A marking method and system applied to storage of involved property

Abstract

The present disclosure relates to the technical field of case-related property management, and particularly relates to a marking method and system applied to case-related property storage, which comprises: target detection and image recognition on a real-time shooting picture of case-related property through a first multi-modal large model to obtain an overall feature description of a target object in the real-time shooting picture, comparing the overall feature description with information of an object to be stored, and triggering a camera to shoot in the case of matching the overall feature description with the information of the object to be stored; obtaining an encrypted watermark based on a shooting time, user information and the overall feature description, embedding the encrypted watermark into a shooting picture, and storing the watermark picture and the object information into a data entry corresponding to the case-related object. Thus, the uniformity among case-related object information, actually shot objects and picture watermarks can be improved, the anti-fake property of the watermark is combined with the uniqueness of the state of the case-related object, the anti-fake strength of the watermark is improved, and the storage safety of the case-related object is further improved.

Description

Technical Field

[0001] This disclosure relates to the field of property management technology, specifically to a marking method and system for the entry of property into custody. Background Technology

[0002] The management of seized assets is a crucial aspect of judicial activities, with its core objective being to ensure the security, standardization, and traceability of the entire process from seizure, warehousing, storage, release, to disposal of seized assets. With the development of information technology, traditional paper-based ledger management has been gradually replaced by digital management systems. Currently, to improve the reliability of management and the evidentiary value of seized assets, digital watermarking technology is widely used when they are put into storage.

[0003] However, there is a disconnect between the generation of watermark information in the relevant technology and the physical state of the seized property. Operators may mistakenly or maliciously bind the watermark information of seized property A in the system to the actual photographed seized property B. Although the generated watermark itself is encrypted reliably and cannot be tampered with, the underlying object it is bound to is wrong from the source, rendering subsequent tracing meaningless and potentially misleading the investigation. Summary of the Invention

[0004] To overcome the problems existing in the related technologies, this disclosure provides a marking method and system for the entry of seized property into a warehouse, thereby solving the defects in the related technologies.

[0005] According to a first aspect of the present disclosure, a marking method for marking seized property upon its entry into a vault is provided, comprising:

[0006] Obtain information on seized items awaiting warehousing from the seized property management platform;

[0007] In response to the first user's operation to turn on the camera in the first terminal device, the system performs target detection and image recognition on the real-time captured image by the camera using a first multimodal large model to obtain a first analysis result. Based on the first analysis result, it generates an overall feature description of the target object in the real-time captured image. The system performs a semantic comparison between the overall feature description and the object information to obtain a semantic similarity. If the semantic similarity is higher than or equal to a preset similarity threshold, the system triggers the camera to take a picture of the target object to obtain a captured image.

[0008] Determine the shooting time of the captured image and the user information of the first user, combine at least the shooting time, the user information and the overall feature description into plaintext, encrypt the plaintext to obtain an encrypted watermark, and embed the encrypted watermark into the captured image to obtain a first watermark image;

[0009] The first watermark image and the item information are associated and stored in the data entry corresponding to the item in the database of involved property.

[0010] According to a second aspect of the present disclosure, a marking system for the storage of seized property is provided, including a seized property management platform, a first terminal device, a watermark service engine, and a seized property database;

[0011] The platform for managing seized assets is used to input information about seized items that are to be put into storage.

[0012] The first terminal device is used to respond to the first user's operation of turning on the camera, perform target detection and image recognition on the real-time shooting scene captured by the camera through the first multimodal large model, obtain a first analysis result, and generate an overall feature description of the target object in the real-time shooting scene based on the first analysis result. The overall feature description is semantically compared with the object information to obtain a semantic similarity. If the semantic similarity is higher than or equal to a preset similarity threshold, the camera is triggered to take a picture of the target object to obtain a captured image.

[0013] The watermarking service engine is used to determine the shooting time of the captured image and the user information of the first user. It combines the shooting time, the user information and the overall feature description into plaintext, encrypts the plaintext to obtain an encrypted watermark, and embeds the encrypted watermark into the captured image to obtain a first watermarked image.

[0014] The database of seized assets is used to associate and store the first watermark image and the item information in the data entries corresponding to the seized items.

[0015] The technical solutions provided in this disclosure may have the following beneficial effects:

[0016] The marking method for warehousing seized assets provided in this disclosure introduces a multimodal large model to identify and semantically compare real-time images captured by a camera. This ensures that the real-time photographed items match the seized assets to be warehoused, thereby improving the consistency between seized asset information, the actual photographed items, and image watermarks in the seized asset management platform from the source, and thus improving the accuracy of seized asset warehousing. Furthermore, when the real-time photographed items match the seized assets to be warehoused, the camera is automatically triggered to take a picture, reducing manual intervention and thus improving the efficiency of seized asset warehousing.

[0017] In addition, the overall feature description extracted by a multimodal large model is added to the watermark of the items involved in the case when they are put into storage. This combines the anti-counterfeiting properties of the watermark with the uniqueness of the status of the items involved in the case. This means that the watermark information not only contains the static business data of the items involved in the case, but also the dynamic status characteristics of the items involved in the case. This allows for more accurate identification of attackers tampering with the images of the items involved in the case, improves the anti-counterfeiting strength of the watermark images of the items involved in the case, and thus improves the security of the items involved in the case when they are put into storage. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0019] Figure 1 This is a flowchart illustrating an exemplary embodiment of the present disclosure of a marking method applied to the storage of seized property;

[0020] Figure 2 This is a block diagram illustrating an exemplary embodiment of the present disclosure of a marking system applied to the storage of involved property. Detailed Implementation

[0021] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure.

