Security image feature enhancement model training method and security image processing method

Abstract

The embodiment of the application provides a security check image feature enhancement model training method and a security check image processing method. The security check image feature enhancement model comprises an image encoder, an image generation module and an image decoder. A first sample projection image and a text label of the first sample projection image are obtained. The image encoder is trained by using the first sample projection image and the text label. A second sample projection image and a picture label of the second sample projection image are obtained. The second sample projection image is input into the image encoder, and a second sample image encoding result is output. The image generation module and the image decoder are trained based on the second sample image encoding result and the picture label of the second sample projection image. The image encoder is first trained, and then the trained image encoder is used to train the image generation module and the image decoder, so as to obtain a trained security check image feature enhancement model. The model has a good optimization effect on image quality.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a training method for a security inspection image feature enhancement model and a security inspection image processing method. Background Technology

[0002] Millimeter-wave imaging technology is frequently used in security checks at airports, customs, and rail transit. However, millimeter-wave imaging is prone to unavoidable coherence patterns, which significantly reduce image quality and severely interfere with the detection of contraband. Current image denoising algorithms fall into two categories: one uses filters from digital signal processing to suppress noise and optimize image quality; the other uses convolutional neural networks, such as CNNs, for denoising and image quality optimization. While both methods effectively suppress global noise, they cannot remove irregular image distortions caused by the imaging principle itself, such as coherence patterns, resulting in poor image quality and an inability to accurately detect contraband. Summary of the Invention

[0003] The purpose of this application is to provide a training method for a security inspection image feature enhancement model and a security inspection image processing method for feature enhancement of millimeter-wave imaging. The specific technical solution is as follows:

[0004] In a first aspect, embodiments of this application provide a training method for a security inspection image feature enhancement model, wherein the security inspection image feature enhancement model includes an image encoder, an image generation module, and an image decoder, and the method includes:

[0005] Obtain a first sample projection image and a text label for the first sample projection image, wherein the first sample projection image is obtained by projecting first sample millimeter-wave three-dimensional point cloud data, and the text label is used to describe information about contraband in the first sample projection image;

[0006] The image encoder is trained using the first sample projection image and the text label. The image encoder is used to perform dimensionality reduction processing on the input image and output the dimensionality-reduced encoded image corresponding to the input image. The dimensionality reduction processing includes image feature dimensionality reduction processing and / or image channel dimensionality reduction processing.

[0007] Obtain a second sample projection image and an image label for the second sample projection image, wherein the second sample projection image is obtained by projecting second sample millimeter-wave 3D point cloud data, and the image label is a feature-enhanced sample image corresponding to the second sample projection image;

[0008] The second sample projection image is input into the image encoder, and the second sample image encoding result is output.

[0009] Based on the encoding result of the second sample image and the image label of the second sample projected image, the image generation module and the image decoder are trained; wherein, the image generation module is used to perform feature enhancement processing on the input encoded image, and the image decoder is used to decode the encoded image output by the generation module to obtain the feature-enhanced image.

[0010] In one possible implementation, training the image encoder using the first sample projection image and the text label includes:

[0011] Obtain the label vector encoder;

[0012] The first sample projection image is input into the image encoder to obtain the first sample image encoding result;

[0013] The text labels are vectorized, and the vectorized text labels are input into the label vector encoder to obtain the text encoding result;

[0014] The parameters of the image encoder are adjusted based on the text encoding result and the encoding result of the first sample image.

[0015] In one possible implementation, adjusting the parameters of the image encoder based on the text encoding result and the first sample image encoding result includes:

[0016] Calculate the semantic similarity between the text encoding result and the encoding result of the first sample image to obtain the first loss;

[0017] The parameters of the image encoder are adjusted based on the first loss.

[0018] In one possible implementation, the image generation module and image decoder are trained based on the encoding result of the second sample image and the image label of the second sample projected image, including:

[0019] The encoding result of the second sample image is input into the image generation module to obtain the encoding result of the predicted sample image;

[0020] The encoding result of the predicted sample image is input into the image decoder to obtain the enhanced predicted image of the second sample projection image;

[0021] Based on the feature difference between the enhanced predicted image and the image label, the parameters of the image generation module and the image decoder are adjusted.

[0022] In one possible implementation, the image generation module includes a diffusion sub-model, and the step of inputting the encoding result of the second sample image into the image generation module to obtain the encoding result of the predicted sample image includes:

[0023] The diffusion sub-model is subjected to autoregressive training to obtain a pre-trained diffusion sub-model;

[0024] Using the pre-trained diffusion sub-model, the encoding result of the second sample image is denoised to obtain the encoding result of the predicted sample image.

[0025] In one possible implementation, the parameters of the image generation module and the image decoder are adjusted based on the feature difference between the enhanced predicted image and the image label, including:

[0026] Determine the second pixel point corresponding to the first pixel point on the image label at the same position on the image label, and calculate the difference between the pixel values ​​of the first pixel point and the second pixel point;

[0027] Based on the difference in pixel values ​​between the first pixel and the second pixel, the feature difference between the enhanced prediction image and the image label is determined.

[0028] Based on the feature difference, the parameters of the image generation module and the image decoder are adjusted.

[0029] Secondly, embodiments of this application provide a security inspection image processing method, which is implemented based on a security inspection image feature enhancement model. The security inspection image feature enhancement model is implemented based on any of the training methods described above. The security inspection image processing method includes:

[0030] Acquire millimeter-wave images of human bodies for security checks;

[0031] The millimeter-wave human body security inspection image is input into the image encoder in the security inspection image feature enhancement model to obtain security inspection image encoded data;

[0032] The security inspection image encoding data is input into the image generation module in the security inspection image feature enhancement model to obtain enhanced security inspection image encoding data;

[0033] The enhanced security inspection image encoding data is input into the image decoder in the security inspection image feature enhancement model to obtain the feature-enhanced security inspection image corresponding to the millimeter-wave human body security inspection image;

[0034] The enhanced security inspection image is then identified to determine information about prohibited items.

