A self-supervised reparameterizable terahertz image dangerous goods instance segmentation method

Abstract

The invention discloses a self-supervised renumberable terahertz image dangerous goods instance segmentation method, comprising: performing data enhancement on a terahertz human body concealed dangerous goods security inspection image data set; pre-training a self-supervised learning image mask modeling enhancement model through the enhanced data set, and obtaining encoder parameters of the self-supervised learning image mask modeling enhancement model; using the encoder of the self-supervised learning image mask modeling enhancement model as an initial feature extraction backbone network of an instance segmentation model; migrating encoder parameters, fine-tuning the initial feature extraction backbone network of the instance segmentation model through an unenhanced terahertz human body concealed dangerous goods security inspection image data set, and obtaining a feature extraction backbone network of the instance segmentation model; inputting a human body concealed dangerous goods security inspection image to be segmented into the feature extraction backbone network of the instance segmentation model, extracting multi-scale features, integrating the multi-scale features, and performing dynamic decoupling to obtain dangerous goods detection segmentation results.

Description

Technical Field

[0001] The invention belongs to the field of terahertz image dangerous goods instance segmentation, and in particular relates to a self-supervised reparameterizable terahertz image dangerous goods instance segmentation method. Background Art

[0002] Terahertz is a type of radiation between microwaves and infrared waves, with a wavelength range of , due to its strong penetrability and harmlessness to the human body, it has been widely used. ) uses terahertz rays to perform non-contact scanning of the human body, which can penetrate a variety of materials to detect hidden objects. Compared with traditional metal detectors, It can not only detect explosive liquids and powders, but also display the shape and location of suspicious objects, and has great application potential in the field of security inspection. Most of the early terahertz security inspection image dangerous goods detection technologies were based on manual recognition and image processing solutions. A few solutions using deep learning technology relied on a large amount of high-quality data and could not achieve low-cost, high-precision detection. These problems have had a significant impact on the practicality of terahertz security detection equipment. With the development of deep learning technology and self-supervised learning, the use of self-supervised learning can enhance the feature extraction ability of the model without relying on a large amount of labeled data. In addition, the balance between reasoning speed and accuracy can be achieved based on the structural reparameterization method. Therefore, the development of dangerous goods instance segmentation that can achieve low-cost and high-precision with a small amount of labeled sample data is very helpful to the application and development of terahertz image detection segmentation technology. Summary of the invention

[0003] In order to solve the problems of high cost and low precision of dangerous goods instance segmentation in terahertz images due to poor terahertz image quality and difficulty in sample annotation, the present invention proposes a self-supervised reparameterized terahertz image dangerous goods instance segmentation method, which reduces the segmentation detection training cost of dangerous goods in terahertz images and improves the detection accuracy by combining a self-supervised learning image mask modeling method and a deep learning instance segmentation algorithm.

[0004] To achieve the above object, the present invention provides a self-supervised reparameterized terahertz image dangerous goods instance segmentation method, which specifically includes the following steps:

[0005] Acquire a terahertz human body hidden dangerous goods security inspection image dataset, and perform data enhancement on the terahertz human body hidden dangerous goods security inspection image dataset;

[0006] Pre-training a self-supervised learning image mask modeling enhancement model through the enhanced terahertz human body concealed dangerous goods security inspection image data set, and obtaining encoder parameters of the self-supervised learning image mask modeling enhancement model;

[0007] Using the encoder of the self-supervised learning image mask modeling enhancement model as the initial feature extraction backbone network of the instance segmentation model;

[0008] Migrating the encoder parameters, fine-tuning the initial feature extraction backbone network of the instance segmentation model through the unenhanced terahertz human body concealed dangerous goods security inspection image dataset, and obtaining the feature extraction backbone network of the instance segmentation model;

[0009] The human body concealed dangerous goods security inspection image to be segmented is input into the feature extraction backbone network of the instance segmentation model, multi-scale features are extracted, the multi-scale features are integrated, and dynamically decoupled to obtain dangerous goods detection segmentation results.

