Millimeter wave SAR human body security image contraband detection method and device

Abstract

This invention provides a method and apparatus for detecting contraband in millimeter-wave SAR human security inspection images, belonging to the field of millimeter-wave SAR detection and recognition. The steps are as follows: Step 1: Acquire human security inspection images and construct a dataset; Step 2: Label the targets in the dataset to obtain a labeled dataset; Step 3: Iteratively train the FSRCNN model based on the dataset from Step 1 to obtain a trained FSRCNN model, and input the labeled dataset from Step 2 into the trained FSRCNN model to obtain a super-resolution reconstructed image; Step 4: Train the super-resolution reconstructed image based on the improved YOLOv5s algorithm, when the loss function l... MPDIoU When the minimum value is reached, the contraband detection model is obtained; Step 5: Input the security inspection image of the human body to be detected into the trained FSRCNN model for super-resolution reconstruction, input the super-resolution reconstructed image into the contraband detection model, and obtain the detection result; Combine the FSRCNN super-resolution reconstruction algorithm with the improved YOLOv5s target detection framework to enhance image quality and improve target detection accuracy.

Description

Technical Field

[0001] This invention relates to the field of millimeter-wave SAR detection and identification technology, and in particular to a method and apparatus for detecting contraband in millimeter-wave SAR human body security images. Background Technology

[0002] Synthetic Aperture Radar (SAR) images have strong anti-jamming characteristics. Millimeter-wave SAR technology can provide more detailed images than traditional metal detectors, but it still has the following problems: (1) The resolution of millimeter-wave SAR imaging is usually still lower than that of X-ray scanning technology. Low resolution may result in insufficient image detail and difficulty in identifying small or closely packed objects. (2) Millimeter-wave SAR imaging technology is very sensitive to environmental conditions, especially temperature and humidity. Temperature changes will alter the dielectric constant and density of air, thereby affecting the propagation speed and direction of millimeter waves, leading to attenuation or scattering of millimeter waves. Increased humidity will also cause similar problems because moisture in the air will affect the dielectric constant of the air. Fluctuations in these environmental variables may lead to unstable millimeter-wave SAR image quality, manifested as reduced resolution and loss of detail, resulting in impaired imaging effects and ultimately a deterioration in the detection and identification of hazardous materials. Summary of the Invention

[0003] The technical problem to be solved by this invention is how to improve the efficiency and accuracy of security checks.

[0004] This invention solves the above-mentioned technical problems through the following technical solution: a millimeter-wave SAR human body security inspection image contraband detection method, comprising:

[0005] Step 1: Collect images of people undergoing security checks and construct a dataset;

[0006] Step 2: Label the targets in the dataset to obtain the labeled dataset;

[0007] Step 3: Iteratively train the FSRCNN model based on the dataset from Step 1 to obtain a trained FSRCNN model. Input the labeled dataset from Step 2 into the trained FSRCNN model to obtain the super-resolution reconstructed image.

[0008] Step 4: Train the super-resolution reconstructed image based on the improved YOLOv5s algorithm, when the loss function... When the minimum value is reached, a contraband detection model is obtained;

[0009] Step 5: Input the human body security inspection image to be detected into the trained FSRCNN model for super-resolution reconstruction, and input the super-resolution reconstructed image into the contraband detection model to obtain the detection result.

[0010] This invention combines the FSRCNN super-resolution reconstruction algorithm with the improved YOLOv5s target detection framework to enhance image quality and improve the accuracy of target detection. It can effectively detect dangerous items carried by the human body, especially in high-precision millimeter-wave SAR human security inspection scenarios. It improves the efficiency and accuracy of the contraband detection system and helps to better identify and prevent potential security threats.

[0011] Preferably, in step two, LabelImg image annotation software is used to annotate the rectangular boundaries of the targets in the dataset according to the PASCAL VOC data format.

[0012] Preferably, the structure of the FSRCNN model in step three includes a feature extraction layer, a shrinking layer, a nonlinear feature mapping layer, an expanding layer, and a deconvolution layer. The feature extraction layer extracts useful features from the input dataset. The kernel size of the feature extraction layer is 5, and the number of output channels is d. The shrinking layer reduces the number of feature channels from d to s using a 1×1 convolution kernel. Each nonlinear feature mapping layer is a 3×3 convolution kernel. The nonlinear feature mapping layer refines and enhances the image features. The expanding layer is a 1×1 convolution layer used to expand the number of feature channels from s back to d. The deconvolution layer uses ConvTranspose2d convolution to upsample the feature map to a higher resolution.

