CT Image Processing Method and Apparatus, and Inspection Method and Apparatus for International Express Parcels

Abstract

The present invention relates to a CT image processing method and apparatus, and a method and apparatus for inspecting international express parcels. The CT image processing method includes: a preprocessing step of uniformly sampling and coordinate normalizing three-dimensional data of a CT image to obtain preprocessed data; an object instance obtaining step of obtaining object instance data after instance segmentation combined with semantic information for the preprocessed data; and a structured feature obtaining step of obtaining structured features of the object instance data by using a feature extractor obtained by training based on a neural network.

Description

Technical Field

[0001] The present invention relates to the field of CT image processing, and particularly to the intelligent structured description technology of CT images for the inspection of international express parcels. Background Art

[0002] With the wide popularization of e-commerce, emerging business forms such as cross-border e-commerce have developed rapidly, and import and export trade has become increasingly prosperous. The import and export volume of our country is continuously increasing, and the workload of express parcel supervision and inspection is also gradually increasing. The traditional on-site supervision mode of express parcels: For any parcel entering the supervision center, it is necessary to use the traditional inspection system to check the actual information of the parcel, usually by comparing the X-ray perspective view with the parcel declaration information to complete. When illegal merchants make entry declarations, they make false statements and conceal the true situation of the goods, so as to achieve the purpose of paying less or even not paying taxes. In addition, some lawbreakers misreport goods clearly stipulated by the customs as items that cannot be mailed or carried and hide them in parcels in order to enter the country smoothly. Most of the existing inspection systems rely on the perspective images scanned by X-ray machines, and the position of the X-ray machine in the inspection equipment is fixed and cannot penetrate the parcel from more angles, which affects the imaging effect and leads to more missed inspections and misinspections.

[0003] In addition, in the past, only physical characteristic values such as density and atomic number of CT scan images were used for item inspection, and the three-dimensional shape information scanned by CT was not fully utilized, so the image characteristics of the declared goods in the parcel could not be completely expressed, and thus the comparison between the goods list and the order could not be accurately realized. Summary of the Invention

[0004] In order to overcome the defect in the prior art that the three-dimensional shape information of goods cannot be accurately utilized, the present invention provides a CT image processing method and device, as well as an inspection method and device for international express parcels.

[0005] According to one aspect of the present invention, a CT image processing method is provided, including: a preprocessing step of uniformly sampling and coordinate normalizing the three-dimensional data of the CT image to obtain preprocessing data; an item instance obtaining step of obtaining item instance data after instance segmentation combined with semantic information for the preprocessing data; and a structured feature obtaining step of obtaining the structured features of the item instance data by using a feature extractor obtained by training based on a neural network.

[0006] According to the above CT image processing method of the present invention, the article instance obtaining step includes: obtaining a feature matrix including global features and local features from the preprocessed data; respectively passing the feature matrix through two decoders to implement semantic segmentation and instance segmentation; embedding the semantic features obtained by semantic segmentation into the instance space after instance segmentation to obtain an instance feature matrix with known semantics; and obtaining the article instances after instance segmentation by performing unsupervised clustering on different instances to label instance labels.

[0007] According to the above CT image processing method of the present invention, the construction of the feature extractor includes: constructing a three-dimensional metric learning network based on the siamese network framework that can learn general image features; obtaining source data including paired data as training data; and training the three-dimensional metric learning network with the training data to obtain the feature extractor.

[0008] According to one aspect of the present invention, there is provided an international express inspection method, including: an image acquisition step of acquiring a CT three-dimensional image of an express; a CT image processing step of using the above CT image processing method to obtain the structured features of the CT image; a feature map construction step of constructing a feature map for each article, the feature map including fine-grained field information including attribute information reflecting the content of the express declaration form and the structured features of the CT image; a preset model acquisition step of respectively obtaining a preset feature model for the articles of the feature map; and a judgment step of judging whether the structured features of each article match the preset feature model.

[0009] According to the above international express inspection method of the present invention, wherein the feature map construction step includes: a text parsing step of using a BERT neural network model to obtain the fine-grained field information of the declaration form of the express; a text-image correspondence step of corresponding the fine-grained field information with the structured features of the image extracted by the feature extractor and the physical information of the CT scan image; and a storage step of storing the content established in the text-image correspondence step as a feature map.