[0010] Optionally, obtaining the terahertz human body concealed dangerous goods security inspection image dataset and performing data enhancement on the terahertz human body concealed dangerous goods security inspection image dataset includes:

[0011] Using terahertz imaging equipment, obtain several target images;

[0012] Marking the contours of the dangerous goods in several target images, and converting the coordinates of the contour points to generate label data, to obtain the terahertz human body concealed dangerous goods security inspection image data set;

[0013] Based on The automatic data enhancement method is used to automatically enhance the data of the terahertz human body concealed dangerous goods security inspection image dataset;

[0014] Among them, the The automatic data augmentation method consists of pixel augmentation space and spatial augmentation space.

[0015] Optionally, the self-supervised learning image mask modeling enhancement model is used based on The image mask modeling method can extract features;

[0016] Among them, the The image mask modeling method consists of an encoder and a decoder. The enhanced terahertz human body hidden dangerous goods security inspection image dataset is randomly occluded, and the unoccluded pixels are used as sparse voxels and encoded using sparse convolution. The encoder outputs feature maps of several levels and obtains target level features. The multi-level feature fusion module collects several target level features for optimization and passes them to the decoder for image reconstruction.

[0017] Optionally, before the multi-level feature fusion module collects a plurality of features of the target level for optimization and transmits them to the decoder for image reconstruction, it further includes:

[0018] In each layer, based on the features of the target level, each is adjusted through maximum pooling and upsampling operations. level, to achieve feature size alignment, and use convolution to achieve channel number alignment, that is:

[0019]

[0020] in, , , represent Convolution changes the number of channels to , represents bilinear interpolation, represents the identity mapping, Represents the maximum pooling feature map arrive Resolution, For the Layer The feature map output after the level is aligned is For the The reference height of the layer feature map, For the The reference width of the layer feature map, is the number of encoder output layers.

[0021] Optionally, pre-training the self-supervised learning image mask modeling enhancement model with the enhanced terahertz human body concealed dangerous goods security inspection image data set, and obtaining encoder parameters of the self-supervised learning image mask modeling enhancement model includes:

[0022] The self-supervised learning image mask modeling enhancement model includes a feature refinement fusion module;

[0023] The self-supervised learning image mask modeling enhancement model is used as an encoder for self-supervised learning, and the images in the enhanced terahertz human body concealed dangerous goods security inspection image dataset are fused through the feature refinement fusion module to obtain a fused feature map;

[0024] Fill all blank positions of the sparse feature map, use the decoder to decode and reconstruct the fused feature map, complete the pre-training of the self-supervised learning image mask modeling enhancement model, and then obtain the encoder parameters of the self-supervised learning image mask modeling enhancement model.

[0025] Optionally, the pre-training process is:

[0026] ,

[0027] in, and Respectively represent Fusion mechanism implicit function and fused feature map, Indicates output layer, and Represents mask embedding The padding operation and the feature map after padding, represents the implicit function of successive blocks of the decoder, , , They correspond to feature maps of different scales in the backbone.

[0028] Optionally, migrating the encoder parameters, fine-tuning the initial feature extraction backbone network of the instance segmentation model using the unenhanced terahertz human body concealed dangerous goods security inspection image dataset, and obtaining the feature extraction backbone network of the instance segmentation model includes:

[0029] By migrating the encoder parameters, the unenhanced terahertz human body concealed dangerous goods security inspection image dataset is input into the instance segmentation model, and the initial feature extraction backbone network of the instance segmentation model is fine-tuned and trained using a structure reparameterization module to obtain the feature extraction backbone network of the instance segmentation model;

[0030] The structure reparameterization module is used to convolve and separate the given initial feature map, obtain the first feature map and the second feature map, input the first feature map into the DBB module, fuse the feature map output by the DBB module with the first feature map and the second feature map, and output the target feature map through convolution operation;

[0031] The structure reparameterization module is:

[0032]

[0033]

[0034]

[0035] in, is the feature map, for The feature map output by the module, and represents convolution and Implicit functions of modules, and represent separation and fusion operations respectively, is the output feature map.