[0022] The terminology used in this disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of the disclosure. The singular forms “a,” “the,” and “the” used in this disclosure are also intended to include the plural forms unless the context clearly indicates otherwise. It should be understood that although the terms first, second, third, etc., may be used in this disclosure to describe various information, such information should not be limited to these terms. These terms are used only to distinguish information of the same type from one another. For example, first information may also be referred to as second information without departing from the scope of this disclosure, and similarly, second information may also be referred to as first information.

[0023] Currently, the watermark information on seized assets lacks real-time physical verification. It is impossible to verify whether the actual object captured by the camera truly corresponds to the selected item to be stored. Operators may, due to error or malice, bind the watermark information of seized asset A to the actual captured asset B. Although the generated watermark itself is securely encrypted and tamper-proof, the underlying object it is bound to is incorrect from the source, rendering subsequent tracing meaningless and potentially misleading the investigation.

[0024] Based on this, in a first aspect, at least one embodiment of this disclosure provides a marking method applied to the entry of seized property into a vault. Please refer to the appendix. Figure 1 The diagram illustrates the process of the method, including steps S101 to S104.

[0025] In step S101, the information of the items to be put into storage is obtained from the case-related property management platform.

[0026] In step S102, in response to the first user's operation to turn on the camera in the first terminal device, the first multimodal large model performs target detection and image recognition on the real-time shooting image captured by the camera to obtain a first analysis result. Based on the first analysis result, an overall feature description of the target object in the real-time shooting image is generated. The overall feature description is semantically compared with the object information to obtain a semantic similarity. If the semantic similarity is higher than or equal to a preset similarity threshold, the camera is triggered to take a picture of the target object to obtain a captured image.

[0027] In step S103, the shooting time of the captured image and the user information of the first user are determined. At least the shooting time, user information and overall feature description are combined into plaintext, the plaintext is encrypted to obtain an encrypted watermark, and the encrypted watermark is embedded into the captured image to obtain the first watermark image.

[0028] In step S104, the first watermark image and the item information are associated and stored in the data entry corresponding to the item in the database of involved property.

[0029] By introducing a multimodal large-scale model to identify and semantically compare real-time images captured by the camera, it is possible to ensure that the real-time photographed items match the items to be stored in the warehouse. This improves the consistency between the information on items involved in the case, the actual photographed items, and the image watermarks in the case-related asset management platform, thereby enhancing the accuracy of the storage of items involved in the case. Furthermore, when the real-time photographed items match the items to be stored in the warehouse, the camera is automatically triggered to take pictures, reducing manual intervention and improving the efficiency of storing items involved in the case. In addition, by incorporating the overall feature description extracted by the multimodal large-scale model into the watermark of the items being stored in the warehouse, the anti-counterfeiting properties of the watermark are combined with the uniqueness of the status of the items involved in the case. This ensures that the watermark information includes not only the static business data of the items involved in the case but also the dynamic status characteristics of the items. This allows for more accurate identification of attackers tampering with the images of items involved in the case, improving the anti-counterfeiting strength of the watermark images and thus enhancing the security of storing items involved in the case.

[0030] To facilitate understanding, the marking method provided in this disclosure for the entry of seized property into the warehouse will be further illustrated with examples below.

[0031] For example, the case-related property management platform provides a visual interactive page for entering information about case-related items. The item information may include the item name, category, quantity, case number, item number, and the person who handed over the item. In practical applications, the item information may be automatically pushed to the case-related property management platform through the case-handling system, or it may be manually entered by case handlers on the platform; this embodiment does not limit this approach.

[0032] For example, the first terminal device may include a high-definition camera and integrate a program for storing seized assets. This program may have a built-in lightweight multimodal large model (i.e., the first multimodal large model) for real-time visual analysis (such as object detection and image recognition). When a first user turns on the camera in the first terminal device, that is, activates the camera function of the first terminal device, the first terminal device then uses the lightweight first multimodal large model to perform object detection and image recognition on the real-time captured image. This is used to quickly verify whether the items in the real-time captured image are consistent with the seized assets to be stored, thereby achieving the binding of the captured items with the seized assets to be stored.

[0033] For example, the first multimodal large model first performs object detection, identifying one or more objects in the real-time captured image, and determining their bounding box coordinates (used to determine the object's position in the image), preliminary classification labels, and confidence scores. Then, based on the object detection results, the first multimodal large model perceives low-level features (such as color, texture, etc.) and high-level semantic features (such as object parts, states, etc.), and performs semantic understanding and association on the perceived features to obtain the first analysis result. For example, if the perceived visual characteristics are: {black, glass material, rectangle, radial cracks on the screen area, logo pattern on the back cover}, the first analysis result could be: {Subject: smartphone; Attributes: brand-X, color-black, state-screen damage, confidence score: 0.98}.