[0010] According to the above international express inspection method of the present invention, wherein the training of the BERT neural network model includes the following steps: a declaration form data acquisition step of acquiring a plurality of declaration form data as source data; a feature extraction step of extracting the declaration elements of all declaration form data; an attribute annotation step of performing word segmentation on the application elements and performing attribute annotation on each word segment according to an attribute label; and a training step of inputting the word segments with attribute annotation into the BERT network for article attribute training to obtain the BERT neural network model.

[0011] According to the above international express inspection method of the present invention, the judgment step includes: a matching coefficient calculation step of calculating a matching coefficient between the structural features of the CT image and a preset feature model; and a matching judgment step of judging that the article is consistent with the customs declaration information when the matching coefficient is greater than a preset threshold.

[0012] According to one aspect of the present invention, there is provided a CT image processing apparatus, including: a preprocessing unit that performs uniform sampling and coordinate normalization processing on three-dimensional data of a CT image to obtain preprocessed data; an article instance acquisition unit that obtains article instance data after instance segmentation combined with semantic information for the preprocessed data; and a structural feature acquisition unit that obtains structural features of the article instance data by using a feature extractor obtained by training based on a neural network.

[0013] According to the above CT image processing apparatus of the present invention, wherein the article instance acquisition unit includes the following units: obtaining a feature matrix including global features and local features from the preprocessed data; respectively passing the feature matrix through two decoders to implement semantic segmentation and instance segmentation; embedding semantic features obtained by semantic segmentation into an instance space after instance segmentation to obtain an instance feature matrix with known semantics; and obtaining the article instance after instance segmentation by performing unsupervised clustering on different instances to label instance labels.

[0014] According to one aspect of the present invention, there is provided an international express inspection apparatus, which includes: an image acquisition unit that acquires a CT three-dimensional image of an express; a CT image processing unit that uses the above CT image processing apparatus to obtain structural features of the CT image; a feature map construction unit that constructs a feature map for each article, where the feature map includes fine-grained field information including attribute information reflecting the content of the customs declaration of the express and the structural features of the CT image; a preset model acquisition unit that respectively obtains a preset feature model for the articles in the feature map; and a judgment unit that judges whether the structural features of each article match the preset feature model.

[0015] According to the above international express inspection apparatus of the present invention, wherein the feature map construction unit includes: a text parsing unit that uses a BERT neural network model to obtain the fine-grained field information of the customs declaration of the express; a text-image correspondence unit that corresponds the fine-grained field information with the structural features of the image extracted by the feature extractor and the physical information of the CT scan image; and a storage unit that stores the content established in the text-image correspondence unit as a feature map.

[0016] According to the above-mentioned international express inspection device of the present invention, the judgment unit includes: a matching coefficient calculation unit that calculates the matching coefficient between the structural features of the CT image and a preset feature model; and a matching judgment unit that determines that the item is consistent with the customs declaration information when the matching coefficient is greater than a preset threshold.

[0017] According to the above-mentioned international express inspection method and device, an intelligent comparison method based on three-dimensional CT images and customs declaration data is provided. Through the international express inspection method involved in the present invention, compared with the previous manual inspection and judgment, the inspection efficiency of machine inspection is greatly improved. In addition, the present invention effectively utilizes three-dimensional information to accurately judge the machine inspection CT image, improving the accuracy of international express inspection. In addition, according to the above-mentioned international express inspection method, multi-dimensional information such as item density, atomic number, and graphic images can also be combined to accurately describe the machine inspection CT image, improving the accuracy of verifying the customs declaration documents of the package and the machine inspection CT image. Brief Description of the Drawings

[0018] Figure 1 is a flowchart showing the CT image processing method involved in the present invention;

[0019] Figure 2 is a specific flowchart showing the acquisition of item instances;

[0020] Figure 3 is a flowchart showing the construction of a feature extractor;

[0021] Figure 4 is a flowchart showing the international express inspection method of the present invention;

[0022] Figure 5 is a flowchart showing the construction of a feature map;

[0023] Figure 6 is a flowchart showing the training of a BERT neural network model;

[0024] Figure 7 is a flowchart showing the judgment process;

[0025] Figure 8 is a block diagram showing the CT image processing device;

[0026] Figure 9 is a block diagram showing the international express inspection device. Detailed Description of the Invention

[0027] The features and exemplary embodiments of various aspects of the present invention will be described in detail below. In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only configured to explain the present invention and are not configured to limit the present invention. For those skilled in the art, the present invention can be implemented without some of these specific details. The following description of the embodiments is only to provide a better understanding of the present invention by showing examples of the present invention.