[0013] Preferably, in step three, the FSRCNN model is iteratively trained based on the dataset from step one, and the Adam algorithm is used for optimization training. Training stops when the loss function MSE reaches its minimum or the total number of iterations is reached. The loss function MSE is used to measure the error between the super-resolution reconstructed image and the original high-resolution image, and the calculation formula is as follows:

[0014]

[0015] in, These are the pixel values ​​in the original image. These are the pixel values ​​in the super-resolution reconstructed image predicted by the FSRCNN model. n This represents the number of samples in the dataset.

[0016] Preferably, the loss function in step four The calculation formula is:

[0017]

[0018]

[0019] in, , These are the distances to the corresponding vertices. , , ( and( They are the real bounding boxes Top left and bottom right corners, and( Predicted boxes Top left and bottom right corners w , h These represent the width and height of the input super-resolution reconstructed image, respectively.

[0020] This invention also provides a millimeter-wave SAR human body security image contraband detection device, comprising:

[0021] The data acquisition unit is used to acquire images of people undergoing security checks and to build datasets.

[0022] Data annotation units are used to annotate targets in a dataset to obtain an annotated dataset;

[0023] The super-resolution reconstruction unit is used to iteratively train the FSRCNN model based on the dataset from the data acquisition unit to obtain a trained FSRCNN model. The labeled dataset from the data labeling unit is then input into the trained FSRCNN model to obtain a super-resolution reconstructed image.

[0024] The model training unit is used to train super-resolution reconstructed images based on the improved YOLOv5s algorithm, when the loss function... When the minimum value is reached, a contraband detection model is obtained;

[0025] The model detection unit is used to input the security inspection image of the human body to be detected into the trained FSRCNN model for super-resolution reconstruction, and then input the super-resolution reconstructed image into the contraband detection model to obtain the detection result.

[0026] Preferably, the data annotation unit uses LabelImg image annotation software to annotate the rectangular boundaries of the targets in the dataset according to the PASCAL VOC data format.

[0027] Preferably, the structure of the FSRCNN model in the super-resolution reconstruction unit includes a feature extraction layer, a shrinking layer, a nonlinear feature mapping layer, an enlarging layer, and a deconvolution layer. The feature extraction layer extracts useful features from the input dataset. The kernel size of the feature extraction layer is 5, and the number of output channels is d. The shrinking layer reduces the number of feature channels from d to s using a 1×1 convolution kernel. Each nonlinear feature mapping layer is a 3×3 convolution kernel. The nonlinear feature mapping layer refines and enhances the image features. The enlarging layer is a 1×1 convolution layer used to expand the number of feature channels from s back to d. The deconvolution layer uses ConvTranspose2d convolution to upsample the feature map of reduced resolution to high resolution.

[0028] Preferably, in the super-resolution reconstruction unit, the FSRCNN model is iteratively trained based on the dataset from the data acquisition unit, and the Adam algorithm is used for optimization training. Training stops when the loss function MSE reaches its minimum or the number of iterations is reached. The loss function MSE is used to measure the error between the super-resolution reconstructed image and the original high-resolution image, and the calculation formula is as follows:

[0029]

[0030] in, These are the pixel values ​​in the original image. These are the pixel values ​​in the super-resolution reconstructed image predicted by the FSRCNN model. n This represents the number of samples in the dataset.

[0031] Preferably, the loss function in the model training unit The calculation formula is:

[0032]

[0033]

[0034] in, , These are the distances to the corresponding vertices. , , ( and( They are the real bounding boxes Top left and bottom right corners, and( Predicted boxes Top left and bottom right corners w , h These represent the width and height of the input super-resolution reconstructed image, respectively.

[0035] The advantages provided by this invention are:

[0036] (1) This invention combines the FSRCNN super-resolution reconstruction algorithm with the improved YOLOv5s target detection framework to enhance image quality and improve the accuracy of target detection. It can effectively detect dangerous items carried by the human body, especially in high-precision millimeter-wave SAR human security inspection scenarios. It improves the efficiency and accuracy of the contraband detection system and helps to better identify and prevent potential security threats.

[0037] (2) This invention improves the rated resolution and detail richness of the image through super-resolution processing. It uses super-resolution reconstruction technology to recover high-resolution image details from low-resolution millimeter-wave SAR security inspection images, significantly improving the image clarity and the visibility of edges and details. This makes it easier to identify potentially dangerous items hidden under clothing, such as knives and explosives. At the same time, it enhances the ability to distinguish between non-prohibited items and dangerous items such as electronic devices and metal items.