[0035] Thirdly, embodiments of this application provide a training apparatus for a security inspection image feature enhancement model, wherein the security inspection image feature enhancement model includes an image encoder, an image generation module, and an image decoder, and the apparatus includes:

[0036] The first acquisition module is used to acquire a first sample projection image and a text label of the first sample projection image, wherein the first sample projection image is obtained by projecting the first sample millimeter-wave three-dimensional point cloud data, and the text label is used to describe the information of the contraband in the first sample projection image.

[0037] An image encoder training module is used to train the image encoder using the first sample projection image and the text label. The image encoder is used to perform dimensionality reduction processing on the input image and output the dimensionality-reduced encoded image corresponding to the input image. The dimensionality reduction processing includes image feature dimensionality reduction processing and / or image channel dimensionality reduction processing.

[0038] The second acquisition module is used to acquire a second sample projection image and an image label for the second sample projection image, wherein the second sample projection image is obtained by projecting the second sample millimeter-wave three-dimensional point cloud data, and the image label is a feature-enhanced sample image corresponding to the second sample projection image;

[0039] The second sample projection image input module is used to input the second sample projection image into the image encoder and output the second sample image encoding result.

[0040] The training module is used to train the image generation module and the image decoder based on the encoding result of the second sample image and the image label of the second sample projection image; wherein, the image generation module is used to perform feature enhancement processing on the input encoded image, and the image decoder is used to decode the encoded image output by the generation module to obtain the feature-enhanced image.

[0041] In one possible implementation, the image encoder training module includes:

[0042] The tag vector encoder acquisition module is used to acquire the tag vector encoder.

[0043] The first sample image encoding result acquisition module is used to input the first sample projection image into the image encoder to obtain the first sample image encoding result;

[0044] The text encoding result acquisition module is used to vectorize the text label and input the vectorized text label into the label vector encoder to obtain the text encoding result.

[0045] The image encoder training submodule is used to adjust the parameters of the image encoder based on the text encoding result and the encoding result of the first sample image.

[0046] In one possible implementation, the image encoder training submodule includes:

[0047] The loss calculation unit is used to calculate the semantic similarity between the text encoding result and the first sample image encoding result to obtain the first loss;

[0048] An image encoder parameter adjustment unit is used to adjust the parameters of the image encoder according to the first loss.

[0049] In one possible implementation, the training module includes:

[0050] The second sample image encoding result input module is used to input the second sample image encoding result into the image generation module to obtain the predicted sample image encoding result;

[0051] The prediction sample image encoding result input module is used to input the prediction sample image encoding result into the image decoder to obtain the enhanced prediction image of the second sample projection image;

[0052] The training submodule is used to adjust the parameters of the image generation module and the image decoder based on the feature difference between the enhanced predicted image and the image label.

[0053] In one possible implementation, the image generation module includes a diffusion sub-model, and the second sample image encoding result input module includes:

[0054] The diffusion sub-model autoregressive training module is used to perform autoregressive training on the diffusion sub-model to obtain a pre-trained diffusion sub-model.

[0055] The noise reduction module is used to perform noise reduction processing on the second sample image encoding result using the pre-trained diffusion sub-model to obtain the predicted sample image encoding result.

[0056] In one possible implementation, the training submodule includes:

[0057] The pixel pair determination module is used to determine the second pixel point corresponding to the first pixel point on the image label at the same position on the image label, and to calculate the difference between the pixel values ​​of the first pixel point and the second pixel point.

[0058] The feature difference determination module is used to determine the feature difference between the enhanced prediction image and the image label based on the difference between the pixel values ​​of the first pixel and the second pixel.

[0059] The parameter adjustment module is used to adjust the parameters of the image generation module and the image decoder based on the feature difference degree.

[0060] Fourthly, embodiments of this application provide a security inspection image processing device, which is implemented based on a security inspection image feature enhancement model. The security inspection image feature enhancement model is implemented based on any of the training methods described above. The security inspection image processing device includes:

[0061] Millimeter-wave human body security inspection image acquisition module, used to acquire millimeter-wave human body security inspection images;

[0062] A millimeter-wave human body security inspection image input module is used to input the millimeter-wave human body security inspection image into the image encoder in the security inspection image feature enhancement model to obtain security inspection image encoded data;

[0063] The security inspection image encoding data input module is used to input the security inspection image encoding data into the image generation module in the security inspection image feature enhancement model to obtain enhanced security inspection image encoding data.

[0064] An enhanced security inspection image encoding data input module is used to input the enhanced security inspection image encoding data into the image decoder in the security inspection image feature enhancement model to obtain the feature-enhanced security inspection image corresponding to the millimeter-wave human body security inspection image;

[0065] The identification module is used to identify the enhanced security inspection image and determine the information of prohibited items.

[0066] Fifthly, embodiments of this application provide an electronic device, including:

[0067] Memory, used to store computer programs;

[0068] The processor, when executing the program stored in memory, implements the training method for any of the above-mentioned security inspection image feature enhancement models.

[0069] Sixthly, embodiments of this application provide an electronic device, including:

[0070] Memory, used to store computer programs;

[0071] The processor, when executing the program stored in the memory, implements the above-mentioned security inspection image processing method.

[0072] Seventhly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the training method for any of the above-described security inspection image feature enhancement models.

[0073] Eighthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described security inspection image processing method.

[0074] Beneficial effects of the embodiments in this application:

[0075] The training method for a security inspection image feature enhancement model provided in this application includes an image encoder, an image generation module, and an image decoder. The method involves acquiring a first sample projected image and its text labels; training the image encoder using the first sample projected image and its text labels; acquiring a second sample projected image and its image labels; inputting the second sample projected image into the image encoder and outputting its encoding result; and training the image generation module and image decoder based on the encoding result and the image labels of the second sample projected image. By first training the image encoder and then using the trained image encoder to train the image generation module and image decoder, a well-trained security inspection image feature enhancement model is obtained, which exhibits good optimization effects on image quality.