[0028] The present invention provides a method for inspecting international express parcels. The method for inspecting international express parcels can utilize the three-dimensional image features of CT images, thereby enabling more accurate inspection.

[0029] In the first embodiment of the present invention, Figure 1 is a flowchart showing the CT image processing method involved in the present invention, including: a preprocessing step S1 of uniformly sampling and coordinate normalizing the three-dimensional data of the CT image to obtain preprocessing data; an object instance obtaining step S2 of obtaining object instance data after instance segmentation combined with semantic information for the preprocessing data; and a structured feature obtaining step S3 of obtaining the structured features of the object instance data by using a feature extractor obtained by training based on a neural network.

[0030] Pass the international express parcel through a CT machine to obtain a three-dimensional CT image. In CT image processing, first, in the preprocessing step S1, uniformly sample and perform coordinate normalization on the three-dimensional data of the CT image. In the preprocessing step, the resolution of the obtained three-dimensional CT image data is relatively high. To reduce the network model overhead as needed, perform equally spaced uniform sampling to obtain sampling points. Perform coordinate normalization on the sampling points to normalize the coordinates of the sampling points to between (0 and 1), thereby obtaining preprocessing data.

[0031] Figure 2 is a specific flowchart showing the obtaining of object instances. In the object instance obtaining step S2, obtain object instance data after instance segmentation combined with semantic information for the preprocessing data. Specifically, it includes the following steps S21 to S24.

[0032] In step S21, obtain a feature matrix including global features and local features from the preprocessing data. Input the preprocessing data into a pre-trained point cloud feature extraction network to obtain a feature matrix combining global features and local features.

[0033] In step S22, the above-mentioned feature matrix is respectively passed through two decoders to achieve semantic segmentation and instance segmentation. There are two decoders set, and the same feature matrix is synchronously input into the two decoders respectively to achieve semantic segmentation and instance segmentation. Semantic segmentation is performed through one decoder. For example, semantic features are obtained according to the feature data, and the items are milk powder cans and cups. Instance segmentation is performed through the other decoder. For example, instance features are obtained according to the feature data, including five milk powder can instances and five water cup instances.

[0034] In step S23, the semantic features obtained by semantic segmentation are embedded into the instance space after instance segmentation to obtain an instance feature matrix with known semantics. For example, the obtained semantics include milk powder cans and water cups. The semantics of milk powder cans and water cups are respectively embedded into each instance in the instance space after instance segmentation, so as to obtain the instance feature matrices with known semantics.

[0035] In step S24, different instances are labeled with instance labels through unsupervised clustering to obtain the item instances after instance segmentation. In the above embodiment, the above-mentioned sampling points have feature vectors with a length of 5, that is, five-dimensional feature vectors. The sampling points are subjected to unsupervised clustering according to the feature vectors, and the points with the same feature vectors are clustered, and labels are added to each instance, so as to obtain the item instances after instance segmentation according to the instance labels.

[0036] Through the above, item instance data is obtained for the preprocessed data.

[0037] In the structured feature acquisition step S3, a feature extractor obtained by training based on a neural network is used to obtain the structured features of the item instance data.

[0038] Figure 3 It is a flowchart showing the construction of the feature extractor, including the following steps S31 to S33.

[0039] In step S31, a three-dimensional metric learning network capable of learning general image features based on the siamese network framework is constructed. According to the three-dimensional characteristics of the instance data, a three-dimensional metric learning network capable of learning general image features is designed based on the siamese network framework. For example, for a milk powder can, its three-dimensional characteristic is the shape of the milk powder can. In step S32, source data containing paired data is obtained as training data. Standard optimization techniques are used to train the paired data sampled from the source data. The paired data as training data, for example, paired data indicating that item A and item B are the same, paired data indicating that item C and item D are different. In step S33, the three-dimensional metric learning network is trained with the training data to obtain the feature extractor. The feature extractor is obtained by training the paired data through the three-dimensional metric learning network.

[0040] Return to step S3. Using the trained feature extractor mentioned above, the feature extractor obtains the structured features of the item instance data.

[0041] Through the above CT image processing method, for CT images, the structured features of the item instance data of the three-dimensional CT image can be obtained. That is, through the CT image processing method, the three-dimensional image features of the image can be obtained.

