A method for verifying image features of a service robot in a detection scene

Abstract

The application discloses a kind of service robot image feature preservation encryption methods for testing detection scene, comprising the following steps: (1) testing detection scene's secret level division;(2) random key generation;(3) robot image feature preservation encryption and decryption;(4) multi-level scene image feature preservation encryption;(5) multi-level image feature preservation encryption for scene content.The application selects desensitization laxity dynamic calculation different level plaintext weight, then adopts image weighted fusion model to fuse plaintext and ciphertext image, realizes multi-level scene image feature preservation encryption;When multi-level image feature preservation encryption for scene content, dynamic recursion method is used for classified encryption in secret area, high desensitization laxity is set for background area, and image feature preservation edge processing strategy is applied to optimize the irregular segmentation edge of image, balance the sawtooth defect brought by irregular contour and edge encryption effect.

Description

Technical Field

[0001] This invention belongs to the field of robot vision and information security technology, specifically relating to a method for preserving and encrypting image features of service robots in inspection and detection scenarios. Background Technology

[0002] With the intelligent upgrading of the testing and inspection industry, service robots are widely deployed in various laboratories to undertake tasks such as sample transportation and comprehensive inspection, thereby improving the efficiency of laboratory operations and daily maintenance. The core operational processes of service robots in testing and inspection scenarios include fixed-point check-in across confidential areas, environmental photography, point location verification, and image tracking of sample flow. Each time a service robot arrives at a preset inspection point, it must capture on-site images and transmit them back to the backend for confirmation of inspection arrival, generation of electronic ledgers, and traceability of sample flow paths. Laboratories need to maintain complete inspection image archives long-term to support sample dispute review, equipment fault diagnosis, on-site regulatory verification, and internal audits of the laboratory quality system. Service robots can generate a large number of cross-regional on-site images daily. Backend maintenance and quality control personnel need to browse images in batches, quickly distinguish shooting points, verify check-in records, and locate abnormal areas, creating a rigid business demand for online image preview and rapid retrieval. In actual operation, service robots need to continuously move between laboratories and functional office areas with different security levels to complete cross-regional tasks such as sample delivery. However, testing and inspection laboratories contain classified physical spaces, and when the visual sensors on the robots capture images, they inevitably capture classified information such as confidential instruments and equipment, personnel, samples, and sensitive test parameters. Regarding encryption requirements for service robots, two national standards, "Safety Requirements for Service Robots" (GB / T 44253-2024) and "General Requirements for Information Security of Service Robots" (GB / T45502-2025), have already made provisions. This presents a practical business contradiction: back-end maintenance and quality control personnel should not directly view the classified and private content in the images, but at the same time, they need to use the rough visual features retained in the images to complete online previews and quick searches, enabling the identification of shooting locations and verification of attendance records. Therefore, it is urgent to balance the business needs of image preview and quick retrieval while protecting the classified and private information in the images captured by the robots. Traditional encryption algorithms such as AES lose the ability to preview images. While simple visual desensitization methods such as static mosaic and global blur can restore sensitive details through image restoration AI algorithms, they do not achieve standardized reversible encryption based on cryptography and are subject to human tampering risks, making them unsuitable as effective evidence for CMA review and accountability for quality and safety incidents. Therefore, thumbnail-preserving encryption (TPE) can be introduced. This algorithm preserves the coarse-grained scene outline of the image through encryption, allowing for online preview without decryption. Back-end maintenance and quality control personnel can quickly distinguish between the laboratory and inspection points based on prior visual knowledge without exposing privacy in the image. Furthermore, encrypted images can be archived and stored long-term for data transmission and back-end storage. After decryption, the original scene is fully restored, balancing business retrieval and confidentiality requirements.