[0038] (3) Millimeter-wave SAR human security inspection images are usually accompanied by high noise and clutter interference. This noise may make the boundaries of real objects unclear or covered by noise, making the localization of bounding boxes more difficult. The MPDIoU loss function used in this invention directly optimizes the position and size error of the bounding boxes by reducing the average distance between the predicted box and the real box, thereby optimizing the localization accuracy of the bounding boxes. Especially in environments with high noise and complex backgrounds, it can reduce false alarms and false negatives of contraband and optimize target detection performance. In particular, when dealing with overlapping or dense areas of targets, it can significantly improve the accuracy and reliability of the system. Attached Figure Description

[0039] Figure 1 A flowchart of a millimeter-wave SAR human body security inspection image contraband detection method provided for embodiments of the present invention;

[0040] Figure 2 The structural diagram of the FSRCNN model in the millimeter-wave SAR human body security inspection image contraband detection method provided in the embodiments of the present invention;

[0041] Figure 3 A schematic diagram of the MPDIoU loss function in the millimeter-wave SAR human body security inspection image contraband detection method provided in the embodiments of the present invention;

[0042] Figure 4 A simulation diagram of a millimeter-wave SAR human body security inspection image contraband detection method provided for an embodiment of the present invention;

[0043] Figure 5(a) shows a picture of the contraband ceramic knife;

[0044] Figure 5(b) shows a millimeter-wave SAR image of a volunteer carrying a ceramic knife;

[0045] Figure 6(a) shows a picture of the contraband wrench;

[0046] Figure 6(b) shows a millimeter-wave SAR image of a volunteer carrying a wrench;

[0047] Figure 7(a) shows a picture of a contraband handgun;

[0048] Figure 7(b) shows a millimeter-wave SAR image of a volunteer carrying a pistol;

[0049] Figure 8(a) is a picture of a prohibited metal folding knife;

[0050] Figure 8(b) shows a millimeter-wave SAR image of a volunteer carrying a metal folding knife;

[0051] Figure 9(a) shows a picture of the contraband hammer;

[0052] Figure 9(b) shows a millimeter-wave SAR image of a volunteer carrying a hammer. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0054] Example 1

[0055] like Figure 1 As shown, this embodiment provides a method for detecting contraband in millimeter-wave SAR human body security images, including the following steps:

[0056] Step 1: Collect images of human body security checks and construct a dataset. This invention uses a millimeter-wave SAR human body security check device built in the laboratory to collect images of volunteers of different body types (tall, short, thin, fat) carrying common prohibited items and placing them on different parts of their bodies, as shown in Figures 5(b), 6(b), 7(b), 8(b), and 9(b). Figures 5(a), 6(a), 7(a), 8(a), and 9(a) respectively show five common prohibited items, including ceramic knives, wrenches, pistols, metal folding knives, and hammers.

[0057] Step 2: Label the targets in the dataset to obtain a labeled dataset. This invention uses LabelImg image annotation software to label the targets in the dataset with rectangular boundaries according to the PASCAL VOC data format, obtaining a labeled dataset. The labeled dataset is then divided into a training set and a validation set in an 8:2 ratio.

[0058] Step 3: Iteratively train the FSRCNN model based on the dataset from Step 1 to obtain a trained FSRCNN model. Input the labeled dataset from Step 2 into the trained FSRCNN model to obtain a super-resolution reconstructed image.

[0059] like Figure 2As shown, the FSRCNN model structure includes a feature extraction layer, a shrinking layer, a non-linear feature mapping layer, a dilating layer, and a deconvolution layer. The first layer of the network is the feature extraction layer, which uses a convolution with a kernel size of 5 and d output channels to extract useful features from the input millimeter-wave SAR human security inspection image. Appropriate padding is used to maintain the image size, and a PReLU activation function is used to enable the model to capture non-linear relationships.

[0060] The second layer is the shrinking layer, which reduces the feature channels from d to s using a 1×1 convolution kernel to reduce the number of model parameters and computational complexity, thereby improving computational speed, and also uses the PRELU activation function.

[0061] The third layer is a non-linear feature mapping layer, which consists of m layers. Each non-linear feature mapping layer uses a 3×3 convolutional kernel. The size of the feature map is kept unchanged by appropriate padding, and the PReLU activation function is used. Through the m non-linear feature mapping layers, the feature representation of the image is further refined and enhanced.

