maps for each image, and then extract multi-layer object edge feature maps for each image based on these feature maps.

[0052] As Figure 2 shown, for each batch of images (such as 20 images per batch) input from the training set data into the bidirectional feature propagation network, first use the convolutional layer to generate multi-layer initial feature maps for each image. These initial feature maps are input into the deep feature propagation module of the bidirectional feature propagation network, and through the top-down path, object edge feature propagation is performed, so that each layer of object edge feature map contains the object edge feature information of all subsequent layers. Moreover, object edge feature propagation uses a densely connected method to reduce information distortion during the mapping process.

[0053] Among them, when the deep feature propagation module performs object edge feature propagation, the edge feature map of each layer of the object is obtained according to the following formula:

[0054]

[0055] In the above formula, represents the object edge feature map after the m-th layer of initial feature map passes through deep feature propagation, F m represents the m-th layer of object edge feature map, V represents a convolutional layer with a convolution kernel size of 1x1 used to reduce the dimension of the m-th layer of object edge feature map, U i represents the upsampling operation, M represents the total number of layers of the object deep feature propagation module, which is the same as the number of layers of the initial feature map, F m+i represents all layers of object edge feature maps after the m-th layer of object edge feature map, i = 1, 2... s, and s is a positive integer.

[0056] Step S23: Input the multi-layer object edge feature maps in each image into the boundary feature aggregation module to obtain multi-layer boundary feature enhanced feature maps of the object in each image.

[0057] When obtaining the multi-layer boundary feature enhanced feature maps of the object in each image, each layer of boundary feature enhanced feature map is obtained by the boundary feature aggregation module through the following steps.

[0058] Step S230: Copy a layer of object edge feature map to obtain a preset number of corresponding layer object edge feature maps.

[0059] The boundary feature aggregation module copies each layer of the target object edge feature map of the target object in each image into 5 corresponding layer target object edge feature maps, which are specifically obtained according to the following formula:

[0060]

[0061] where C m，n represents the target object edge feature map after deep feature propagation of the m-th layer of the initial feature map The n-th target object edge feature map obtained by copying (n = 2, 3, 4, 5).

[0062] Step S231: Perform multi-directional pooling operations on the copied target object edge feature maps corresponding to each other to extract key information of the target object boundary, and obtain the target object boundary feature maps in the corresponding directions.

[0063] Perform pooling operations in the up, down, left, and right four directions on 4 of the 5 copied corresponding layer target object edge feature maps, and extract key information of the target object boundary in multiple directions to obtain the target object boundary feature maps in the corresponding directions.

[0064] As Figure 3 shown, for the 4 copied target object edge feature maps of any layer of the target object edge feature map, perform pooling operations in the up, down, left, and right four directions respectively, that is, one target object edge feature map corresponds to one direction of pooling operation. Taking the downward pooling operation as an example, for each row of data starting from the top of a certain target object edge feature map, select the current maximum value of each pixel point in its corresponding column in that row to highlight the pixel boundary points in the image. This process is described by the following formula:

[0065]

[0066] where C m，n (i, j) represents the pixel value of the i-th row and j-th column of the feature map after boundary aggregation; represents the maximum value among the pixel points in the column where the pixel point at the i-th row and j-th column is located and all the pixel points before it during the downward pooling operation of the n-th copied target object edge feature map of a certain layer of the target object edge feature map; represents the maximum value among the pixel points in the column where the pixel point at the i-th row and j-th column is located and all the pixel points before it during the upward pooling operation of the n-th copied target object edge feature map of a certain layer of the target object edge feature map; represents the maximum value among the pixel points in the row where the pixel point at the i-th row and j-th column is located and all the pixel points before it during the leftward pooling operation of the n-th copied target object edge feature map of a certain layer of the target object edge feature map; It represents the maximum value among the pixel points in the same row as the pixel point at the i-th row and j-th column during the right pooling operation in the n-th target object edge feature map that replicates the edge feature map of a certain layer, including this pixel point and all the pixel points before it.