[0036] Optionally, fine-tuning the initial feature extraction backbone network of the instance segmentation model further includes: using The function optimizes the loss function, namely:

[0037]

[0038]

[0039] in, is the loss function, is the number of detection layers, is the bounding box regression loss function, is the classification loss function, , is the weight coefficient of the loss function, is the bounding box loss function, is the prediction box loss function.

[0040] Optionally, inputting the human body concealed dangerous goods security inspection image to be segmented into the feature extraction backbone network of the instance segmentation model, extracting multi-scale features, integrating the multi-scale features, and dynamically decoupling them to obtain dangerous goods detection segmentation results includes:

[0041] Inputting the security inspection image of dangerous goods concealed by a human body to be segmented into the instance segmentation model, and extracting multi-scale features through the feature extraction backbone network of the instance segmentation model;

[0042] Integrate the multi-scale features and dynamically decouple them through a dynamic decoupling head module to obtain the dangerous goods detection segmentation result;

[0043] The dangerous goods detection and segmentation results include a dangerous goods detection frame, a dangerous goods category code, a dangerous goods segmentation mask and a result image of the prediction confidence.

[0044] Optionally, the dynamic decoupling head module is used to give a number of feature maps whose sizes are halved in sequence, use convolution to reduce the number of channels to obtain a number of convolved feature maps, and input them into the dynamic attention module, and use the decoupling head to perform dangerous goods feature detection and segmentation;

[0045] The dynamic decoupling head module is:

[0046] , ,

[0047]

[0048]

[0049] in, , , , , , are all feature maps, and , , , , , , Indicated by A continuous The implicit function of the dynamic attention module composed of represents the implicit function of the decoupled detection head, represents the convolution function, Indicates Output feature map of the layer decoupling head, It is the output feature map after being processed by the dynamic attention module.

[0050] The present invention has the following beneficial effects:

[0051] The present invention can enhance the model feature extraction capability under a small amount of sample training through a self-supervised learning image mask modeling method and a multi-level feature refinement and fusion mechanism, thereby reducing data annotation and model training costs; a structural reparameterization module is designed in the feature extraction backbone network, which can improve the reasoning speed and detection accuracy of the instance segmentation model; in addition, a dynamic decoupling head is designed in the detection head, which can effectively reduce the parameter amount of the instance segmentation model and improve the detection accuracy; the present invention is conducive to the application research of terahertz security inspection image dangerous goods instance segmentation technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0053] Figure 1 A flowchart of a self-supervised reparameterizable terahertz image dangerous goods instance segmentation method according to an embodiment of the present invention;

[0054] Figure 2 A schematic diagram of a network architecture for self-supervised learning image mask modeling proposed in an embodiment of the present invention;

[0055] Figure 3 A schematic diagram of the network architecture of the instance segmentation algorithm proposed in an embodiment of the present invention;

[0056] Figure 4 A schematic diagram of the structure reparameterization module architecture proposed in an embodiment of the present invention;

[0057] Figure 5 This is a schematic diagram of the dynamic decoupling head module architecture proposed in an embodiment of the present invention. DETAILED DESCRIPTION

[0058] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0059] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0060] like Figure 1 As shown, this embodiment provides a self-supervised reparameterized terahertz image dangerous goods instance segmentation method, including: firstly, making a terahertz human body hidden dangerous goods security inspection image dataset, then automatically enhancing the dataset to construct another dataset for self-supervised learning; then using the enhanced dataset to perform self-supervised learning image mask modeling, which is used as the backbone of the pre-training instance segmentation model; finally, migrating the parameters of the self-supervised pre-training backbone, fine-tuning the training instance segmentation algorithm on the unenhanced dataset, and obtaining the human body hidden dangerous goods instance segmentation image, the specific implementation steps are as follows:

[0061] Step 1: Create a terahertz human body hidden dangerous goods security inspection image dataset, and automatically enhance the terahertz human body hidden dangerous goods security inspection image dataset to construct another dataset for self-supervised learning;

[0062] Photographed using terahertz imaging equipment target images, each target image is denoted as . Use the point The contours of dangerous goods in the target image are marked, and the coordinates of the contour points are transformed into label data after the points are connected into irregular contour lines. and The input image set of images and Used for instance segmentation training and verification. Automatic data augmentation methods such as Figure 1 As shown. The method consists of two enhancement spaces, namely the pixel enhancement space containing six enhancement operations and a spatial enhancement space containing eight enhancement operations A highly randomized operation sampling strategy is used to sample operations from the two enhancement spaces, assuming that each enhancement branch consists of sequentially executed data augmentation operations, generating Seed from and The combination of sampling in the corresponding enhanced branch number .Will After automatic data augmentation, the training images can be obtained A dataset of images Used for self-supervised learning training.

[0063] Step 2: Use the enhanced dataset to perform self-supervised learning of image mask modeling, which is used as the backbone of the pre-trained instance segmentation model;

[0064] like Figure 2 As shown, self-supervised learning image mask modeling enhances the model's ability to extract features based on This method is implemented by an image mask modeling method. This method consists of an encoder and a decoder. The image in the enhanced dataset is used as the input image, and the input image is randomly masked. The unmasked pixels are used as The sparse voxels of the point cloud are encoded using sparse convolution. The hierarchical decoder of the style reconstructs the image from the multi-scale encoded features. After the self-supervised learning training is completed, only the parameters of the encoder are retained. The feature extraction backbone in the instance segmentation model is used as the encoder of the image mask reconstruction model, and the parameters retained after training are transferred to the instance segmentation model for fine-tuning training. This can enhance the feature extraction ability of the instance segmentation model in the case of few samples, which is conducive to reducing data annotation and model training costs.

[0065] In the feature fusion process of encoder and decoder, a multi-level feature refinement fusion mechanism is designed ( ), and integrate it into In this method, the model's ability to capture multi-level semantic information is improved, thereby improving the quality of learned representations in instance segmentation tasks. Figure 2 As shown in the figure, the designed multi-level feature refinement fusion mechanism is implemented in the following way: for an image in an enhanced dataset as the input image , encoder output The feature map of each level is used Indicates level of features, where , Module collection features The features of level l are directly passed to the decoder. In, use For reference, by max pooling ( ), the upsampling operation adjusts each level, implementation and The feature sizes are aligned, and 1×1 convolution is used to achieve channel number alignment, which is specifically expressed as follows:

[0066]

[0067] in, , , represent Convolution changes the number of channels to , represents bilinear interpolation, represents the identity mapping, Represents the maximum pooling feature map arrive Resolution, For the Layer The feature map output after the level is aligned is For the The reference height of the layer feature map, For the The reference width of the layer feature map, is the number of encoder output layers.

[0068] In all the After the feature maps of each level are aligned, Figure 2 The feature refinement module shown ( ) Filter information and fuse feature maps to obtain , in order to reduce conflicting information and improve the representation of tiny features. Taking the second level as an example, Feature size and input feature map The size of is aligned, then the expression is as follows:

[0069]

[0070] in, Represents the implicit function of the feature refinement module.

[0071] The complete self-supervised learning training process can be expressed as follows: the backbone of the instance segmentation model is used as the encoder for self-supervised learning training, and the feature map is fused using the designed multi-level feature refinement fusion mechanism. , and After filling all the blank positions of the sparse feature map, a block consisting of three consecutive blocks is used. The decoder composed of the upsampling layer decodes the fused feature map and reconstructs the image, and then completes the pre-training. After the pre-training is completed, only the parameters of the encoder are retained and migrated to the backbone of the detector for fine-tuning training. The complete pre-training process is expressed as follows:

[0072] ,

[0073] in, and Respectively represent Fusion mechanism implicit function and fused feature map, Indicates output layer, and Represents mask embedding The padding operation and the feature map after padding, Implicit functions representing successive blocks of the decoder.