[0042] The second embodiment of the present invention provides a method for inspecting international express deliveries using the above CT image processing method. Figure 4 It is a flowchart showing the international express delivery inspection method of the present invention. The international express delivery inspection method includes: an image acquisition step S100 of acquiring a CT three-dimensional image of the express delivery; a CT image processing step S200 of using the CT image processing method in the above embodiment to obtain the structured features of the CT image; a feature map construction step S300 of constructing a feature map for each type of item, the feature map including fine-grained field information containing attribute information reflecting the content of the express delivery customs declaration form and the structured features of the CT image; a preset model acquisition step S400 of respectively acquiring a preset feature model for the items in the feature map; and a judgment step S500 of judging whether the structured features match the preset feature model.

[0043] In step S100, a CT three-dimensional image of the international express delivery is acquired by scanning the international express delivery with a CT machine.

[0044] In step S200, the CT image processing method of the above embodiment is used to obtain the structured features of the CT image, and the specific process is not described in detail here.

[0045] In step S300, a feature map is constructed for each type of item, the feature map including fine-grained field information containing attribute information reflecting the content of the express delivery customs declaration form and the structured features of the CT image.

[0046] Figure 5 It is a flowchart showing the construction of the feature map, including step S301 to step S303. In the text parsing step S301, the BERT neural network model is used to obtain the fine-grained field information of the customs declaration form of the express delivery.

[0047] Figure 6It is a flowchart showing the training of the BERT neural network model, including step S3001 to step S3004. In the customs declaration data acquisition step S3001, multiple customs declaration data as source data are acquired. A large amount of customs declaration data is acquired as training data. In the element extraction step S3002, the declaration elements of all customs declaration data are extracted. In the attribute annotation step S3003, the application elements are tokenized, and each token is annotated with an attribute label according to the attribute tag. For example, the declaration form is 50 barrels of Meiji milk powder, 100 grams per barrel. The tokenization result is milk powder, Meiji, 50 barrels, 100 grams per barrel. The attribute tags include, for example, commodity name, brand, quantity, specification, etc. After attribute annotation: milk powder - commodity name; Meiji - brand; 50 barrels - quantity; 100 grams per barrel - specification. In the training step S3004, the tokens with attribute annotation are input into the BERT network for item attribute training to obtain the BERT neural network model. Through the above training, the BERT neural network model can identify the attributes of the customs declaration words, and a parsing algorithm is designed according to the BERT neural network model to obtain fine-grained fields including attribute information such as the commodity name, brand, quantity, and specification corresponding to the customs declaration data.

[0048] Return to Figure 5 , in the text-image correspondence step S302, the fine-grained field information is corresponded to the structured features of the image extracted by the feature extractor and the physical information of the CT scan image. The fine-grained field information is precisely coupled with the structured features of the image extracted by the feature extractor and the physical information of the CT scan image to construct a feature map of text and image bimodal information. The physical information of the patent CT scan image refers to the item weight information calculated according to the density.

[0049] In step S302, the content established in the text-image correspondence step S301 is stored as a feature map.

[0050] According to the above Figure 5 steps, a feature map is constructed for each commodity of the customs declaration.

[0051] Return to Figure 4 , in the preset model acquisition step S400, a preset feature model is acquired for each item in the feature map. For each item in the feature map, the feature model of the item is acquired in advance and stored. Here, the feature model of the item can be stored in the feature map corresponding to the item in the feature map, or stored separately.

[0052] In the judgment step S500, it is judged whether the structured features of each item match the preset feature model. Figure 7It is a flowchart showing a judgment process, including step S501 and step S502. In the matching coefficient calculation step S501, the matching coefficient between the structural features of the CT image and a preset feature model is calculated. The matching coefficient between the feature data of the input CT image and the preset feature model is calculated according to the following formula:

[0053] r = max[r(X, Y i )], i = 1, 2, 3, …, n

[0054]

[0055] where r is the matching coefficient between the feature data of the CT image and the preset feature model, X is the feature data of the CT image, Y is the preset feature model, and r(X, Y i ) is the correlation coefficient between the feature data of the CT image and the i-th preset feature model, n is the number of preset feature models, Cov(X, Y i ) is the covariance between the feature data of the CT image and the i-th preset feature model, Var[X] is the variance of the feature data of the CT image, and Var[Y i is the variance of the i-th preset feature model.