[0034] For example, the text decoder of the first multimodal large model can receive a structured first analysis structure, and then organize it into a fluent, accurate, and grammatically correct text description according to the learned language patterns. For instance, using the above example, the overall feature description of the target object in the real-time captured image can be obtained as: {a black X-brand smartphone with a broken screen}.

[0035] Therefore, the original pixel data is transformed into text descriptions rich in semantic information through the first multimodal large model, laying a solid foundation for subsequent semantic comparison, thereby realizing the binding of photographed items with items to be put into storage and improving the accuracy of the storage of items involved in the case.

[0036] For example, considering that the item information may include case number, handover person, and other information unrelated to the item's characteristics, when semantically comparing the overall feature description with the item information, we can first extract the item feature information from the item information, and then perform a semantic comparison between the overall feature description and the item feature information. For instance, continuing with the above example, the overall feature description is: {a black XX brand smartphone with a broken screen}, and the item feature information extracted from the item information is: {X brand Y model mobile phone, black}. Then, we perform a semantic comparison between the two to obtain the semantic similarity.

[0037] For example, the preset similarity threshold can be set according to actual needs, such as 0.9, and this embodiment of the present disclosure does not limit this. If the semantic similarity is higher than or equal to the preset similarity threshold, it means that the photographed item matches the item to be put into storage, so the subsequent storage process can continue, that is, the camera is triggered to photograph the target item in the real-time shooting scene to obtain the photographed image.

[0038] In some embodiments, when the semantic similarity is lower than a preset similarity threshold, a first intelligent agent can analyze the reason why the semantic similarity is lower than the preset similarity threshold based on the first analysis result, the first user's current entry operation information, the first user's historical entry operation information, and the item information entry operation information; a second intelligent agent can determine the shooting strategy for the target item based on the reason; and the target item can be photographed based on the shooting strategy to obtain the photographed image.

[0039] It should be understood that a semantic similarity lower than the preset similarity threshold may be due to errors in the entry of item information into the asset management platform, or the photographed item may be unusual or complex (such as a circuit board with an unidentifiable brand or a handwritten document), making it impossible for the model to determine whether it matches the asset to be stored. Based on this, this embodiment of the disclosure utilizes collaboration between intelligent agents to perform in-depth diagnosis of the reasons for semantic similarity lower than the preset similarity threshold, thereby automatically triggering different refined processing procedures.

[0040] For example, the first analysis result includes a confidence level. The first agent can analyze the reason why the semantic similarity is lower than the preset similarity threshold based on the numerical relationship between this confidence level and a preset confidence threshold (e.g., 0.95). It should be understood that if the confidence level is greater than the preset confidence threshold, and there is a fundamental, category-level contradiction between the item information in the first analysis result and the item information of the items to be put into storage (e.g., different brands), considering that a well-trained model is unlikely to misidentify the brand, it is very likely that the item information entered by the case-related property management platform is incorrect.

[0041] For example, the first agent can analyze the first user's current inventory entry sequence. If the scan is taken immediately afterward, without prolonged pauses or other abnormal operations, the likelihood of picking the wrong item decreases. Simultaneously, the first agent analyzes the first user's historical operation accuracy. If the first user's historical operation accuracy is extremely high, the probability of picking the wrong item is low, while the possibility of incorrect item information entry relatively increases.

[0042] For example, the first intelligent agent can synchronize the information of seized items to be put into storage from the seized property management platform. If it is found that the item information has been modified by different people during the entry process, the probability of incorrect item information entry increases.

[0043] For example, the first intelligent agent can have a built-in decision function to fuse the above analysis results to determine the reason why the semantic similarity is lower than the preset similarity threshold. For instance, if the item to be stored is a Z brand mobile phone, and the target item in the real-time captured image is an X brand mobile phone, the first intelligent agent can output the reason why the semantic similarity is lower than the preset similarity threshold as: {Diagnostic criteria: 1. The item to be stored is a Z brand mobile phone, but the item captured in the real-time image is an X brand mobile phone, which is a fundamental conflict; 2. The data entry operation information shows that there were multiple modifications during the data entry process. Based on comprehensive judgment, this is a systemic data error that occurred in the information entry process.}

[0044] In some embodiments, the second intelligent agent is used to determine the shooting strategy for the target item as a first shooting strategy when the cause characterization item information is entered incorrectly. The first shooting strategy is used to generate an information correction request and send the information correction request to the case-related property management platform to prompt the administrator of the case-related property management platform to review and correct the item information in the case-related property management platform. In response to receiving the review and correction result of the item information, the agent performs a semantic comparison between the overall feature description and the corrected item information to obtain a new semantic similarity. If the new semantic similarity is higher than or equal to a preset similarity threshold, the agent triggers the camera to shoot the target item and obtain a captured image. The information correction request includes the overall feature description, item information, and the real-time captured image corresponding to the overall feature description.

[0045] For example, a prompt first appears on the first terminal device interface: "The photographed item may not match the items to be stored, suggesting a possible system information entry error." This helps the first user understand that the photographed item may not match the items to be stored. Simultaneously, the second intelligent agent can automatically submit an information correction request to the case-related property management platform, including a comparison of old and new information, on-site photos, and operator information, prompting the administrator to review and correct the item information database. After review and correction, a semantic comparison is performed again based on the corrected information. This automatically detects and corrects anomalies in the storage identification process, improving the intelligence level of case-related property storage and preventing the same error from recurring.