[0076] Of course, implementing any product or method of this application does not necessarily require achieving all of the advantages described above at the same time. Attached Figure Description

[0077] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other embodiments can be obtained based on these drawings.

[0078] Figure 1 A training method for a security inspection image feature enhancement model provided in this application embodiment;

[0079] Figure 2 Another training method for the security inspection image feature enhancement model provided in the embodiments of this application;

[0080] Figure 3 This is a schematic diagram of the network structure during the noise reduction process;

[0081] Figure 4 This is a schematic diagram of a first flowchart of a security inspection image processing method provided in an embodiment of this application;

[0082] Figure 5 This is a schematic diagram of a second flowchart of the security inspection image processing method provided in the embodiments of this application;

[0083] Figure 6 This is a schematic diagram of a third process for a security inspection image processing method provided in an embodiment of this application;

[0084] Figure 7 This is a schematic diagram comparing traditional object detection methods with the object detection method proposed in this application;

[0085] Figure 8 A schematic diagram of a training device for a security inspection image feature enhancement model provided in an embodiment of this application;

[0086] Figure 9 A schematic diagram of the security inspection image processing device provided in the embodiments of this application;

[0087] Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0088] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art based on this application are within the scope of protection of this application.

[0089] First, the terms used in this application will be explained.

[0090] Millimeter waves: electromagnetic waves with wavelengths in the range of 1-10mm, corresponding to frequencies in the range of 30-300GHz, which can penetrate clothing of a certain thickness when used for human security checks.

[0091] Generative models: a type of deep learning model;

[0092] Diffusion model (DM): A deep learning generative model that can learn the probability distribution of samples and generate information such as images and audio based on input guidance information (images, text, signals, etc.).

[0093] Millimeter-wave imaging is a technique that uses the intensity and phase information of the echo generated by millimeter waves acting on the object being measured to create an image. Due to the non-contact and low-radiation characteristics of millimeter waves, it is often used for personnel security checks at airports, customs, and rail transit. Compared to transmission imaging, such as X-ray imaging, millimeter-wave imaging quality is inferior. Furthermore, millimeter-wave imaging of the human body is easily affected by numerous factors, such as clothing wrinkles, the subject's posture, and other accidental factors. Additionally, due to the inherent characteristics of millimeter-wave imaging technology, unavoidable coherence patterns are prone to occur. These interference factors significantly reduce image quality and severely interfere with the detection of contraband. Currently, optimizing image quality often removes features of contraband along with the image, hindering its identification and detection. Moreover, existing image denoising algorithms fall into two categories: one uses filters from digital signal processing to suppress noise and optimize image quality; the other uses convolutional neural networks, such as CNNs, for denoising and image quality optimization. While both methods effectively suppress global noise, they cannot remove irregular image distortions caused by the imaging principle itself, such as coherence patterns, resulting in poor image quality. Even with poor-quality images, accurate detection and identification of contraband is impossible.

[0094] To optimize millimeter-wave imaging, in the first aspect of this application, a training method for a security inspection image feature enhancement model is provided, see [link to relevant documentation]. Figure 1 The security inspection image feature enhancement model includes an image encoder, an image generation module, and an image decoder, and the method includes:

[0095] S101. Obtain the first sample projection image and the text label of the first sample projection image.

[0096] The first sample projected image is obtained by projecting first sample millimeter-wave 3D point cloud data. The 3D point cloud data generally includes the following attribute information: 3D spatial coordinates of the point cloud (X, Y, Z); laser reflection intensity value (indicating the density of points); color values ​​(R, G, B); transparency; and the normal vector of the point cloud (nx, ny, nz). The first sample projected image is an image of the personnel undergoing security checks. Text labels are used to describe information about contraband in the first sample projected image, such as describing the person being checked and the contraband in the first sample projected image. In one example, the text label may include textual descriptions such as the height of the person being checked, the location of the contraband, and the type of the contraband.

[0097] S102. Using the first sample projection image and the text label, train the image encoder.

[0098] The image encoder is used to perform dimensionality reduction processing on the input image and output the dimensionality-reduced encoded image corresponding to the input image. The dimensionality reduction processing includes image feature dimensionality reduction processing and / or image channel dimensionality reduction processing. For example, a three-channel image with depth information and a size of 640*640 is reduced to a single-channel image with a size of 320*320. The image encoder is used to encode the first sample projection image, thereby mapping the first sample projection image to a low-dimensional space. Exemplarily, the image encoder in this embodiment can adopt a ResNet structure.

[0099] S103. Obtain the second sample projection image and the image label of the second sample projection image.

[0100] The second sample projected image may be the same as or different from the first sample projected image. The second sample projected image is obtained by projecting second sample millimeter-wave 3D point cloud data. The 3D point cloud data generally includes the following attribute information: 3D spatial coordinates of the point cloud (X, Y, Z); laser reflection intensity value (indicating the density of points); color values ​​(R, G, B); transparency; and the normal vector of the point cloud (nx, ny, nz). The second sample projected image is an image of a person undergoing security screening, and the image label is the feature-enhanced sample image corresponding to the second sample projected image. In this embodiment, for example, image feature enhancement may include at least one of the following operations: removing coherent speckles, removing artifacts, removing noise, enhancing image details, and enhancing image grayscale. For example, feature enhancement may include enhancing edge details of the image using the Canny operator, denoising the image using the BM3D algorithm, and enhancing the grayscale of the region of interest by stretching the local grayscale histogram. One possible implementation is to manually enhance the features of the second sample projection image, for example, by using a mean filter, adaptive Wiener filter, or median filter to process the second sample projection image to obtain a feature-enhanced second sample projection image. Alternatively, the location of the contraband can be manually marked, and the remaining area can be smoothed to obtain a feature-enhanced second sample projection image.

[0101] S104. Input the second sample projection image into the image encoder and output the second sample image encoding result.

[0102] The image encoder is used to encode the second sample projection image, thereby mapping the second sample projection image to a low-dimensional space and obtaining the second sample image encoding result.