[0003] However, the introduction of the Thumbnail Preservation Encryption (TPE) algorithm still has many shortcomings in adapting to cross-regional, multi-level testing and inspection laboratories:

[0004] 1) The current classification scheme only classifies the confidentiality level for a single independent laboratory, and does not cover the entire space of the testing and inspection laboratory (public areas, office areas, general laboratories, key classified laboratories, etc.). It does not take into account the characteristics of service robots continuously inspecting across the entire laboratory, lacks a unified confidentiality level classification system for the entire testing and inspection field, and cannot configure gradient and hierarchical encryption strength according to the regional confidentiality level, or set different encryption strengths for different confidentiality level scenarios, which is prone to cross-regional image data leakage risk;

[0005] 2) Although the Thumbnail Preservation Encryption (TPE) algorithm can use dynamic encryption sub-blocks to adapt to the encryption requirements of different levels in inspection and detection scenarios, the size of the encryption sub-blocks is generally limited to an integer power of 2 due to considerations such as memory alignment, parallel computing acceleration, and compatibility with existing image encoding and decoding standard macroblocks. Combined with the diverse image resolutions in the work scenarios, this can easily lead to problems such as large parameter jumps, excessively large intervals between levels, and discontinuous parameter adjustments. As a result, it is difficult to accurately control the detail processing effect of image feature preservation encryption. In addition, there are hierarchical encryption requirements in actual multi-level inspection and detection scenarios. High-density scenarios need to shield privacy details, medium-to-high-secrecy scenarios need to adaptively balance security and scene visibility, and low-density scenarios have high visibility and moderate encryption.

[0006] 3) In some complex robot inspection scenarios, the robot inspection images contain a large amount of differentiated scene content. The images include highly confidential core areas such as laboratory staff, precision testing instruments, and experimental samples to be tested, as well as ordinary background areas with no confidentiality value such as open ground, walls, tables and chairs. The Thumbnail Preservation Encryption (TPE) algorithm uses a uniform encryption mode to encrypt the entire image in a uniform block, applying the same encryption strength to confidential key areas and invalid backgrounds. It cannot achieve differentiated encryption control of multiple areas and multiple levels within a single image, which can easily lead to over-encryption of details in highly confidential core areas, while the background areas with high discrimination but no encryption significance are blurred and distorted. It does not take into account both encryption security and image readability, and does not achieve the effect of keeping key areas confidential and background areas clearly visible. This is not conducive to legitimate users with prior knowledge of the scene conducting accurate reasoning and recognition analysis of the image content. Summary of the Invention

[0007] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a method for encrypting and preserving image features of service robots in inspection and detection scenarios.

[0008] This invention provides a novel method for preserving and encrypting image features of service robots in inspection and testing scenarios.

[0009] The above-mentioned objective of the present invention can be achieved through the following technical solution: a method for preserving and encrypting image features of a service robot in an inspection and detection scenario, comprising the following steps:

[0010] (1) Classification of confidentiality levels for inspection and testing scenarios:

[0011] Service robot operation inspection and testing scenarios are divided into five confidentiality levels from L1 to L5, and corresponding desensitization relaxation value ranges are set for each level.

[0012] (2) Random key generation:

[0013] Select unique operation process parameters of the robot as key parameters and generate a random key stream;

[0014] (3) Robot image feature preservation encryption and decryption:

[0015] The Thumbnail Preservation Encryption (TPE) algorithm is used to encrypt the features of images captured by the service robot, generating ciphertext images for transmission and storage. The original image, i.e., the plaintext image, is then decrypted using the same key stream used for encryption.

[0016] (4) Encryption of multi-level scene image features:

[0017] Based on the confidentiality level of the inspection and testing scenario in which the service robot is located, the desensitization leniency is determined within the corresponding level range. A three-stage differentiated plaintext weight model is adopted, and the plaintext weights of different levels are dynamically calculated based on the desensitization leniency. Then, an image weighted fusion model is used to fuse plaintext and ciphertext images to achieve encryption while maintaining the image features of multi-level scenarios.

[0018] The three-stage differentiated plaintext weight model includes a tenth power concave function model, an adaptive controllable Sigmoid model, and a bounded exponential convergence model. The L5 scenario uses the tenth power concave function model, the L2-L4 scenarios use the adaptive controllable Sigmoid model, and the L1 scenario uses the bounded exponential convergence model.