[0046] In some embodiments, the second intelligent agent is used to determine a second shooting strategy for shooting the target item when the cause characterizes the target item as a difficult item. The second shooting strategy is used to generate a collaborative verification request and send the collaborative verification request to the second user's second terminal device to prompt the second user to judge whether the target item matches the item information. In response to receiving the target feedback result sent by the second terminal device, the camera is triggered to shoot the target item and obtain the captured image. The collaborative verification request includes item information, real-time captured image corresponding to the first analysis result, and user information of the first user. The target feedback result characterizes that the target item matches the item information.

[0047] For example, a second intelligent agent automatically initiates a collaborative verification request, which is pushed in real-time via the intranet to the terminal device (i.e., the second terminal device) of an expert police officer in a designated field (i.e., the second user). The request includes: item information of the items to be stored, corresponding real-time captured images, and the information of the police officer initiating the verification (i.e., the user information of the first user). The expert police officer reviews the request on their own device, makes a judgment ("compliant" or "non-compliant"), and clicks feedback. If the feedback indicates compliance, the subsequent process can continue, triggering the camera to capture images of the target items. This transforms the single-point anomaly decision for capturing items into a distributed, network-based group decision, improving the ability to identify and process complex items involved in cases during the storage process.

[0048] In some embodiments, the real-time captured image may be captured by the camera at a first resolution. Accordingly, triggering the camera to capture an image of the target item includes triggering the camera to capture an image of the target item at a second resolution, wherein the second resolution is higher than the first resolution. Thus, smooth video stream analysis and rapid identification can be achieved through low-resolution real-time capture. After the real-time captured item is matched with the items to be stored, the camera's high-level interface can be invoked to command it to capture and save a still image at the highest hardware-supported resolution and lowest compression rate. This process can also enable optimization algorithms such as autofocus and automatic white balance to ensure image clarity, making the physical state of the captured image more consistent with the actual state of the items involved in the case.

[0049] After obtaining the captured image, the capture time and the user information of the first user can be determined. Then, at least the capture time, the user information, and the overall feature description are combined into plaintext. For example, the combined plaintext could be: {Capture Time: 20241008T154601.123Z, User Information: JC12345, Overall Feature Description: #Damaged Screen#}. Then, the plaintext is encrypted to generate an encrypted watermark. Finally, this encrypted watermark is embedded into the captured image to obtain the first watermarked image.

[0050] In some embodiments, the overall feature description and the captured image can be input into a second multimodal large model for pixel-level depth image analysis to obtain a second image analysis result. Based on the second image analysis result, a detailed feature description of the target object can be generated. Accordingly, the shooting time, user information, overall feature description, and detailed feature description can be combined into plaintext.

[0051] It should be understood that existing technologies lack the intelligent identification and verification capabilities for the unique characteristics of the seized property itself. They primarily rely on operators visually comparing the physical object with system information, which is inefficient and prone to errors due to subjective fatigue and negligence. Furthermore, the unique condition of the seized property (such as a phone with a cracked screen) (e.g., crack patterns, scratch distribution) is its most reliable identifying fingerprint. Existing technologies cannot automatically extract, record, and compare these visual features, nor can they form a verifiable state chain throughout the lifecycle of the seized item. When the condition of the seized item changes (whether due to natural aging or deliberate substitution), it cannot be automatically and accurately detected and alerted. Additionally, the first multimodal large model is a lightweight model, fast but potentially missing some unique details of the seized item.

[0052] Therefore, this embodiment of the disclosure uses a second multimodal large model to perform pixel-level depth image analysis, which facilitates the identification of detailed features of the items in question. Furthermore, while traditional watermarks can be removed or forged, the physical state of the items is difficult to perfectly reproduce. Therefore, adding detailed features to the watermark of the item image can improve the anti-counterfeiting capabilities of the item image, thereby enhancing the security of the items' storage.

[0053] For example, the second multimodal big data model can be deployed in the cloud for in-depth analysis of the physical condition of items (such as screen cracking, casing scratches, liquid leakage, special markings, etc.), outputting a detailed description in JSON format. In practical applications, the captured image and the recognition results of the first multimodal big data model can be combined into a prompt (e.g., "This is a smartphone involved in a case; please carefully check its appearance for damage, scratches, stains, or other abnormalities.") and submitted to the cloud-based second multimodal big data model. The second multimodal big data model can output a detailed feature description: {Screen cracked, severe scratches, suspected bloodstains in the lower right corner}. This fully utilizes the collaborative analysis capabilities of the local and cloud-based big data models to obtain a more refined and objective description of the features of the items involved in the case. While improving efficiency, it can also generate watermark information that better reflects the actual condition of the items, thereby enhancing the anti-counterfeiting strength of the watermark on the images of the items involved in the case, and ultimately improving the security of the items being stored in the warehouse.

[0054] In some embodiments, in response to the label printing operation of the target item, a physical label containing the item number and QR code of the target item can be generated based on the item information and detailed feature description. The physical label is used to affix to the target item for warehousing identification. The item information includes the item number, and the QR code is used to link to and display the detailed feature description corresponding to the target item.