[0103] S105. Based on the encoding result of the second sample image and the image label of the second sample projection image, train the image generation module and the image decoder.

[0104] In this embodiment, an image encoder is first trained, and then the trained image encoder is used to train an image generation module and an image decoder. The image generation module performs feature enhancement processing on the input encoded image, and the image decoder decodes the encoded image output by the generation module to obtain a feature-enhanced image. For example, image feature enhancement may include at least one of the following operations: removing coherent speckles, removing artifacts, removing noise, enhancing image details, enhancing image grayscale, etc., thereby achieving image quality optimization.

[0105] As can be seen, the training method of the security inspection image feature enhancement model provided in this application can train an image encoder based on the first sample projection image and the text label of the first sample projection image, and then use the trained image encoder to encode the second sample projection image to obtain the encoding result of the second sample image; based on the encoding result of the second sample image and the image label of the second sample projection image, an image generation module and an image decoder are trained. When the security inspection image feature enhancement model including the image encoder, the image generation module and the image decoder is applied, the image optimization effect can be achieved.

[0106] In one example, step S102 above may include the following steps:

[0107] Step 1: Obtain the label vector encoder;

[0108] A label vector encoder can be a pre-trained encoder obtained from an existing neural network model, used to encode input vectors (such as word vectors); or it can be an encoder co-trained with an image encoder. The label vector encoder employs fully connected layers with self-attention.

[0109] Step 2: Input the first sample projection image into the image encoder to obtain the first sample image encoding result;

[0110] An image encoder is used to encode the first sample projection image, thereby mapping the first sample projection image to a low-dimensional space. This significantly reduces computational cost; for convolutional networks, the computational cost is approximately proportional to the size of the input features. Furthermore, it can remove coherent patterns and other noise from the first sample projection image. In this embodiment, the image encoder can employ a ResNet structure.

[0111] Step 3: Vectorize the text labels and input the vectorized text labels into the label vector encoder to obtain the text encoding result;

[0112] The text tags are vectorized to obtain a one-dimensional tag vector (e.g., 001111001010...). The tag vector encoder encodes the one-dimensional tag vector to obtain the text encoding result.

[0113] Step 4: Adjust the parameters of the image encoder based on the text encoding result and the encoding result of the first sample image.

[0114] Specifically, the parameters of the image encoder can be adjusted based on the semantic difference or similarity between the text encoding result corresponding to the text label and the first sample image encoding result corresponding to the first sample projection image, so that the semantic similarity between the first sample image encoding result output by the image encoder after parameter adjustment and the text encoding result gradually approaches.

[0115] For example, the semantic similarity between the text encoding result and the encoding result of the first sample image can be calculated to obtain a first loss; the parameters of the image encoder can be adjusted according to the first loss.

[0116] Regarding the calculation of semantic similarity, this application does not impose specific limitations on the embodiments. For example, cosine similarity can be used as a method for calculating semantic similarity. Cosine similarity calculation can reduce the influence of the encoding value on the similarity and focus more on the semantic (directional) distance. In one example, the first loss can be obtained by calculating the cosine similarity between the text encoding result and the encoding result of the first sample image. In addition, similarity calculation methods such as Euclidean distance and Manhattan distance can also be used. It is understood that the convergence of the image encoder is determined based on the first loss. If it converges, it means that the image encoder has been trained well; if it does not converge, the parameters of the image encoder need to be adjusted and training continues until convergence. In one case, the number of training iterations can be preset. When the number of times "the parameters of the image encoder are adjusted according to the text encoding result and the encoding result of the first sample image" reaches the preset number of training iterations, the trained image encoder is obtained.

[0117] When calculating the similarity between the text encoding result and the first sample image encoding result, the first sample image encoding result can be one-dimensional (for example, the first sample image encoding result with an encoding size of n*n can be one-dimensionalized to n*n*1). The advantage of doing this is that it is convenient to calculate the similarity with the text encoding result, and the input to the image generation module is still the sample image encoding result before one-dimensionalization.

[0118] In this embodiment, an image encoder is trained using a pre-trained label vector encoder, making the image encoding result as close as possible to the text encoding result. This allows for manual control over the content to be focused on during the image encoding process. Furthermore, using the text encoding result as guidance to determine the image encoding result can also remove unwanted noise from the image.

[0119] In one example, see Figure 2 Step S105 above may include:

[0120] S201. Input the second sample image encoding result into the image generation module to obtain the predicted sample image encoding result.

[0121] The image generation module includes a diffusion sub-model.

[0122] The encoding result of the second sample image is input into the image generation module to obtain the encoding result of the predicted sample image, including:

[0123] Step 1: Perform autoregressive training on the diffusion sub-model to obtain the pre-trained diffusion sub-model.

[0124] Step 2: Using the pre-trained diffusion sub-model, the encoding result of the second sample image is denoised to obtain the encoding result of the predicted sample image.

[0125] Autoregressive training refers to using past information to predict current information; that is, using itself to predict itself. The diffusion sub-model is an autoregressive model that does not require labels during training. A trained diffusion sub-model can have its input effect controlled by explicitly controlling the number of iterations. Other models, such as VAEs and GANs, require retraining. When deploying a diffusion sub-model, the number of iterations can be dynamically selected to meet different needs. In extreme cases, it can be iterated only once, in which case the diffusion sub-model will degenerate into a simple denoising model, still achieving the relevant functionality. The diffusion sub-model is based on DDPM (Denoising Diffusion Probabilistic Model), offering higher interpretability than other generative models for denoising tasks. The diffusion sub-model can adapt to different computing platforms.

[0126] Autoregressive training of the diffusion sub-model includes two processes: forward noise addition and backward denoising. The forward and backward processes of the diffusion sub-model constitute autoregressive training. The forward process (adding noise to the original image) generates a "regression target" image corresponding to the original image. This "regression target" image is then input into the backward denoising diffusion sub-model to fit the noise distribution, thus obtaining a fitting result similar to the original image. This can be simplified into the following formula:

[0127] 1. Original input x, forward process y = g(x), y represents the result after adding noise;

[0128] 2. Autoregression result x', reverse process x' = r(y);

[0129] 3. The complete autoregressive process: x' = r(g(x));

[0130] The following is a detailed explanation.