[0074] Step 3: Transfer the parameters of the self-supervised pre-trained backbone, fine-tune the training instance segmentation algorithm on the unenhanced dataset, and obtain the segmentation image of the dangerous goods hidden in the human body;

[0075] like Figure 3 As shown in the figure, the instance segmentation algorithm obtains the instance segmentation image of the dangerous goods hidden in the human body by fine-tuning the instance segmentation algorithm model network composed of a reparameterizable feature extraction backbone pre-trained by self-supervised learning, a neck that can integrate multi-scale features, and a lightweight dynamically decoupled head. The unenhanced dataset is input into the instance segmentation model as the training dataset. Specifically, the image of the target object in the training dataset is Input into the instance segmentation network model, and then the output is the segmentation image of the dangerous goods hidden in the human body , its mathematical model can be expressed as follows:

[0076]

[0077] in, represents a dangerous goods detection image, represents the neural network model of the instance segmentation algorithm, are the parameters of the neural network, Represents the pixel horizontal and vertical coordinates of the output detection box, represents the dangerous goods segmentation mask, , , Respectively represent the number of channels, height and width of the image. If there are dangerous goods in the detected target image, an image with a detection frame and a segmentation mask will be output. In addition, the category code of the dangerous goods and its prediction confidence will be further displayed on the detection frame; on the contrary, if there are no dangerous goods in the detected target image, the output image is consistent with the input image.

[0078] In the feature extraction backbone, the backbone parameters after self-supervised pre-training are transferred through learning, and the instance segmentation algorithm is trained using structural reparameterization and fine-tuning training, which is beneficial to enhance the model's feature extraction ability and improve generalization, thereby improving the detection accuracy of hidden dangerous goods.

[0079] A structural reparameterization module is designed in the feature extraction process , which enables the network to decouple the complex multi-branch structure in the training phase and the flat structure in the reasoning phase, thereby improving the reasoning speed and detection accuracy of the instance segmentation model. Figure 4 As shown in the figure, the designed structure reparameterization module is implemented in the following way: given a Feature map , first convolve and separate the feature map to obtain two Feature map of size and , and then one of the feature maps enter Such as Figure 4 Shown module. The module consists of four branches during the model training phase, and is equivalently converted to a single-branch structure during the inference phase to improve the inference speed. indivual Feature map of module output and and After fusion and convolution operation, the feature map is output . Its mathematical model can be expressed as follows:

[0080]

[0081]

[0082]

[0083] in, for The feature map output by the module, and represents convolution and Implicit functions of modules, and Represent separation and fusion operations respectively.

[0084] In such Figure 3 The feature fusion neck shown uses the structure reparameterization module Designing a neck that can integrate multi-scale features is beneficial to reducing the overall parameter amount of the instance segmentation model, as well as enhancing the model's feature extraction capability and improving the detection accuracy of hidden dangerous goods.

[0085] like Figure 3 As shown in Figure 2, in order to reduce the number of parameters in the instance segmentation model and improve detection accuracy, the following are designed in the detection head: Figure 5 The dynamic decoupling head shown. Given three input feature maps whose sizes are halved successively , , the feature map , , Use The convolution reduces the number of channels to obtain , , , and enter the Such as Figure 5 shown The dynamic attention module is composed of Figure 5 The decoupled head shown in the figure performs dangerous goods feature detection and segmentation. Its mathematical model can be expressed as follows:

[0086] , ,

[0087]

[0088]

[0089] in, is the output feature map after processing by the dynamic attention module, Indicated by A continuous The implicit function of the dynamic attention module composed of represents the implicit function of the decoupled detection head, represents the convolution function, Indicates Output feature map of the layer decoupling head.