[0056] In the matching judgment step S502, when the matching coefficient is greater than a preset threshold, it is judged that the item is consistent with the customs declaration information. When the matching coefficient r is greater than the preset threshold, it indicates that the image is consistent with the customs declaration; when the matching coefficient r is below the preset threshold, it indicates that the image is inconsistent with the customs declaration.

[0057] According to the above international express inspection method, an intelligent comparison method based on three-dimensional CT images and customs declaration data is provided. Through the international express inspection method involved in the present invention, compared with the previous manual inspection and judgment, the inspection efficiency of machine inspection is greatly improved.

[0058] In addition, the present invention effectively utilizes three-dimensional information to accurately judge the machine inspection CT image, improving the accuracy of international express inspection.

[0059] In addition, according to the above international express inspection method, multi-dimensional information such as item density, atomic number, and graphic images can also be combined to accurately describe the machine inspection CT image, improving the accuracy of verifying the customs declaration documents of the package and the machine inspection CT image.

[0060] The third embodiment of the present invention also provides a CT image processing device 1, Figure 8is a block diagram showing a CT image processing apparatus, including: a preprocessing unit 10 that performs uniform sampling and coordinate normalization processing on three-dimensional data of a CT image to obtain preprocessed data; an article instance acquisition unit 20 that acquires article instance data after instance segmentation combined with semantic information for the preprocessed data; and a structured feature acquisition unit 30 that obtains structured features of the article instance data by using a feature extractor obtained through training based on a neural network. The article instance acquisition unit 20 includes the following units: obtaining a feature matrix including global features and local features from the preprocessed data; respectively passing the feature matrix through two decoders to implement semantic segmentation and instance segmentation; embedding semantic features obtained by semantic segmentation into an instance space after instance segmentation to obtain an instance feature matrix with known semantics; and obtaining the article instance after instance segmentation by performing unsupervised clustering on different instances to label instance tags. For a specific description of the CT image processing apparatus of this embodiment, refer to the description of the above CT image processing method, which will not be elaborated here.

[0061] A fourth embodiment of the present invention further provides an international express inspection apparatus 100, Figure 9 is a block diagram showing an international express inspection apparatus, including: an image acquisition unit 101 that acquires a CT three-dimensional image of an express; a CT image processing unit 102 that uses the CT image processing apparatus of the above embodiment to obtain structured features of the CT image; a feature map construction unit 103 that constructs a feature map for each article, where the feature map includes fine-grained field information including attribute information reflecting the content of the express customs declaration form and the structured features of the CT image; a preset model acquisition unit 104 that respectively acquires a preset feature model for the articles in the feature map; and a judgment unit 105 that judges whether the structured features of each article match the preset feature model. The feature map construction unit 103 includes: a text parsing unit that uses a BERT neural network model to obtain the fine-grained field information of the customs declaration form of the express; a text-image correspondence unit that corresponds the fine-grained field information with the structured features of the image extracted by the feature extractor and the physical information of the CT scan image; and a storage unit that stores the content established in the text-image correspondence unit as a feature map. The judgment unit 105 includes: a matching coefficient calculation unit that calculates a matching coefficient between the structured features of the CT image and the preset feature model; and a matching judgment unit that judges that the article is consistent with the customs declaration form information when the matching coefficient is greater than a preset threshold. For a specific description of the international express inspection apparatus of this embodiment, refer to the description of the above international express inspection method, which will not be elaborated here.

[0062] Based on the above-mentioned international express inspection device, an intelligent comparison method based on 3D CT images and customs declaration form data is provided. Through the international express inspection method involved in the present invention, compared with the previous manual inspection and judgment, the inspection efficiency of machine inspection is greatly improved.

[0063] In addition, the present invention effectively utilizes three-dimensional information to accurately judge the machine inspection CT images, improving the accuracy of international express inspection.

[0064] In addition, based on the above-mentioned international express inspection device, multi-dimensional information such as the density of the item, atomic number, and graphic image can also be combined to accurately describe the machine inspection CT images, improving the accuracy of the verification of the customs declaration documents of the package and the machine inspection CT images.

[0065] As mentioned above, although the embodiments and specific examples of the present invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and deformations without departing from the spirit and scope of the present invention, and such modifications and deformations all fall within the scope defined by the claims.