[0019] (5) Encryption of multi-level image features for scene content preservation:

[0020] The service robot first identifies the inspection and testing scenario it is in and determines the corresponding confidentiality level. Based on the confidentiality level scenario, it uses a panoramic image segmentation algorithm to segment and a recognition and classification algorithm to identify, classify and fuse the scenario content, divide the scenario content into confidential areas and background areas, divide it into different levels according to the confidentiality risk level of the scenario content, and then perform image feature preservation edge division.

[0021] A dynamic recursive method is used to perform hierarchical encryption of classified areas, setting a low desensitization leniency for classified areas and a high desensitization leniency for background areas;

[0022] A scene content edge processing strategy is adopted to encrypt the scene content edges, thereby achieving multi-level image feature preservation encryption oriented towards scene content.

[0023] In the above-mentioned service robot image feature preservation encryption method for inspection and testing scenarios:

[0024] Preferably, in step (1), the desensitization leniency range for scenario L1 is (0.8, 1], for scenario L2 it is (0.6, 0.8], for scenario L3 it is (0.4, 0.6], for scenario L4 it is (0.2, 0.4], and for scenario L5 it is (0, 0.2]. Among them, scenario L1 is a public open area, scenario L2 is an office area, scenario L3 is a functional laboratory, scenario L4 is a confidential office and a medium-risk laboratory, and scenario L5 is a high-confidential laboratory and a core testing area.

[0025] Preferably, step (2) of selecting unique operation process parameters of the robot as key parameters and generating a random key stream includes: selecting unique operation process parameters of the service robot as key parameters and normalizing them, and using a mapping-coupling iterative algorithm to generate an encrypted key stream to ensure that the key stream generation result is random each time, specifically including:

[0026] (2.1) Unique key parameter acquisition: Select the service robot equipment serial number, inspection three-dimensional space coordinates, fuselage heading angle and inspection timestamp as unique operation process parameters of the service robot, and use the unique operation process parameters as the original input parameters for key generation;

[0027] (2.2) Key parameter normalization: The four types of original key parameters, namely service robot equipment serial number, inspection three-dimensional space coordinates, fuselage heading angle and inspection timestamp, are uniformly constrained to the interval [0, 1) and normalized to eliminate the interference of heterogeneous parameter magnitude differences on the key generation algorithm iteration;

[0028] (2.3) Random key stream generation: The normalized unique key parameters are used as the external perturbation input of the algorithm. The mapping-coupled iterative algorithm is adopted. Relying on the chaotic divergence effect generated by the algorithm, a random key stream with non-repeating results in each iteration is generated through multiple rounds of iterative calculation.

[0029] Preferably, step (3) employs the Thumbnail Preservation Encryption (TPE) algorithm to encrypt the image features captured by the service robot, generating a ciphertext image for transmission and storage. During the viewing phase, the original image, i.e., the plaintext image, is decrypted using the same key stream used for encryption. Specifically, this includes:

[0030] (3.1) Selection of encrypted sub-block size: The side length N of the encrypted sub-block is selected as:

[0031] (1)

[0032] Where N is the side length of the encrypted sub-block, round(·) means rounding to the nearest integer, and Pm and Pn refer to the width and height of the image, respectively;

[0033] (3.2) The Thumbnail Preservation Encryption (TPE) algorithm is adopted to divide the image grid into blocks. Using the key stream, the permutation sequence is generated and the permutation encryption sub-block pixels are rearranged in a loop. The full-domain block permutation encryption image is output for transmission and storage. While completing the image pixel encryption, the coarse-grained scene features of the image are preserved.

[0034] (3.3) Decryption while preserving robot image features: Solve the inverse permutation index using the same key stream as encryption, cyclically permutate to recover sub-block pixels, fully decrypt and recover the original image.