[0055] Therefore, once the watermark is embedded, it can respond to the first user's label printing operation on the target item, driving the connected label printer to print a physical label containing information such as the item number and QR code. The first user then affixes this physical label to the target item, thus forming a mapping between the physical object, the label, and the watermark image, building a robust traceability chain. Furthermore, in addition to basic item information, the QR code can also contain a URL (Uniform Resource Locator) pointing to the target item's details page, where detailed feature descriptions generated by the second multimodal large model can be viewed, facilitating an understanding of the physical condition of the item in question.

[0056] In some embodiments, a second watermark image corresponding to the seized item can be obtained from the database of seized property. The encrypted watermark of the second watermark image is extracted, and the encrypted watermark of the second watermark image is decrypted to obtain the plaintext corresponding to the second watermark image. Information of specific fields is extracted from the plaintext corresponding to the second watermark image to obtain a first item feature description. A hash calculation is performed based on the first item feature description to obtain a first feature hash value. The second watermark image is input into a second multimodal large model for deep image analysis to obtain a third image analysis result. Based on the third image analysis result, a second item feature description corresponding to the second watermark image is generated. A hash calculation is performed based on the second item feature description to obtain a second feature hash value. If the first feature hash value and the second feature hash value are not equal, a tampering prompt message is output. The tampering prompt message is used to indicate that the watermark image corresponding to the seized item in the database of seized property has been tampered with.

[0057] For example, for a second watermarked image obtained from a database of seized assets, a dedicated watermark extraction tool is used to analyze the image's pixel data, searching in the frequency domain to extract the previously embedded, invisible encrypted watermark. This encrypted watermark is then sent to the decryption interface of the watermark service engine. This interface accesses the hardware security module and uses the same key used when embedding the watermark to decrypt it, obtaining the original plaintext information. Feature description fields (such as "damaged screen") are extracted from this plaintext information, and a hash calculation is performed based on these fields to obtain the first feature hash value. Additionally, the second watermarked image is input into a second multimodal big data model, instructing it to analyze the state of the items in the image. Based on the current content of the second watermarked image, the second multimodal big data model generates a new, structured description of the item's features, and then performs a hash calculation based on this description to obtain the second feature hash value representing the current state of the seized items in the second watermarked image.

[0058] It should be understood that when the first feature hash value and the second feature hash value are equal, the item status of the watermarked image in the database of the involved property is completely consistent with the state when the watermark was embedded, and the image is highly likely not to have been tampered with. When the first feature hash value and the second feature hash value are not equal, it means that there is a difference between the two, which may indicate that the image content has been tampered with, thus prompting the output of a tampering warning message.

[0059] In some embodiments, when the first feature hash value and the second feature hash value are not equal, a tampering prompt message is output, including: when the first feature hash value and the second feature hash value are not equal, verifying whether the second watermark image has been tampered with based on the first item feature description, the second item feature description, the plaintext corresponding to the second watermark image, the operation log corresponding to the item in question, and the warehouse monitoring video corresponding to the item in question; and outputting a tampering prompt message when the second watermark image has been tampered with.

[0060] It should be understood that the first feature hash value and the second feature hash value are not equal, which could indicate that the image content has been tampered with, or that the item itself has subsequently suffered new damage. Therefore, to ensure the accuracy of the tampering warning, it is possible to further verify whether the watermark image of the item has been altered.

[0061] In some embodiments, verifying whether the second watermark image has been tampered with, based on a first item feature description, a second item feature description, plaintext corresponding to the second watermark image, operation logs corresponding to the item in question, and monitoring video of the storage location corresponding to the item in question, includes: comparing the first item feature description and the second item feature description to obtain a comparison result; determining the storage status of the item in question based on the operation logs corresponding to the item in question; determining the target time point for obtaining the second watermark image from the database of the item in question, and capturing target monitoring footage within a preset time range before and after the target time point from the monitoring video of the storage location corresponding to the item in question, comparing the visual features of the item in question in the target monitoring footage with those in the second watermark image to obtain a visual feature comparison result; and verifying whether the second watermark image has been tampered with based on the comparison result, the storage status, and the visual feature comparison result.

[0062] For example, natural language processing techniques can be used to compare the differences between the first and second item feature descriptions. If the difference is local and subtle (e.g., the original description is "the screen has a 3 cm crack," while the new description is "the screen has a 5 cm crack and an additional scratch"), it is preliminarily judged that the object's condition is likely due to natural evolution or accidental damage. If the difference is global and disruptive (e.g., the original description is "the screen has a 3 cm crack," while the new description is "the screen is intact"), the engine preliminarily judges that the image is extremely likely to have been maliciously altered.

[0063] Additionally, all operation records for the item since it entered the warehouse can be retrieved. If the operation records show legitimate outbound-return records within that time period, the likelihood of the item's physical condition changing during the outbound period increases dramatically. If the item has remained in inventory since entering the warehouse without any outbound records, the credibility of a change in physical condition decreases, while the possibility of image tampering increases.