Claims

1. An international express inspection method, wherein, it includes: An image acquisition step of acquiring a CT three-dimensional image of the express; A CT image processing step of using a CT image processing method to obtain the structured features of the CT image. The CT image processing method includes: a preprocessing step of uniformly sampling and coordinate normalizing the three-dimensional data of the CT image to obtain preprocessing data; an object instance acquisition step of obtaining object instance data after instance segmentation combined with semantic information for the preprocessing data; and a structured feature acquisition step of obtaining the structured features of the object instance data by using a feature extractor obtained through training based on a neural network; A feature map construction step of constructing a feature map for each type of object. The feature map includes fine-grained field information containing attribute information reflecting the content of the express declaration form and the structured features of the CT image; A preset model acquisition step of respectively obtaining preset feature models for the objects in the feature map; and A judgment step of judging whether the structured features of each object match the preset feature model.

2. The international express inspection method according to claim 1, wherein, the feature map construction step includes: A text parsing step of using a BERT neural network model to obtain the fine-grained field information of the declaration form of the express; A text-image correspondence step of corresponding the fine-grained field information with the structured features of the image extracted by the feature extractor and the physical information of the CT scan image; A storage step of storing the corresponding content established in the text-image correspondence step as a feature map.

3. The international express inspection method according to claim 2, wherein, the training of the BERT neural network model includes the following steps: A declaration form data acquisition step of acquiring multiple declaration form data as source data; An element extraction step of extracting the declaration elements of all declaration form data; An attribute annotation step of performing word segmentation on the application elements and performing attribute annotation on each word segment according to attribute labels; A training step of inputting the word segments with attribute annotation into the BERT network for object attribute training to obtain the BERT neural network model.

4. The international express inspection method according to claim 1, wherein, the judgment step includes: A matching coefficient calculation step of calculating the matching coefficient between the structured features of the CT image and the preset feature model; A matching judgment step of judging that the object is consistent with the declaration form information when the matching coefficient is greater than a preset threshold.

5. The international express inspection method according to claim 1, wherein, the object instance acquisition step includes: Obtaining a feature matrix containing global features and local features from the preprocessing data; Respectively passing the feature matrix through two decoders to implement semantic segmentation and instance segmentation; Embedding the semantic features obtained by semantic segmentation into the instance space after instance segmentation to obtain an instance feature matrix with known semantics; and Obtaining the object instance after instance segmentation by performing unsupervised clustering on different instances to label instance labels.

6. The international express inspection method according to claim 1, wherein, the construction of the feature extractor includes: Construct a 3D metric learning network based on the Siamese network framework that can learn general image features; Obtain source data containing paired data as training data; Train the 3D metric learning network using the training data to obtain the feature extractor.

7. An international express inspection device, wherein, comprising: an image acquisition unit that acquires a CT 3D image of the express; a CT image processing unit that uses a CT image processing device to obtain the structured features of the CT image. The CT image processing device includes: a preprocessing unit that performs uniform sampling and coordinate normalization on the 3D data of the CT image to obtain preprocessed data; an object instance acquisition unit that obtains object instance data after instance segmentation combined with semantic information for the preprocessed data; and a structured feature acquisition unit that obtains the structured features of the object instance data using a feature extractor obtained through neural network-based training; a feature map construction unit that constructs a feature map for each item, where the feature map includes fine-grained field information containing attribute information reflecting the content of the express declaration form and the structured features of the CT image; a preset model acquisition unit that respectively obtains preset feature models for the items in the feature map; and a judgment unit that judges whether the structured features of each item match the preset feature models.

8. The international express inspection device according to claim 7, wherein, the feature map construction unit includes: a text parsing unit that uses a BERT neural network model to obtain the fine-grained field information of the express declaration form of the express; a text-image correspondence unit that corresponds the fine-grained field information with the structured features of the image extracted by the feature extractor and the physical information of the CT scan image; a storage unit that stores the corresponding content established in the text-image correspondence unit as a feature map.

9. The international express inspection device according to claim 7, wherein, the judgment unit includes: a matching coefficient calculation unit that calculates the matching coefficient between the structured features of the CT image and the preset feature model; a matching judgment unit that, when the matching coefficient is greater than a preset threshold, judges that the item is consistent with the express declaration information.

10. The international express inspection device according to claim 7, wherein, the object instance acquisition unit includes the following units: obtain a feature matrix containing global features and local features from the preprocessed data; pass the feature matrix through two decoders respectively to achieve semantic segmentation and instance segmentation; embed the semantic features obtained by semantic segmentation into the instance space after instance segmentation to obtain an instance feature matrix with known semantics; and obtain the object instances after instance segmentation by performing unsupervised clustering to label different instances with instance labels.

Images