[0090] In the fine-tuning training process of the instance segmentation algorithm, Function to loss function The optimization process is as follows:

[0091]

[0092]

[0093] in, is the number of detection layers, is the bounding box regression loss function, is the classification loss function, , is the weight coefficient of the loss function, is the bounding box loss function, is the prediction box loss function.

[0094] go through After training, the optimized parameters can be obtained .

[0095] For an image taken by a terahertz security scanning camera After the feature information is extracted by the reparameterized feature extraction backbone, the multi-scale feature neck can be integrated, and the lightweight dynamic decoupling head is processed, the segmentation result of dangerous goods detection can be obtained. The result image containing the dangerous goods detection frame, dangerous goods category code, dangerous goods segmentation mask and prediction confidence is obtained, where: are the parameters of the neural network after training and optimization.

[0096] The above is only a preferred specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A self-supervised reparameterized terahertz image dangerous goods instance segmentation method, characterized in that: The specific steps include: Acquire a terahertz human body hidden dangerous goods security inspection image dataset, and perform data enhancement on the terahertz human body hidden dangerous goods security inspection image dataset; Pre-training a self-supervised learning image mask modeling enhancement model through the enhanced terahertz human body concealed dangerous goods security inspection image data set, and obtaining encoder parameters of the self-supervised learning image mask modeling enhancement model; The self-supervised learning image mask modeling enhancement model is used to The image mask modeling method can extract features; Among them, the The image mask modeling method is composed of an encoder and a decoder, the enhanced terahertz human body concealed dangerous goods security inspection image data set is randomly blocked, and the unblocked pixels are used as sparse voxels and encoded by sparse convolution. The encoder outputs several levels of feature maps and obtains target level features. The multi-level feature fusion module collects several target level features for optimization and transmits them to the decoder for image reconstruction; The multi-level feature fusion module collects a number of features of the target level for optimization and transmits them to the decoder for image reconstruction, and also includes: In each layer, based on the features of the target level, each is adjusted through maximum pooling and upsampling operations. level, to achieve feature size alignment, and use convolution to achieve channel number alignment, that is: in, , , represent Convolution changes the number of channels to , represents bilinear interpolation, represents the identity mapping, Represents the maximum pooling feature map arrive Resolution, For the Layer The feature map output after the level is aligned is For the The reference height of the layer feature map, For the The reference width of the layer feature map, is the number of encoder output layers; Pre-training the self-supervised learning image mask modeling enhancement model through the enhanced terahertz human body concealed dangerous goods security inspection image data set, and obtaining the encoder parameters of the self-supervised learning image mask modeling enhancement model includes: The self-supervised learning image mask modeling enhancement model includes a feature refinement fusion module; The self-supervised learning image mask modeling enhancement model is used as an encoder for self-supervised learning, and the images in the enhanced terahertz human body concealed dangerous goods security inspection image dataset are fused through the feature refinement fusion module to obtain a fused feature map; Fill all blank positions of the sparse feature map, use the decoder to decode and reconstruct the fused feature map, complete the pre-training of the self-supervised learning image mask modeling enhancement model, and then obtain the encoder parameters of the self-supervised learning image mask modeling enhancement model; Using the encoder of the self-supervised learning image mask modeling enhancement model as the initial feature extraction backbone network of the instance segmentation model; Migrating the encoder parameters, fine-tuning the initial feature extraction backbone network of the instance segmentation model through the unenhanced terahertz human body concealed dangerous goods security inspection image dataset, and obtaining the feature extraction backbone network of the instance segmentation model; The human body concealed dangerous goods security inspection image to be segmented is input into the feature extraction backbone network of the instance segmentation model, multi-scale features are extracted, the multi-scale features are integrated, and dynamically decoupled to obtain dangerous goods detection segmentation results.