[0035] Preferably, the three-segment differentiated plaintext weight model described in step (4) is as shown in formula (8):

[0036] (8);

[0037] in: W ori (α) represents the three-stage differentiated plaintext weights. The first stage uses the tenth-power concave function model for plaintext weight calculation in the L5 scenario; the second stage uses the adaptive controllable sigmoid model for plaintext weight calculation in the L2-L4 scenarios; and the third stage uses the bounded exponential convergence model for plaintext weight calculation in the L1 scenario. α represents the desensitization leniency. V Max V The desensitization leniency thresholds for L5, L2-L4, and L1 scenarios are set to 0.2, 0.8, and W, respectively. L W M Let α be the plaintext weight with threshold values ​​of 0.2 and 0.8 respectively, and k, λ, and Max be the weights. W The steepness coefficient, adaptive convexity coefficient, and global scaling coefficient in the second adaptive controllable Sigmoid model are [4.9, 5.1], [0.68, 0.72], and [0.6, 0.7], respectively, with optimal values ​​of 5.0, 0.7, and 0.65. k2 and Max...END These are the steepness coefficient and the maximum convergence upper limit coefficient in the third bounded exponential convergence model, respectively [5.8, 6.1] and [0.75, 0.83], with 6.0 and 0.8 being preferred.

[0038] Furthermore, the encryption of the multi-level scene image features in step (4) specifically includes:

[0039] (4.1) High-security scenario settings: For the L5 scenario, a tenth power concave function model is used to obtain plaintext weights to maximize image encryption strength;

[0040] (4.2) Controlled Scene Setup: For L2-L4 scenes, an adaptive controllable Sigmoid model is used to obtain plaintext weights. Based on the Sigmoid model, a steepness coefficient k is introduced to adjust the rate of increase of weights in the interval. An adaptive convexity coefficient λ is introduced to control the amplitude of the bulge in the middle of the curve. A global scaling coefficient Max is introduced. W Constrain the range of plaintext weights to achieve controllable plaintext and ciphertext weights;

[0041] (4.3) Open Scenario Settings: For the L1 scenario, set the steepness coefficient k2 and the maximum convergence upper limit coefficient Max. END The plaintext weights are obtained by using a bounded exponential convergence model, which allows the plaintext weights to rise steadily and converge to the maximum convergence limit.

[0042] (4.4) Multi-level scene image feature preservation encryption: Randomly select the desensitization leniency in the corresponding interval of different level scenes, use a three-segment differentiated plaintext weight model to calculate the plaintext weight of different levels, and then use an image weighted fusion model to fuse plaintext images and ciphertext images to achieve multi-level scene image feature preservation encryption from L1 to L5.

[0043] Preferably, step (5) employs a panoramic image segmentation algorithm for segmentation and a recognition and classification algorithm for recognition, classification, and fusion. This includes using a panoramic image segmentation algorithm to segment and detect scene image targets, generating a target mask set, using a recognition and classification algorithm to extract target regions within the mask and identify and classify them, and fusing image features of the same category into one region. Specifically, this includes:

[0044] (5.1.1) Inspection and detection scene image segmentation: The panoramic image segmentation algorithm is adopted. Through panoramic grid point sampling and hierarchical cropping inference mechanism, the target in the inspection and detection scene image is segmented in the panoramic view to generate a target mask set;

[0045] (5.1.2) Image feature recognition and classification: The image features of the target area within the mask are extracted using a recognition and classification algorithm. The cosine similarity is calculated between the image features and the text features of each category of scene content. The category corresponding to the maximum similarity is selected as the label content of the current fragment mask.

[0046] (5.1.3) Fusion of features of the same category: For the large number of fragmented masks output by image segmentation, a fusion strategy of bitwise OR of the masks is used to merge the fragments with adjacent identical label content into a complete large region.

[0047] Preferably, the classified area in step (5) includes personnel categories, testing instruments and samples, and other items, and the background area includes scene background, which is divided into different levels according to the classified risk level of the scene content, from high to low as follows: (5.2.1) personnel categories, (5.2.2) testing instruments and samples, (5.2.3) other items, and (5.2.4) scene background.