[0131] First, during the forward noise addition process, after acquiring a batch of training samples, the training samples are preprocessed. Specifically, Gaussian noise is added to the training samples a predetermined number of times until the training samples become random noise. The training samples can be formed by downsampling 3D point cloud data. Gaussian noise is added after downsampling the 3D point cloud until the original signal is submerged. Specifically, Gaussian noise is added to the training samples using the following formula:

[0132]

[0133] q(x t |x t-1 This indicates that each diffusion process is the result of x obtained in the previous step. t-1 The process of adding Gaussian noise, where N() represents a Gaussian distribution, x t This represents an image with t-th order Gaussian noise added according to the above formula, β t This represents the variance of each noise addition; β1 < β2 < ... < β T In other words, the variance of noise increases with each subsequent addition. This characteristic can be understood through the reverse denoising process, where obvious noise is removed first, followed by denoising of subtle noise. In this embodiment, Gaussian noise is added to the training samples a preset number of times to obtain a noisy training sample that completely drowns out the original signal. The entire noise addition process can be viewed as a Markov chain:

[0134]

[0135] x0 represents the training sample, x1 represents the result of adding Gaussian noise to x0 once, and T represents the number of times Gaussian noise is added.

[0136] Secondly, inverse denoising can be viewed as the reverse process of forward noise addition. Since the input data in inverse denoising is obtained from the result of the forward noise addition process, the true distribution of the samples and the Gaussian noise added each time can be determined. The inverse denoising process is a multi-stage inference process, learning the distribution pattern of the real data layer by layer. It can recover the un-noiseed training samples from the noisy samples. Specifically, the following formula is used to perform inverse denoising on the noisy training samples.

[0137]

[0138] p θ (x t-1 |x t This is the inverse denoising process, which mathematically is considered to follow a Gaussian distribution, but with a mean μ. (θ) (x t ,t) and variance ∑ θ (x t (t) requires a neural network to combine with the input x t (The result of step t in the forward process) is used for fitting, therefore p θ (x t-1 |x t The distribution is a parameterized Gaussian distribution, with mean and variance obtained through network training. The specific training process includes: processing the noisy training samples q(x) obtained in the previous step... t |x t-1 As the input to the network, since the variance of the noise added each time is different, the number of times the noise is added, t, also needs to be used as input. See [link to relevant documentation]. Figure 3 The purpose of this is to provide the network with prior knowledge, using the corresponding stage's 't' as guidance information for training at different stages. Since the noise added during the noise-adding process is known, the network can be trained based on the known noise and the noisy training samples, and the network parameters can be adjusted based on the error between the denoised training samples predicted by the network and the known noise. Time-embedding refers to temporal embedding; Unet with timeembedding means that the Unet network is implemented using temporal embedding techniques.

[0139] The reverse denoising process can also be viewed as a Markov chain.

[0140]

[0141] p θ (x 0:T This describes the entire reverse denoising process: the sum of steps from step 0 to step T; p θ (x t-1 |x t) is a parameterized Gaussian distribution. Due to the lack of aftereffect in Markov chains, the entire process can be described by a product of the conditional probabilities of each step; p(x T )=N(x t N(.) represents a Gaussian distribution, where 0, I are the vectorized mean and variance.

[0142] In one example, suppose the original image is The image after adding Gaussian noise for the tth time is They satisfy the following:

[0143]

[0144] The forward process involves gradually adding Gaussian noise. This time-varying distribution then forms a Markov chain. Based on its stationarity, we know that if a probability distribution changes over time, under the influence of a Markov chain, it will eventually tend towards a stationary distribution, which in this case is a Gaussian distribution. Therefore:

[0145] The reverse process uses Unet to simulate the transition probability distribution generated in the reverse process: q(x t-1 |x t ).

[0146] Once the diffusion sub-model has been trained, it can be used to denoise the encoding results of the second sample image to obtain the encoding results of the predicted sample image.

[0147] S202. Input the encoding result of the predicted sample image into the image decoder to obtain the enhanced predicted image of the second sample projection image.

[0148] Since the previous operations were all performed in low-dimensional space, the image needs to be restored to obtain a final output. The encoded result of the predicted sample image output by the diffusion sub-model is input into the image decoder to obtain the enhanced predicted image corresponding to the second sample projection image.

[0149] S203. Based on the feature difference between the enhanced predicted image and the image label, adjust the parameters of the image generation module and the image decoder.

[0150] The image label is the feature-enhanced sample image corresponding to the second sample projection image. A simple alternative implementation is to manually mark the location of the prohibited items in the projection image, and then smooth the remaining area to obtain the image label. Based on the error (i.e., feature difference) between the enhanced predicted image and the image label, the parameters of the image generation module and the image decoder are adjusted in reverse.

[0151] If the error between the enhanced predicted image and the image label is less than a preset threshold, the image generation module and the image decoder are considered to have converged. If the error between the denoised predicted image and the image label is not less than the preset threshold, the image generation module and the image decoder are considered to have not converged. At this time, the parameters of the diffusion model module and the image decoder are adjusted, and iterative training is performed until convergence.

[0152] In one example, step S203 can be implemented as follows:

[0153] S301. Determine the second pixel point corresponding to the first pixel point on the enhanced prediction image at the same position on the image label, and calculate the difference in pixel values ​​between the first pixel point and the second pixel point.

[0154] S302. Based on the difference in pixel values ​​between the first pixel and the second pixel, determine the feature difference between the enhanced prediction image and the image label.