[0062] The fourth layer is the expansion layer, which follows the non-linear feature mapping layer. The expansion layer is a 1×1 convolutional layer used to expand the number of feature channels from s back to d, which helps to recover the important information lost in the shrinking layer. The PReLU activation function is also used in the expansion layer.

[0063] The fifth layer is a deconvolutional layer, which uses ConvTranspose2d convolution to upsample the reduced-resolution feature maps to a higher resolution.

[0064] In step three, the FSRCNN model is iteratively trained based on the dataset from step one, and the Adam algorithm is used for optimization training. Training stops when the loss function MSE reaches its minimum or the total number of iterations is reached. The loss function MSE is used to measure the error between the super-resolution reconstructed image and the original high-resolution image, and the calculation formula is as follows:

[0065]

[0066] in, These are the pixel values ​​in the original image. These are the pixel values ​​in the super-resolution reconstructed image predicted by the FSRCNN model. n This represents the number of samples in the dataset.

[0067] In step three, the performance of the trained FSRCNN model is evaluated using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). PSNR measures the image reconstruction quality by comparing the difference between the original image and the distorted image, while SSIM measures the similarity between the reconstructed image and the original high-resolution image. The formula for calculating PSNR is as follows:

[0068]

[0069] in, MAX Represents the maximum value of a pixel. Therefore, we can conclude that:

[0070]

[0071] The Structural Similarity Index (SSIM) is used to measure the similarity between two images. The formula for calculating the SSIM is as follows:

[0072]

[0073] in, , Images x ,image y The mean, , Images x ,image y standard deviation For image x and images y covariance, , Images x ,image y variance , It is a constant used for stability. , , The dynamic range of pixel values. , It is a very small constant.

[0074] This invention applies the weights of the trained FSRCNN model to the labeled dataset in step two, improves the image's rated resolution and detail richness through super-resolution processing, and recovers high-resolution image details from low-resolution millimeter-wave SAR security inspection images using super-resolution reconstruction technology. This significantly improves the image's clarity and the visibility of edges and details, making it easier to identify potentially dangerous items hidden under clothing, such as knives and explosives. It also enhances the ability to distinguish between non-prohibited items and dangerous items such as electronic devices and metal objects, which is particularly crucial in security inspection work.

[0075] Step 4: Train the super-resolution reconstructed image based on the improved YOLOv5s algorithm, when the loss function... When the value is minimized, the contraband detection model is obtained; loss function The calculation formula is:

[0076]

[0077]

[0078] in, , These are the distances to the corresponding vertices. , , ( and( They are the real bounding boxes Top left and bottom right corners, and( Predicted boxes Top left and bottom right corners w , h These represent the width and height of the input super-resolution reconstructed image, respectively.

[0079] Millimeter-wave SAR images of human security checks are often accompanied by high noise and clutter interference. This noise can make the boundaries of real objects indistinct or obscured by noise, making bounding box localization more difficult. Although the CIoU loss function considers the overlapping area, center distance, and aspect ratio of the bounding boxes, its performance may be affected in high-noise environments because the model may have difficulty accurately distinguishing objects from noise or background. The CIoU loss function considers the center distance and aspect ratio of the bounding boxes, but for small objects, even a small positional deviation can lead to a large CIoU loss, which may hinder the model from learning to accurately detect small objects.

[0080] The MPDIoU loss function directly optimizes the position and size errors of bounding boxes by reducing the average distance between the predicted and ground truth boxes. Figure 3As shown, this method is more direct and effective than the CIoU loss function in reducing prediction errors because the MPDIoU loss function focuses directly on distance rather than optimizing through multiple indirect parameters (such as aspect ratio). By integrating the MPDIoU loss function into the YOLOv5s algorithm, the localization accuracy of the bounding boxes is optimized. This is particularly effective in noisy and complex background environments, reducing false positives and false negatives for contraband and improving target detection performance. Especially when handling overlapping or densely populated areas, it significantly improves the system's accuracy and reliability.

[0081] Step 5: Input the human body security inspection image to be detected into the trained FSRCNN model for super-resolution reconstruction, and input the super-resolution reconstructed image into the contraband detection model to obtain the detection result.

[0082] The human body image to be detected during security checks is input into a trained FSRCNN model for 4x super-resolution reconstruction. The weights of the contraband detection model are then used to detect contraband in the super-resolution reconstructed image, and the detection results can be displayed on the user interface. Super-resolution technology can improve the quality of the human body image to be detected during security checks, help improve the accuracy of contraband detection, reduce false positives, and improve overall security check efficiency.