2. The self-supervised reparameterized terahertz image dangerous goods instance segmentation method according to claim 1, characterized in that: Acquiring the terahertz human body concealed dangerous goods security inspection image dataset and performing data enhancement on the terahertz human body concealed dangerous goods security inspection image dataset comprises: Using terahertz imaging equipment, obtain several target images; Marking the contours of the dangerous goods in several target images, and converting the coordinates of the contour points to generate label data, to obtain the terahertz human body concealed dangerous goods security inspection image data set; Based on The automatic data enhancement method is used to automatically enhance the data of the terahertz human body concealed dangerous goods security inspection image dataset; Among them, the The automatic data augmentation method consists of pixel augmentation space and spatial augmentation space.

3. The self-supervised reparameterized terahertz image dangerous goods instance segmentation method according to claim 1, characterized in that: The pre-training process is: , in, and Respectively represent Fusion mechanism implicit function and fused feature map, Indicates output layer, and Represents mask embedding The padding operation and the feature map after padding, represents the implicit function of successive blocks of the decoder, , , They correspond to feature maps of different scales in the backbone.

4. The self-supervised reparameterized terahertz image dangerous goods instance segmentation method according to claim 1, characterized in that: Migrating the encoder parameters, fine-tuning the initial feature extraction backbone network of the instance segmentation model through the unenhanced terahertz human body concealed dangerous goods security inspection image data set, and obtaining the feature extraction backbone network of the instance segmentation model includes: By migrating the encoder parameters, the unenhanced terahertz human body concealed dangerous goods security inspection image dataset is input into the instance segmentation model, and the initial feature extraction backbone network of the instance segmentation model is fine-tuned and trained using a structure reparameterization module to obtain the feature extraction backbone network of the instance segmentation model; The structure reparameterization module is used to convolve and separate the given initial feature map, obtain the first feature map and the second feature map, input the first feature map into the DBB module, fuse the feature map output by the DBB module with the first feature map and the second feature map, and output the target feature map through convolution operation; The structure reparameterization module is: in, is the feature map, for The feature map output by the module, and represents convolution and Implicit functions of modules, and represent separation and fusion operations respectively, is the output feature map.

5. The self-supervised reparameterized terahertz image dangerous goods instance segmentation method according to claim 4, characterized in that: Fine-tuning the initial feature extraction backbone network of the instance segmentation model further includes: using The function optimizes the loss function, namely: in, is the loss function, is the number of detection layers, is the bounding box regression loss function, is the classification loss function, , is the weight coefficient of the loss function, is the bounding box loss function, is the prediction box loss function.

6. The self-supervised reparameterized terahertz image dangerous goods instance segmentation method according to claim 1, characterized in that: Inputting the human body concealed dangerous goods security inspection image to be segmented into the feature extraction backbone network of the instance segmentation model, extracting multi-scale features, integrating the multi-scale features, and dynamically decoupling them to obtain dangerous goods detection segmentation results includes: Inputting the security inspection image of dangerous goods concealed by a human body to be segmented into the instance segmentation model, and extracting multi-scale features through the feature extraction backbone network of the instance segmentation model; Integrate the multi-scale features and dynamically decouple them through a dynamic decoupling head module to obtain the dangerous goods detection segmentation result; The dangerous goods detection and segmentation results include a dangerous goods detection frame, a dangerous goods category code, a dangerous goods segmentation mask and a result image of the prediction confidence.

7. The self-supervised reparameterized terahertz image dangerous goods instance segmentation method according to claim 6, characterized in that: The dynamic decoupling head module is used to give a number of feature maps whose sizes are halved in sequence, use convolution to reduce the number of channels to obtain a number of convolved feature maps, and input them into the dynamic attention module, and use the decoupling head to perform dangerous goods feature detection and segmentation; The dynamic decoupling head module is: in, , , , , , are all feature maps, and , , , , , , Indicated by A continuous The implicit function of the dynamic attention module composed of represents the implicit function of the decoupled detection head, represents the convolution function, Indicates Output feature map of the layer decoupling head, It is the output feature map after being processed by the dynamic attention module.

Images