[0067] Performing multi-directional pooling operations on the replicated target object edge feature map can highlight the pixel mutation parts in a certain area of the image and can completely cover the boundary area of the target object in the image.

[0068] Step S232: Fuse the multi-directional target object boundary feature maps with the corresponding layer target object edge feature maps to obtain the feature maps with enhanced boundary features in the corresponding layer.

[0069] As Figure 4 shown, after performing pooling operations on a layer of target object edge feature maps in four directions, the pixel values of the pixel points at the corresponding positions in the four-direction target object boundary feature maps obtained are added to the pixel values of the corresponding layer target object edge feature maps, strengthening the boundary feature information in the feature maps to obtain the feature maps with enhanced boundary features in the corresponding layer. Specifically, it is obtained according to the following formula.

[0070]

[0071] Among them, B m represents the feature map with enhanced boundary features in the m-th layer, B m，n represents all the pixel values of the target object boundary feature maps in each direction obtained by performing multi-directional pooling operations on the target object edge feature map in the m-th layer, and C m，1 represents all the pixel values of the target object edge feature map in the m-th layer. Since the target object edge features after dimensionality reduction incorporate boundary features, the generated feature maps with enhanced boundary features can help accurately locate the area where the target object is located in the image and predict its position information.

[0072] Step S24: Input the multi-layer feature maps with enhanced boundary features of the target object in each image into the bidirectional feature propagation network to achieve fusion with the multi-layer initial feature maps, obtaining the feature maps with enhanced boundary features for target object prediction in each image.

[0073] The shallow feature propagation module in the bidirectional feature propagation network maps the multi-layer feature maps with enhanced boundary features of the target object in each image through a bottom-up path in a densely connected form to the subsequent features of the bidirectional feature propagation network, generating the feature maps after shallow information propagation, and performing feature fusion with the multi-layer initial feature maps. Finally, this is used as the feature maps with enhanced boundary features for target object prediction in each image. The shallow propagation module can generate the feature maps after shallow information propagation according to the following formula:

[0074]

[0075] In the above formula, represents the feature map after the shallow feature propagation module processes the feature map of the boundary feature enhancement of the target object in the m-th layer, generating the feature map after the shallow information propagation. B m represents the feature map of the boundary feature enhancement of the target object in the m-th layer, V represents a convolutional layer with a convolutional kernel size of 1x1 used to reduce the dimension of the feature map of the boundary feature enhancement in the m-th layer, D i represents the downsampling operation, B m-i represents the feature maps of the boundary feature enhancement of all subsequent layers of the target object after the feature map of the boundary feature enhancement of the target object in the m-th layer, where i = 1, 2... s, and s is a positive integer.

[0076] Add the pixel values at the corresponding positions in the feature maps of the multi-layer shallow information propagation of each image generated by passing the feature maps of the multi-layer boundary feature enhancement of each image through the shallow feature propagation module to the pixel values of the corresponding positions in the multi-layer initial feature maps of each image, obtaining the feature maps of the boundary feature enhancement for target object prediction of each image.

[0077] Step S25: Repeat steps S22 to S24 until the preset number of iterations is reached, obtaining a multi-feature fusion model based on the aggregation of target object edge information.

[0078] In the two-way feature propagation network, each time a batch of training set images is input, and the process of steps S22 to S24 is executed once, the parameters of the two-way feature propagation network will be updated once. After the network parameters are iteratively updated to reach the preset number of times, a multi-feature fusion model based on the aggregation of target object edge information is obtained.

[0079] Step S26: Use the test set data to test the accuracy of the multi-feature fusion model based on the aggregation of target object edge information.

[0080] Input the images in the test set data into the multi-feature fusion model based on the aggregation of target object edge information obtained after training, and verify the accuracy of the multi-feature fusion model based on the aggregation of target object edge information. Using the test set data to test the accuracy of the multi-feature fusion model based on the aggregation of target object edge information is an existing mature technology (such as using a scoring method), which will not be elaborated here.