[0083] like Figure 4 As shown in the figure, the original YOLOv5s performs best on millimeter-wave SAR human body security screening data after 4x super-resolution, and the improved YOLOv5s performs best on the same data. The horizontal axis represents the number of iterations, and the vertical axis represents the average accuracy. The figure shows that the data without super-resolution performs the worst on YOLOv5s, the data with 4x super-resolution performs better than the data without super-resolution, and the improved YOLOv5s performs the best.

[0084] This invention combines the FSRCNN super-resolution reconstruction algorithm with the improved YOLOv5s target detection framework to enhance image quality and improve target detection accuracy. It can effectively detect dangerous items carried by the human body and demonstrates superior processing capabilities compared to traditional single technologies in practical operation. In particular, it improves the efficiency and accuracy of the contraband detection system in high-precision millimeter-wave SAR human security inspection scenarios, helping to better identify and prevent potential security threats. It not only reduces the inconvenience and cost caused by false alarms, but also significantly improves the overall system performance and practical value.

[0085] Example 2

[0086] This embodiment provides a millimeter-wave SAR human body security image contraband detection device, including:

[0087] The data acquisition unit is used to collect images of people undergoing security checks and to build datasets.

[0088] The data annotation unit is used to annotate the targets in the dataset, resulting in a labeled dataset. The LabelImg image annotation software is used to annotate the rectangular boundaries of the targets in the dataset according to the PASCAL VOC data format.

[0089] The super-resolution reconstruction unit is used to iteratively train the FSRCNN model based on the dataset from the data acquisition unit to obtain a trained FSRCNN model. The labeled dataset from the data annotation unit is then input into the trained FSRCNN model to obtain a super-resolution reconstructed image.

[0090] The FSRCNN model structure includes a feature extraction layer, a shrinking layer, a non-linear feature mapping layer, an enlarging layer, and a deconvolution layer. The feature extraction layer extracts useful features from the input dataset. The kernel size of the feature extraction layer is 5, and the number of output channels is d. The shrinking layer reduces the number of feature channels from d to s using a 1×1 convolution kernel. Each non-linear feature mapping layer is a 3×3 convolution kernel. The non-linear feature mapping layer refines and enhances the image features. The enlarging layer is a 1×1 convolution layer used to expand the number of feature channels from s back to d. The deconvolution layer uses ConvTranspose2d convolution to upsample the feature maps from the reduced resolution to a higher resolution.

[0091] Iterative training uses the Adam algorithm for optimization. Training stops when the loss function MSE reaches its minimum or the required number of iterations is reached. The loss function MSE measures the error between the super-resolution reconstructed image and the original high-resolution image, and its calculation formula is as follows:

[0092]

[0093] in, These are the pixel values ​​in the original image. These are the pixel values ​​in the super-resolution reconstructed image predicted by the FSRCNN model. n This represents the number of samples in the dataset.

[0094] The model training unit is used to train super-resolution reconstructed images based on the improved YOLOv5s algorithm, when the loss function... When the value is minimized, the contraband detection model is obtained; loss function The calculation formula is:

[0095]

[0096]

[0097] in, , These are the distances to the corresponding vertices. , , ( and( They are the real bounding boxes Top left and bottom right corners, and( Predicted boxes Top left and bottom right corners w , h These represent the width and height of the input super-resolution reconstructed image, respectively.

[0098] The model detection unit is used to input the security inspection image of the human body to be detected into the trained FSRCNN model for super-resolution reconstruction, and then input the super-resolution reconstructed image into the contraband detection model to obtain the detection result.

[0099] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting contraband in millimeter-wave SAR human body security images, characterized in that: include: S1. Collect images of human bodies during security checks and construct a dataset; S2. Label the targets in the dataset to obtain a labeled dataset; S3. Iteratively train the FSRCNN model based on the dataset from step S1 to obtain a trained FSRCNN model. Input the labeled dataset from step S2 into the trained FSRCNN model to obtain a super-resolution reconstructed image. S4. Training is performed on the super-resolution reconstructed image based on the improved YOLOv5s algorithm, when the loss function... When the value is minimized, the contraband detection model is obtained; loss function for: in, , These are the distances to the corresponding vertices. , , and They are the real bounding boxes Top left and bottom right corners and Predicted boxes Top left and bottom right corners w , h These represent the width and height of the input super-resolution reconstructed image, respectively. S5. Input the human body security inspection image to be detected into the trained FSRCNN model for super-resolution reconstruction, input the super-resolution reconstructed image into the contraband detection model, and obtain the detection result.