[0064] Additionally, the system can automatically call the warehouse video management system's interface to quickly locate and capture keyframes of the surveillance footage around the target time point, based on the location number of the items in question and the target time point for obtaining the second watermark image. Then, a rapid visual feature similarity comparison is performed between the second watermark image and the actual item in the surveillance video. If the actual item's condition in the surveillance footage is highly consistent with the leaked image (high feature similarity), it can be determined that the second watermark image is genuine, but the actual item's condition has been altered. If the actual item's condition in the surveillance footage is inconsistent with the second watermark image (e.g., the surveillance screen is still damaged, while the leaked image shows an intact screen), it proves that the second watermark image has been tampered with or forged.

[0065] Finally, based on the above automatic analysis, a structured tampering verification result can be output: {Preliminary diagnosis: High probability of physical state alteration, based on: the state change pattern conforms to natural evolution characteristics, and the item has an outbound record}. Alternatively, a structured tampering verification result can be output as: {Preliminary diagnosis: High probability of image tampering, based on: the state change pattern shows an unreasonable reversal (damaged to intact), and the video surveillance footage is inconsistent with the leaked image}.

[0066] Therefore, it can automatically aggregate evidence from multiple sources such as watermark information, operation logs, and monitoring footage, which greatly facilitates the investigation of watermark image tampering. This allows for a reliable preliminary classification of the root cause of watermark image tampering, improving the efficiency and accuracy of subsequent investigations.

[0067] The marking method for the entry of seized property provided in this disclosure has the following advantages:

[0068] Strong traceability capability: Once an image is leaked, the specific photographer and the time of the photo can be quickly and accurately identified, providing solid evidence for internal accountability and external investigation.

[0069] A powerful deterrent: When users know that their identity information is linked to an image, they will be much more vigilant, reducing the likelihood of intentional leaks at the source.

[0070] Completeness of the chain of evidence: It provides an tamper-proof digital identity for the visual records of the property involved in the case, enhancing the evidentiary value of the entire management process.

[0071] Operational concealment: The hidden watermark does not affect the observation and viewing of the image, maintaining the original appearance of the evidence image.

[0072] According to a second aspect of the present disclosure, a marking system is provided for the storage of seized property. Please refer to the appendix. Figure 2 The marking system 200 applied to the storage of seized property includes a seized property management platform 201, a first terminal device 202, a watermark service engine 203, and a seized property database 204.

[0073] The case-related property management platform 201 is used to input the item information of case-related items that are to be put into storage;

[0074] The first terminal device 202 is used to respond to the first user's operation of turning on the camera, perform target detection and image recognition on the real-time shooting scene captured by the camera through the first multimodal large model, obtain a first analysis result, and generate an overall feature description of the target object in the real-time shooting scene based on the first analysis result. The overall feature description is semantically compared with the object information to obtain a semantic similarity. If the semantic similarity is higher than or equal to a preset similarity threshold, the camera is triggered to take a picture of the target object to obtain a captured image.

[0075] The watermark service engine 203 is used to determine the shooting time of the captured image and the user information of the first user, and at least combine the shooting time, the user information and the overall feature description into plaintext, encrypt the plaintext to obtain an encrypted watermark, and embed the encrypted watermark into the captured image to obtain a first watermarked image;

[0076] The database of involved assets 204 is used to associate and store the first watermark image and the item information in the data entries corresponding to the involved items.

[0077] In some embodiments of this disclosure, the marking system 200 applied to the storage of seized property further includes a first intelligent agent and a second intelligent agent, wherein the first terminal device 202 is used for:

[0078] If the semantic similarity is lower than the preset similarity threshold, the first intelligent agent analyzes the reason why the semantic similarity is lower than the preset similarity threshold based on the first analysis result, the first user's current entry operation information, the first user's historical entry operation information, and the item information entry operation information.

[0079] Based on the aforementioned reasons, a second intelligent agent determines a shooting strategy for the target item.

[0080] The target object is photographed based on the aforementioned shooting strategy to obtain a photographed image.

[0081] In some embodiments of this disclosure, the second intelligent agent is used to determine a first shooting strategy for the target item when the cause indicates that the item information is entered incorrectly. The first shooting strategy is used to generate an information correction request and send the information correction request to the case-related property management platform to prompt the administrator of the case-related property management platform to review and correct the item information in the case-related property management platform. In response to receiving the review and correction result of the item information, the agent performs a semantic comparison between the overall feature description and the corrected item information to obtain a new semantic similarity. If the new semantic similarity is higher than or equal to the preset similarity threshold, the agent triggers the camera to take a picture of the target item to obtain a captured image. The information correction request includes the overall feature description, the item information, and the real-time captured image corresponding to the overall feature description.

[0082] In some embodiments, the second intelligent agent is configured to determine a second shooting strategy for shooting the target item when the cause indicates that the target item is a questionable item. The second shooting strategy is configured to generate a collaborative verification request and send the collaborative verification request to the second user's second terminal device to prompt the second user to judge whether the target item matches the item information. In response to receiving the target feedback result sent by the second terminal device, the agent triggers the camera to shoot the target item and obtain a captured image. The collaborative verification request includes the item information, the real-time captured image corresponding to the first analysis result, and the user information of the first user. The target feedback result indicates that the target item matches the item information.