[0081] It should be emphasized that the method for detecting prohibited items in security inspection images provided by the present invention can also be used in training various X-ray target detection models, which can realize feature extraction and enhancement of object boundaries in X-ray images, providing strong discriminative information for accurate target detection, and can also be applicable in cases where targets are mutually occluded and features are not obvious, which will not be elaborated one by one here.

[0082] In addition, as Figure 5As shown in the figure, an embodiment of the present invention further provides a device for detecting prohibited items in security inspection images, which includes a processor 32 and a memory 31. According to actual needs, it may further include a communication component, a sensor component, a power supply component, a multimedia component, and an input / output interface. Among them, the memory, the communication component, the sensor component, the power supply component, the multimedia component, and the input / output interface are all connected to the processor 32. As previously mentioned, the memory 31 may be a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, etc.; the processor 32 may be a central processing unit (CPU), a graphics processing unit (GPU), a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), a digital signal processing (DSP) chip, etc. Other communication components, sensor components, power supply components, multimedia components, etc. can all be implemented by general components in existing security inspection devices, and will not be specifically described here.

[0083] In addition, the device for detecting prohibited items in security inspection images provided by an embodiment of the present invention includes a processor 32 and a memory 31. The processor 32 reads computer programs or instructions in the memory 31 and is used to perform the following operations:

[0084] Obtain the security inspection image to be detected.

[0085] In a place where security inspection is required, a security inspection device is used to scan luggage to collect the security inspection image of the luggage (referred to as the luggage image). Among them, in the present invention, luggage is a general term for luggage, parcels, bags, boxes, backpacks, handbags, etc.

[0086] Input the security inspection image to be detected into a pre-trained multi-feature fusion model based on the aggregation of target object edge information to obtain a feature map with enhanced boundary features for target object prediction of the image.

[0087] In addition, an embodiment of the present invention further provides a computer-readable storage medium. Instructions are stored on the readable storage medium. When it runs on a computer, it causes the computer to execute as described above Figure 1 The method for detecting prohibited items in security inspection images will not be elaborated on its specific implementation here.

[0088] In addition, an embodiment of the present invention further provides a computer program product containing instructions. When it runs on a computer, it causes the computer to execute as described above Figure 1 The method for detecting prohibited items in security inspection images will not be elaborated on its specific implementation here.

[0089] The method and device for detecting prohibited items in security inspection images provided by the present invention obtain the security inspection images to be detected and input them into a pre-trained multi-feature fusion model based on the aggregation of target object edge information, so as to obtain a feature map with enhanced boundary features for target object prediction in the image. The feature map has rich features and precise boundaries, which is convenient for the detector to detect prohibited items in luggage, and effectively solves the problems of blurred and distorted boundaries of target objects, monotonous image colors and lack of luster in the security inspection images taken by X-rays.

[0090] The above has described in detail the method and device for detecting prohibited items in security inspection images provided by the present invention. For those of ordinary skill in the art, any obvious changes made without departing from the essence of the present invention will fall within the protection scope of the patent right of the present invention.