2. The method for detecting contraband in millimeter-wave SAR human body security images according to claim 1, characterized in that: In step S2, the LabelImg image annotation software is used to annotate the rectangular boundaries of the targets in the dataset according to the PASCAL VOC data format.

3. The method for detecting contraband in millimeter-wave SAR human body security images according to claim 1, characterized in that: In step S3, the FSRCNN model structure includes a feature extraction layer, a shrinking layer, a non-linear feature mapping layer, an enlarging layer, and a deconvolution layer. The feature extraction layer extracts useful features from the input dataset. The kernel size of the feature extraction layer is 5, and the number of output channels is d. The shrinking layer reduces the number of feature channels from d to s using a 1×1 convolution kernel. Each non-linear feature mapping layer is a 3×3 convolution kernel. The non-linear feature mapping layer refines and enhances the image features. The enlarging layer is a 1×1 convolution layer used to expand the number of feature channels from s back to d. The deconvolution layer uses ConvTranspose2d convolution to upsample the feature map from a lower resolution to a higher resolution.

4. The method for detecting contraband in millimeter-wave SAR human body security images according to claim 1, characterized in that: In step S3, the FSRCNN model is iteratively trained based on the dataset from step S1, and the Adam algorithm is used for optimization training. Training stops when the loss function MSE reaches its minimum or the total number of iterations is reached. The loss function MSE is used to measure the error between the super-resolution reconstructed image and the original high-resolution image, and the calculation formula is as follows: in, These are the pixel values ​​in the original image. These are the pixel values ​​in the super-resolution reconstructed image predicted by the FSRCNN model. n This represents the number of samples in the dataset.

5. A millimeter-wave SAR human body security inspection image contraband detection device, characterized in that: include: The data acquisition unit is used to acquire images of people undergoing security checks and to build datasets. Data annotation units are used to annotate targets in a dataset to obtain an annotated dataset; The super-resolution reconstruction unit is used to iteratively train the FSRCNN model based on the dataset from the data acquisition unit to obtain a trained FSRCNN model. The labeled dataset from the data labeling unit is then input into the trained FSRCNN model to obtain a super-resolution reconstructed image. The model training unit is used to train super-resolution reconstructed images based on the improved YOLOv5s algorithm, when the loss function... When the value is minimized, the contraband detection model is obtained; loss function for: in, , These are the distances to the corresponding vertices. , , and They are the real bounding boxes Top left and bottom right corners and Predicted boxes Top left and bottom right corners w , h These represent the width and height of the input super-resolution reconstructed image, respectively. The model detection unit is used to input the security inspection image of the human body to be detected into the trained FSRCNN model for super-resolution reconstruction, and then input the super-resolution reconstructed image into the contraband detection model to obtain the detection result.

6. The millimeter-wave SAR human body security inspection image contraband detection device according to claim 5, characterized in that: The data annotation unit uses LabelImg image annotation software to annotate the rectangular boundaries of targets in the dataset according to the PASCAL VOC data format.

7. The millimeter-wave SAR human body security inspection image contraband detection device according to claim 5, characterized in that: The FSRCNN model in the super-resolution reconstruction unit includes a feature extraction layer, a shrinking layer, a nonlinear feature mapping layer, an enlarging layer, and a deconvolution layer. The feature extraction layer extracts useful features from the input dataset. The kernel size of the feature extraction layer is 5, and the number of output channels is d. The shrinking layer reduces the number of feature channels from d to s using a 1×1 convolution kernel. Each nonlinear feature mapping layer is a 3×3 convolution kernel. The nonlinear feature mapping layer refines and enhances the image features. The enlarging layer is a 1×1 convolution layer used to expand the number of feature channels from s back to d. The deconvolution layer uses ConvTranspose2d convolution to upsample the feature map from the reduced resolution to a higher resolution.

8. The millimeter-wave SAR human body security inspection image contraband detection device according to claim 5, characterized in that: The super-resolution reconstruction unit iteratively trains the FSRCNN model based on the dataset from the data acquisition unit, using the Adam algorithm for optimization. Training stops when the loss function MSE reaches its minimum or the required number of iterations is reached. The loss function MSE measures the error between the super-resolution reconstructed image and the original high-resolution image, and its calculation formula is as follows: in, These are the pixel values ​​in the original image. These are the pixel values ​​in the super-resolution reconstructed image predicted by the FSRCNN model. n This represents the number of samples in the dataset.

Images