[0083] In some embodiments, the watermarking service engine 203 is used to input the overall feature description and the captured image into a second multimodal large model for pixel-level deep image analysis to obtain a second image analysis result, and based on the second image analysis result, generate a detailed feature description of the target item; and combine the shooting time, the user information, the overall feature description and the detailed feature description into plaintext.

[0084] In some embodiments, the involved property management platform 201 is used to generate a physical label containing the item number and a QR code based on the item information and the detailed feature description in response to the label printing operation of the target item. The physical label is used to affix to the target item for warehousing identification. The item information includes the item number, and the QR code is used to link to and display the detailed feature description corresponding to the target item.

[0085] In some embodiments, the case-related property management platform 201 is used to obtain a second watermark image corresponding to the case-related item from the case-related property database, extract the encrypted watermark of the second watermark image, decrypt the encrypted watermark of the second watermark image to obtain the plaintext corresponding to the second watermark image, extract information of specific fields from the plaintext corresponding to the second watermark image to obtain a first item feature description, and perform hash calculation based on the first item feature description to obtain a first feature hash value;

[0086] The second watermark image is input into the second multimodal large model for deep image analysis to obtain a third image analysis result. Based on the third image analysis result, a second item feature description corresponding to the second watermark image is generated. Based on the second item feature description, a hash calculation is performed to obtain a second feature hash value.

[0087] If the first feature hash value and the second feature hash value are not equal, a tampering prompt message is output, wherein the tampering prompt message is used to indicate that the watermark image corresponding to the case-related item in the case-related property database has been tampered with.

[0088] In some embodiments, the case-related property management platform 201 is used to verify whether the second watermark image has been tampered with, based on the first item feature description, the second item feature description, the plaintext corresponding to the second watermark image, the operation log corresponding to the case-related item, and the warehouse monitoring video corresponding to the case-related item, when the first feature hash value and the second feature hash value are not equal.

[0089] If the second watermark image is tampered with, a tampering warning message will be output.

[0090] In some embodiments, the case-related property management platform 201 is used to compare the first item feature description and the second item feature description to obtain a difference comparison result;

[0091] Based on the operation logs corresponding to the items involved in the case, the storage status of the items involved in the case is determined;

[0092] Determine the target time point for obtaining the second watermark image from the database of involved assets, and extract the target monitoring screen within a preset time range before and after the target time point from the monitoring video of the storage location corresponding to the involved items. Compare the visual features of the involved items in the target monitoring screen with the involved items in the second watermark image to obtain the visual feature comparison result.

[0093] Based on the difference comparison results, the storage status, and the visual feature comparison results, it is verified whether the second watermark image has been tampered with.

[0094] Regarding the system in the above embodiments, the specific manner in which each module performs its operations has been described in detail in the embodiments of the method in the first aspect, and will not be elaborated upon here.

[0095] The preferred embodiments of this disclosure have been described in detail above with reference to the accompanying drawings. However, this disclosure is not limited to the specific details of the above embodiments. Within the scope of the technical concept of this disclosure, various simple modifications can be made to the technical solutions of this disclosure, and these simple modifications all fall within the protection scope of this disclosure.

[0096] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, this disclosure will not describe the various possible combinations separately.

[0097] Furthermore, various different embodiments of this disclosure can be combined in any way, as long as they do not violate the spirit of this disclosure, they should also be regarded as the content disclosed in this disclosure.

Claims

1. A marking method applied to the storage of seized property, characterized in that, include: Obtain information on seized items awaiting warehousing from the seized property management platform; In response to the first user's operation to turn on the camera in the first terminal device, a first multimodal large model is used to perform target detection and image recognition on the real-time captured image by the camera to obtain a first analysis result. Based on the first analysis result, an overall feature description of the target item in the real-time captured image is generated. The overall feature description is semantically compared with the item information to obtain a semantic similarity. If the semantic similarity is higher than or equal to a preset similarity threshold, the camera is triggered to take a picture of the target item to obtain a captured image. If the semantic similarity is lower than the preset similarity threshold, a first intelligent agent analyzes the reason why the semantic similarity is lower than the preset similarity threshold based on the first analysis result, the first user's current entry operation information, the first user's historical entry operation information, and the item information entry operation information. A second intelligent agent determines a shooting strategy for the target item based on the reason, and takes a picture of the target item based on the shooting strategy to obtain a captured image. The shooting time of the captured image and the user information of the first user are determined. The overall feature description and the captured image are input into a second multimodal large model for pixel-level depth image analysis to obtain a second image analysis result. Based on the second image analysis result, a detailed feature description of the target item is generated. The shooting time, the user information, the overall feature description and the detailed feature description are combined into plaintext. The plaintext is encrypted to obtain an encrypted watermark. The encrypted watermark is embedded in the captured image to obtain a first watermark image. The first watermark image and the item information are associated and stored in the data entry corresponding to the item in the database of involved property.

2. The marking method for the entry of seized property into the warehouse according to claim 1, characterized in that, The second intelligent agent is used to determine the shooting strategy for the target item as a first shooting strategy when the cause indicates that the item information is entered incorrectly. The first shooting strategy is used to generate an information correction request and send the information correction request to the case-related property management platform to prompt the administrator of the case-related property management platform to review and correct the item information in the case-related property management platform. In response to receiving the review and correction result of the item information, the agent performs a semantic comparison between the overall feature description and the corrected item information to obtain a new semantic similarity. If the new semantic similarity is higher than or equal to the preset similarity threshold, the agent triggers the camera to shoot the target item and obtain a captured image. The information correction request includes the overall feature description, the item information, and the real-time captured image corresponding to the overall feature description.