Claims

1. A method for detecting prohibited items in security inspection images, characterized in that It includes the following steps: Step S1: Obtain the security inspection image to be detected; Step S2: Input the security inspection image to be detected into a pre-trained multi-feature fusion model based on the aggregation of target object edge information to obtain a feature map with enhanced boundary features for target object prediction of the image; Among them, the multi-feature fusion model based on the aggregation of target object edge information is obtained through the following steps: Step S21: Establish an image data set containing prohibited items, and divide the training set data and the test set data; Step S22: For each batch of images input from the training set data into the bidirectional feature propagation network, after generating multi-layer initial feature maps of each image, the deep feature propagation module of the bidirectional feature propagation network extracts multi-layer target object edge feature maps of each image based on the multi-layer initial feature maps; Step S23: Input the multi-layer target object edge feature maps in each image into the boundary feature aggregation module, copy one layer of the target object edge feature map to obtain a preset number of corresponding layer target object edge feature maps, perform multi-directional pooling operations on the copied target object edge feature maps to obtain target object boundary feature maps in corresponding directions; fuse the multi-directional target object boundary feature maps with the corresponding layer target object edge feature maps to obtain multi-layer boundary feature enhanced feature maps of the target object in each image; Step S24: Input the multi-layer boundary feature enhanced feature maps of the target object in each image into the shallow feature propagation module of the bidirectional feature propagation network to achieve fusion with the multi-layer initial feature maps, and obtain a feature map with enhanced boundary features for target object prediction of each image; Step S25: Repeat steps S22 to S24 until the preset number of iterations is reached to obtain a multi-feature fusion model based on the aggregation of target object edge information; Step S26: Use the test set data to test the accuracy of the multi-feature fusion model based on the aggregation of target object edge information.

2. The method for detecting prohibited items in a security inspection image according to claim 1, wherein: When establishing an image data set containing prohibited items, collect several security inspection images, label the items in each image, and then select the images containing prohibited items from them to establish an image data set.

3. The method for detecting prohibited items in security inspection images according to claim 1, characterized in that In step S22, Use the convolutional layer for each batch of input images to generate multi-layer initial feature maps of each image. The multi-layer initial feature maps are input into the deep feature propagation module of the bidirectional feature propagation network, and through the top-down path, target object edge feature propagation is performed, so that each layer of the target object edge feature map contains the target object edge feature information of all subsequent layers.

4. The method for detecting prohibited items in security inspection images according to claim 3, wherein In step S231, When performing multi-directional pooling operations on the copied target object edge feature maps, it is implemented according to the following formula; Among them, represents the pixel value at the i-th row and j-th column of the feature map after boundary aggregation; represents the maximum value among the pixel at the i-th row and j-th column and all the pixels before it in the column where the pixel at the i-th row and j-th column is located during the downsampling operation in the n-th replicated target object edge feature map of a certain layer of the target object edge feature map; represents the maximum value among the pixel at the i-th row and j-th column and all the pixels before it in the column where the pixel at the i-th row and j-th column is located during the upsampling operation in the n-th replicated target object edge feature map of a certain layer of the target object edge feature map; represents the maximum value among the pixel at the i-th row and j-th column and all the pixels before it in the row where the pixel at the i-th row and j-th column is located during the left pooling operation in the n-th replicated target object edge feature map of a certain layer of the target object edge feature map; represents the maximum value among the pixel at the i-th row and j-th column and all the pixels before it in the row where the pixel at the i-th row and j-th column is located during the right pooling operation in the n-th replicated target object edge feature map of a certain layer of the target object edge feature map.

5. The method for detecting prohibited items in security inspection images according to claim 4, characterized in that In step S232, Add the pixel point values at the corresponding positions in the multi-directional target object boundary feature maps and the corresponding layer target object edge feature maps to obtain a feature map with enhanced boundary features for the corresponding layer.

6. The method for detecting prohibited items in security inspection images according to claim 1, characterized in that In step S24, In the shallow feature propagation module of the bidirectional feature propagation network, the feature maps with enhanced multi-layer boundary features of the target objects in each image are used to generate the feature maps after shallow information propagation through a bottom-up path, and the pixel values at the corresponding positions in the multi-layer initial feature maps of the corresponding images are added to obtain the feature maps with enhanced boundary features for target object prediction in each image.

7. A detection device for prohibited items in security inspection images, characterized in that It includes a processor and a memory. The processor reads the computer program or instructions in the memory and is used to perform the operations of the method for detecting contraband items in security inspection images according to any one of claims 1 to 6.

8. A computer-readable storage medium, on which instructions are stored. When the instructions run on a computer, the computer is made to perform the operations of the method for detecting contraband items in security inspection images according to any one of claims 1 to 6.

Images