[0155] In other words, the feature difference between the enhanced predicted image and the image label can be determined based on the difference in pixel values ​​between pixels. For example, for each pair of pixels (i.e., the aforementioned first pixel and its corresponding second pixel), the difference in pixel values ​​between the first pixel and the second pixel is calculated. All the differences are squared, summed, and then divided by the number of pixel pairs. The resulting value is used as the feature difference, which can be used to describe the similarity between the enhanced predicted image and the image label. The smaller the feature difference, the higher the similarity between the enhanced predicted image and the image label.

[0156] S303. Based on the feature difference degree, adjust the parameters of the image generation module and the image decoder.

[0157] Adjust the parameters of the image generation module and the image decoder based on the feature difference, for example, find the parameters of the image generation module and the image decoder that minimize the feature difference.

[0158] In a second aspect, this application provides a security inspection image processing method. This method is based on a security inspection image feature enhancement model, which is implemented using any of the training methods described above. For details not explained in the following embodiments, please refer to the relevant descriptions in the above embodiments. Figure 4 The security inspection image processing method includes:

[0159] S401. Acquire millimeter-wave human body security inspection images.

[0160] Millimeter-wave body scan images are images obtained by scanning and detecting the human body using millimeter-wave technology. They contain information about the human body and information about prohibited items carried by the person.

[0161] S402. Input the millimeter-wave human body security inspection image into the image encoder in the security inspection image feature enhancement model to obtain security inspection image encoded data.

[0162] Image encoders are used for encoding millimeter-wave human body security inspection images, mapping millimeter-wave human body security inspection images to a low-dimensional space.

[0163] S403. Input the security inspection image encoding data into the image generation module in the security inspection image feature enhancement model to obtain enhanced security inspection image encoding data.

[0164] S404. Input the enhanced security inspection image encoding data into the image decoder in the security inspection image feature enhancement model to obtain the feature-enhanced security inspection image corresponding to the millimeter-wave human body security inspection image.

[0165] The image decoder is used to restore the image. After processing, the output is a feature-enhanced security image corresponding to the millimeter-wave human body security inspection image. The feature-enhanced security image is an image that has undergone feature enhancement relative to the millimeter-wave human body security inspection image, such as removing coherence speckles, artifacts, and noise, as well as enhancing image details and grayscale. The feature-enhanced security image is clearer and has better imaging quality, and can be used for subsequent processing such as target detection. Target detection using the feature-enhanced security image has better recognition results.

[0166] S405. Identify the enhanced security inspection image to determine the information on prohibited items.

[0167] In one example, see Figure 5 Here is a flowchart illustrating a security inspection image feature enhancement method provided in this application embodiment:

[0168] The process involves collecting 3D point cloud data and labels, where labels refer to text labels, such as the height of the person undergoing security checks, the location of prohibited items, and a textual description of the prohibited items. Based on the 3D point cloud data, a back-projection imaging technique is used to form an original image. This original image is then preprocessed, including manually annotating the locations of prohibited items and smoothing other areas of the original image to obtain a target image, which is the feature-enhanced version of the original image. The original labels are then vectorized to obtain original label vectors. Based on these vectors, the original image, and the target image, an original dataset is constructed. With the original dataset, a generative model is set up. Before training the security image feature enhancement model, the diffusion sub-model needs to be autoregressively trained to obtain a pre-trained diffusion sub-model. The autoregressive training of the diffusion sub-model includes a forward addition of Gaussian noise. The security image feature enhancement model updates its parameters based on error propagation. The training of the generative model involves three parts: "setting up the generative model," "image optimization model," and "training the generative model." The image optimization model includes the training processes for the image encoder, image generation module, and image decoder. After the security inspection image feature enhancement model is trained, it is predicted and output as an image. Here, the output image refers to an image with reduced coherence speckles, artifacts, and noise.

[0169] In one example, see Figure 6 Here is another flowchart illustrating the security inspection image feature enhancement method provided in this application embodiment:

[0170] Word vectors are input into the label vector encoder, which is obtained by vectorizing the original labels, to obtain the text encoding result. The original image is input into the original image encoder, which is obtained by projecting 3D point cloud data, to obtain the image encoding result. The cosine similarity between the text encoding result and the image encoding result is calculated to obtain the first loss, and the original image encoder is trained based on the first loss. After the original image encoder is trained, a new original image can be input into the original image encoder to obtain a new image encoding result; the new image encoding result is the text encoding result. Figure 6 In the image encoding process, x0 to x1 requires adding Gaussian noise, x1 to x2 requires adding Gaussian noise, and so on. The final result z can be considered as pure noise data. Deducing x0 from z is a reverse denoising process; that is, a predicted image encoding result can be obtained from the new image encoding result. Finally, the decoder decodes the predicted image encoding result and outputs the target image, which refers to an image with reduced coherence speckles, artifacts, and noise.

[0171] See Figure 7 The diagram below illustrates a comparison between traditional image denoising methods and the image denoising method proposed in this application.

[0172] In traditional image denoising methods, the raw data received by the antenna is processed by an imaging algorithm to obtain a 3D point cloud. The original image is then projected onto this point cloud, and image processing is performed to obtain a high-resolution image. Target detection is then performed on this high-resolution image to obtain the location information of contraband and the human body contour map. The human body contour map and location information are then output. Therefore, traditional image denoising methods involve processing the original image to obtain a high-resolution image, and then performing human body contour and contraband location detection on this high-resolution image.

[0173] In the image denoising method proposed in this application, the raw data received by the antenna is processed by an imaging algorithm to obtain a three-dimensional point cloud. The original image is obtained by projection based on the three-dimensional point cloud. Diffusion image processing is performed on the original image to obtain a clear approximate human body contour map. Target detection is performed based on the clear approximate human body contour map to obtain the output result.

[0174] In the third aspect of this application, see Figure 8 A training device for a security inspection image feature enhancement model is provided. The security inspection image feature enhancement model includes an image encoder, an image generation module, and an image decoder. The device includes:

[0175] The first acquisition module 501 is used to acquire a first sample projection image and a text label of the first sample projection image, wherein the first sample projection image is obtained by projecting the first sample millimeter-wave three-dimensional point cloud data, and the text label is used to describe the information of the contraband in the first sample projection image.