3. The marking method for the entry of seized property into the warehouse according to claim 1, characterized in that, The second intelligent agent is used to determine a second shooting strategy for the target item when the cause indicates that the target item is a questionable item. The second shooting strategy is used to generate a collaborative verification request and send the collaborative verification request to the second user's second terminal device to prompt the second user to judge whether the target item matches the item information. In response to receiving the target feedback result sent by the second terminal device, the camera is triggered to shoot the target item and obtain a captured image. The collaborative verification request includes the item information, the real-time captured image corresponding to the first analysis result, and the user information of the first user. The target feedback result indicates that the target item matches the item information.

4. The marking method for the entry of seized property into the warehouse according to claim 1, characterized in that, Also includes: In response to the label printing operation of the target item, a physical label containing the item number and QR code of the target item is generated based on the item information and the detailed feature description. The physical label is used to affix to the target item for warehousing identification. The item information includes the item number, and the QR code is used to link to and display the detailed feature description corresponding to the target item.

5. The marking method for the entry of seized property into the warehouse according to claim 1, characterized in that, Also includes: The second watermark image corresponding to the involved item is obtained from the database of involved property. The encrypted watermark of the second watermark image is extracted. The encrypted watermark of the second watermark image is decrypted to obtain the plaintext corresponding to the second watermark image. Information of specific fields is extracted from the plaintext corresponding to the second watermark image to obtain the first item feature description. A hash calculation is performed based on the first item feature description to obtain the first feature hash value. The second watermark image is input into the second multimodal large model for deep image analysis to obtain a third image analysis result. Based on the third image analysis result, a second item feature description corresponding to the second watermark image is generated. Based on the second item feature description, a hash calculation is performed to obtain a second feature hash value. If the first feature hash value and the second feature hash value are not equal, a tampering prompt message is output, wherein the tampering prompt message is used to indicate that the watermark image corresponding to the case-related item in the case-related property database has been tampered with.

6. The marking method for the entry of seized property into the warehouse according to claim 5, characterized in that, The step of outputting a tampering warning message when the first feature hash value and the second feature hash value are not equal includes: If the first feature hash value and the second feature hash value are not equal, the second watermark image is verified as having been tampered with based on the first item feature description, the second item feature description, the plaintext corresponding to the second watermark image, the operation log corresponding to the item in question, and the warehouse monitoring video corresponding to the item in question. If the second watermark image is tampered with, a tampering warning message will be output.

7. The marking method for the entry of seized property into the warehouse according to claim 6, characterized in that, The step of verifying whether the second watermark image has been tampered with, based on the first item feature description, the second item feature description, the plaintext corresponding to the second watermark image, the operation log corresponding to the item in question, and the warehouse monitoring video corresponding to the item in question, includes: The first item feature description and the second item feature description are compared to obtain the comparison result. Based on the operation logs corresponding to the items involved in the case, the storage status of the items involved in the case is determined; Determine the target time point for obtaining the second watermark image from the database of involved assets, and extract the target monitoring screen within a preset time range before and after the target time point from the monitoring video of the storage location corresponding to the involved items. Compare the visual features of the involved items in the target monitoring screen with the involved items in the second watermark image to obtain the visual feature comparison result. Based on the difference comparison results, the storage status, and the visual feature comparison results, it is verified whether the second watermark image has been tampered with.

8. A marking system applied to the storage of seized property, characterized in that, This includes a platform for managing seized assets, primary terminal equipment, a watermarking service engine, and a database of seized assets. The platform for managing seized assets is used to input information about seized items that are to be put into storage. The first terminal device is used to respond to the first user's camera activation operation, perform target detection and image recognition on the real-time captured image by the camera using a first multimodal large model, obtain a first analysis result, and generate an overall feature description of the target item in the real-time captured image based on the first analysis result. The overall feature description is semantically compared with the item information to obtain semantic similarity. If the semantic similarity is higher than or equal to a preset similarity threshold, the camera is triggered to take a picture of the target item to obtain a captured image. If the semantic similarity is lower than the preset similarity threshold, the first intelligent agent analyzes the reason why the semantic similarity is lower than the preset similarity threshold based on the first analysis result, the first user's current entry operation information, the first user's historical entry operation information, and the item information entry operation information. The second intelligent agent determines a shooting strategy for the target item based on the reason, and takes a picture of the target item based on the shooting strategy to obtain a captured image. The watermarking service engine is used to determine the shooting time of the captured image and the user information of the first user. It inputs the overall feature description and the captured image into a second multimodal large model for pixel-level deep image analysis to obtain a second image analysis result. Based on the second image analysis result, it generates a detailed feature description of the target item. It combines the shooting time, the user information, the overall feature description, and the detailed feature description into plaintext, encrypts the plaintext to obtain an encrypted watermark, and embeds the encrypted watermark into the captured image to obtain a first watermarked image. The database of seized assets is used to associate and store the first watermark image and the item information in the data entries corresponding to the seized items.

Images