[0176] The image encoder training module 502 is used to train the image encoder using the first sample projection image and the text label. The image encoder is used to perform dimensionality reduction processing on the input image and output the dimensionality-reduced encoded image corresponding to the input image. The dimensionality reduction processing includes image feature dimensionality reduction processing and / or image channel dimensionality reduction processing.

[0177] The second acquisition module 503 is used to acquire a second sample projection image and an image label for the second sample projection image, wherein the second sample projection image is obtained by projecting the second sample millimeter-wave three-dimensional point cloud data, and the image label is a feature-enhanced sample image corresponding to the second sample projection image.

[0178] The second sample projection image input module 504 is used to input the second sample projection image into the image encoder and output the second sample image encoding result.

[0179] Training module 505 is used to train the image generation module and image decoder based on the encoding result of the second sample image and the image label of the second sample projection image; wherein, the image generation module is used to perform feature enhancement processing on the input encoded image, and the image decoder is used to decode the encoded image output by the generation module to obtain the feature-enhanced image.

[0180] In one example, the image encoder training module includes:

[0181] The tag vector encoder acquisition module is used to acquire the tag vector encoder.

[0182] The first sample image encoding result acquisition module is used to input the first sample projection image into the image encoder to obtain the first sample image encoding result;

[0183] The text encoding result acquisition module is used to vectorize the text label and input the vectorized text label into the label vector encoder to obtain the text encoding result.

[0184] The image encoder training submodule is used to adjust the parameters of the image encoder based on the text encoding result and the encoding result of the first sample image.

[0185] In one example, the image encoder training submodule includes:

[0186] The loss calculation unit is used to calculate the semantic similarity between the text encoding result and the first sample image encoding result to obtain the first loss;

[0187] An image encoder parameter adjustment unit is used to adjust the parameters of the image encoder according to the first loss.

[0188] In one example, the training module includes:

[0189] The second sample image encoding result input module is used to input the second sample image encoding result into the image generation module to obtain the predicted sample image encoding result;

[0190] The prediction sample image encoding result input module is used to input the prediction sample image encoding result into the image decoder to obtain the enhanced prediction image of the second sample projection image;

[0191] The training submodule is used to adjust the parameters of the image generation module and the image decoder based on the feature difference between the enhanced predicted image and the image label.

[0192] In one example, the image generation module includes a diffusion sub-model, and the second sample image encoding result input module includes:

[0193] The diffusion sub-model autoregressive training module is used to perform autoregressive training on the diffusion sub-model to obtain a pre-trained diffusion sub-model.

[0194] The noise reduction module is used to perform noise reduction processing on the second sample image encoding result using the pre-trained diffusion sub-model to obtain the predicted sample image encoding result.

[0195] In one example, the training submodule includes:

[0196] The pixel pair determination module is used to determine the second pixel point corresponding to the first pixel point on the image label at the same position on the image label, and to calculate the difference between the pixel values ​​of the first pixel point and the second pixel point.

[0197] The feature difference determination module is used to determine the feature difference between the enhanced prediction image and the image label based on the difference between the pixel values ​​of the first pixel and the second pixel.

[0198] The parameter adjustment module is used to adjust the parameters of the image generation module and the image decoder based on the feature difference degree.

[0199] In the fourth aspect of this application, see [link to application]. Figure 9 A security inspection image processing device is provided. The device is based on a security inspection image feature enhancement model, which is implemented using any of the training methods described above. The security inspection image processing device includes:

[0200] Millimeter-wave human body security inspection image acquisition module 601, used to acquire millimeter-wave human body security inspection images;

[0201] The millimeter-wave human body security inspection image input module 602 is used to input the millimeter-wave human body security inspection image into the image encoder in the security inspection image feature enhancement model to obtain security inspection image encoded data;

[0202] The security inspection image encoding data input module 603 is used to input the security inspection image encoding data into the image generation module in the security inspection image feature enhancement model to obtain enhanced security inspection image encoding data.

[0203] The enhanced security inspection image encoding data input module 604 is used to input the enhanced security inspection image encoding data into the image decoder in the security inspection image feature enhancement model to obtain the feature-enhanced security inspection image corresponding to the millimeter wave human body security inspection image;

[0204] The identification module 605 is used to identify the enhanced security inspection image and determine the information of prohibited items.

[0205] In the technical solution of this application, the operations of obtaining, storing, using, processing, transmitting, providing and disclosing user personal information are all carried out with the user's authorization.

[0206] This application also provides an electronic device, such as... Figure 10 As shown, it includes:

[0207] Memory 701 is used to store computer programs;

[0208] The processor 702, when executing the program stored in the memory 701, implements the training method of any of the security inspection image feature enhancement models described above.

[0209] Furthermore, the aforementioned electronic device may also include a communication bus and / or a communication interface, with the processor 702, the communication interface, and the memory 701 communicating with each other via the communication bus.

[0210] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0211] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0212] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0213] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0214] In another embodiment provided in this application, a computer-readable storage medium is also provided, which stores a computer program that, when executed by a processor, implements the training method for any of the above-described security inspection image feature enhancement models.

[0215] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute the training method of any of the security inspection image feature enhancement models in the above embodiments.

[0216] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or other media (e.g., solid state disk (SSD)).

[0217] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0218] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the apparatus embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0219] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application are included within the scope of protection of this application.

Claims

1. A training method for a security inspection image feature enhancement model, characterized in that, The security inspection image feature enhancement model includes an image encoder, an image generation module, and an image decoder; the method includes: Obtain a first sample projection image and a text label for the first sample projection image, wherein the first sample projection image is obtained by projecting first sample millimeter-wave three-dimensional point cloud data, and the text label is used to describe information about contraband in the first sample projection image; Obtain a label vector encoder; input the first sample projection image into the image encoder to obtain the first sample image encoding result; vectorize the text label and input the vectorized text label into the label vector encoder to obtain the text encoding result; adjust the parameters of the image encoder according to the text encoding result and the first sample image encoding result, wherein the image encoder is used to perform dimensionality reduction processing on the input image and output the dimensionality reduction encoded image corresponding to the input image, the dimensionality reduction processing includes image feature dimensionality reduction processing and / or image channel dimensionality reduction processing; Obtain a second sample projection image and an image label for the second sample projection image, wherein the second sample projection image is obtained by projecting second sample millimeter-wave 3D point cloud data, and the image label is a feature-enhanced sample image corresponding to the second sample projection image; The second sample projection image is input into the image encoder, and the second sample image encoding result is output. Based on the encoding result of the second sample image and the image label of the second sample projected image, the image generation module and the image decoder are trained; wherein, the image generation module is used to perform feature enhancement processing on the input encoded image, and the image decoder is used to decode the encoded image output by the image generation module to obtain the feature-enhanced image.

2. The method according to claim 1, characterized in that, Adjusting the parameters of the image encoder based on the text encoding result and the encoding result of the first sample image includes: Calculate the semantic similarity between the text encoding result and the encoding result of the first sample image to obtain the first loss; The parameters of the image encoder are adjusted based on the first loss.

3. The method according to claim 1, characterized in that, Based on the encoding result of the second sample image and the image label of the second sample projected image, the image generation module and image decoder are trained, including: The encoding result of the second sample image is input into the image generation module to obtain the encoding result of the predicted sample image; The encoding result of the predicted sample image is input into the image decoder to obtain the enhanced predicted image of the second sample projection image; Based on the feature difference between the enhanced predicted image and the image label, the parameters of the image generation module and the image decoder are adjusted.

4. The method according to claim 3, characterized in that, The image generation module includes a diffusion sub-model. The step of inputting the encoding result of the second sample image into the image generation module to obtain the encoding result of the predicted sample image includes: The diffusion sub-model is subjected to autoregressive training to obtain a pre-trained diffusion sub-model; Using the pre-trained diffusion sub-model, the encoding result of the second sample image is denoised to obtain the encoding result of the predicted sample image.

5. The method according to claim 3, characterized in that, Based on the feature difference between the enhanced predicted image and the image label, the parameters of the image generation module and the image decoder are adjusted, including: Determine the second pixel point corresponding to the first pixel point on the image label at the same position on the image label, and calculate the difference between the pixel values ​​of the first pixel point and the second pixel point; Based on the difference in pixel values ​​between the first pixel and the second pixel, the feature difference between the enhanced prediction image and the image label is determined. Based on the feature difference, the parameters of the image generation module and the image decoder are adjusted.

6. A security inspection image processing method, characterized in that, The method is based on a security inspection image feature enhancement model, which is implemented based on any one of the training methods described in claims 1-5. The security inspection image processing method includes: Acquire millimeter-wave images of human bodies for security checks; The millimeter-wave human body security inspection image is input into the image encoder in the security inspection image feature enhancement model to obtain security inspection image encoded data; The security inspection image encoding data is input into the image generation module in the security inspection image feature enhancement model to obtain enhanced security inspection image encoding data; The enhanced security inspection image encoding data is input into the image decoder in the security inspection image feature enhancement model to obtain the feature-enhanced security inspection image corresponding to the millimeter-wave human body security inspection image; The enhanced security inspection image is then identified to determine information about prohibited items.

7. A training device for a security inspection image feature enhancement model, characterized in that, The security inspection image feature enhancement model includes an image encoder, an image generation module, and an image decoder. The device includes: The first acquisition module is used to acquire a first sample projection image and a text label of the first sample projection image, wherein the first sample projection image is obtained by projecting the first sample millimeter-wave three-dimensional point cloud data, and the text label is used to describe the information of the contraband in the first sample projection image. An image encoder training module includes: a label vector encoder acquisition module for acquiring a label vector encoder; a first sample image encoding result acquisition module for inputting the first sample projection image into the image encoder to obtain the first sample image encoding result; a text encoding result acquisition module for vectorizing the text labels and inputting the vectorized text labels into the label vector encoder to obtain the text encoding result; and an image encoder training submodule for adjusting the parameters of the image encoder according to the text encoding result and the first sample image encoding result, wherein the image encoder is used to perform dimensionality reduction processing on the input image and output the dimensionality-reduced encoded image corresponding to the input image, and the dimensionality reduction processing includes image feature dimensionality reduction processing and / or image channel dimensionality reduction processing; The second acquisition module is used to acquire a second sample projection image and an image label for the second sample projection image, wherein the second sample projection image is obtained by projecting the second sample millimeter-wave three-dimensional point cloud data, and the image label is a feature-enhanced sample image corresponding to the second sample projection image; The second sample projection image input module is used to input the second sample projection image into the image encoder and output the second sample image encoding result. The training module is used to train the image generation module and the image decoder based on the encoding result of the second sample image and the image label of the second sample projection image; wherein, the image generation module is used to perform feature enhancement processing on the input encoded image, and the image decoder is used to decode the encoded image output by the image generation module to obtain the feature-enhanced image.

8. A security inspection image processing device, characterized in that, The device is based on a security inspection image feature enhancement model, which is implemented based on any one of the training methods described in claims 1-5. The security inspection image processing device includes: Millimeter-wave human body security inspection image acquisition module, used to acquire millimeter-wave human body security inspection images; A millimeter-wave human body security inspection image input module is used to input the millimeter-wave human body security inspection image into the image encoder in the security inspection image feature enhancement model to obtain security inspection image encoded data; The security inspection image encoding data input module is used to input the security inspection image encoding data into the image generation module in the security inspection image feature enhancement model to obtain enhanced security inspection image encoding data. An enhanced security inspection image encoding data input module is used to input the enhanced security inspection image encoding data into the image decoder in the security inspection image feature enhancement model to obtain the feature-enhanced security inspection image corresponding to the millimeter-wave human body security inspection image; The identification module is used to identify the enhanced security inspection image and determine the information of prohibited items.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the method of any one of claims 1-6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method described in any one of claims 1-6.

Images