[0048] Preferably, in step (5), a dynamic recursive method is used to perform hierarchical encryption of classified areas, including setting minimum spacing constraints between classified area levels. The dynamic recursive method is used to transform global constraints into local security sampling intervals by dynamically adjusting the upper and lower limits, ensuring that there is a minimum spacing between the desensitization slackness between levels, thereby improving the distinguishability. The background area randomly obtains L1 scene desensitization slackness encryption to improve scene distinguishability.

[0049] Preferably, in step (5), a dynamic recursive method is used to perform hierarchical encryption of the classified area, setting a low desensitization leniency for the classified area and a high desensitization leniency for the background area. The selection range of the scene content hierarchical encryption desensitization leniency is as follows:

[0050] For scenario L1, the desensitization leniency for the person category U1 is (0.8, 0.899998], the desensitization leniency for the detection instrument and sample U2 is (U1+0.050001, 0.949999], the desensitization leniency for other items U3 is (U2+0.050001, 1], and the desensitization leniency for the scene background is 1;

[0051] For scenario L2, the desensitization leniency for person category U1 is (0.6, 0.699998], the desensitization leniency for detection instrument and sample U2 is (U1+0.050001, 0.749999], the desensitization leniency for other items U3 is (U2+0.050001, 0.8], and the desensitization leniency for scenario background is (0.8, 1.0].

[0052] For scenario L3, the desensitization leniency for person category U1 is (0.4, 0.499998], the desensitization leniency for detection instrument and sample U2 is (U1+0.050001, 0.549999], the desensitization leniency for other items U3 is (U2+0.050001, 0.6], and the desensitization leniency for scenario background is (0.8, 1.0];

[0053] For scenario L4, the desensitization leniency for person category U1 is (0.2, 0.299998], the desensitization leniency for detection instrument and sample U2 is (U1+0.050001, 0.349999], the desensitization leniency for other items U3 is (U2+0.050001, 0.4], and the desensitization leniency for scenario background is (0.8, 1.0];

[0054] For scenario L5, the desensitization leniency for the character category U1 is (0, 0.099998), the desensitization leniency for the detection instrument and sample U2 is (U1+0.050001, 0.149999), the desensitization leniency for other items U3 is (U2+0.050001, 0.2), and the desensitization leniency for the scene background is (0.8, 1.0).

[0055] Preferably, the image feature preservation edge division in step (5) includes: dynamically determining the block type using the mask pixel coverage ratio, dividing the encrypted sub-blocks of the confidential area into three categories: core block, edge block, and background block, which facilitates the subsequent hierarchical configuration of the desensitization leniency parameter and achieves a balance between contour jaggedness suppression and edge encryption effect, wherein:

[0056] (5.2.1) Core block determination: Count the percentage of pixels in the encrypted sub-block that belong to the target object. When the percentage is >95%, it is determined to be a core block.

[0057] (5.2.2) Edge block determination; when the proportion of target pixels in the encrypted sub-block is in the range of [30%, 95%], it is determined to be an edge block;

[0058] (5.2.3) Background block determination: When the pixel ratio of the target object in the encrypted sub-block is less than 30%, it is determined to be a pure background block.

[0059] Preferably, the scene content edge encryption strategy used in step (5) includes: core blocks are encrypted according to the scene content desensitization leniency, edge blocks are encrypted with an offset of 0.029-0.032 above the corresponding core block baseline value, and background blocks are encrypted by randomly obtaining the L1 scene desensitization leniency. The hierarchical encryption achieves a balance between contour jaggedness suppression and edge encryption effect.

[0060] Compared with the prior art, the present invention has the following advantages:

[0061] (1) Multi-level scene image feature preservation encryption: This invention divides the service robot operation scene into five security levels from L1 to L5 and sets corresponding desensitization leniency ranges. The unique operation process parameters of the robot are selected as key parameters and a random key stream is generated. The thumbnail preservation encryption (TPE) algorithm is used to encrypt and decrypt the robot image features. The desensitization leniency is randomly selected to dynamically calculate the plaintext weights of different levels. Then, the image weighted fusion model is used to fuse the plaintext and ciphertext images to achieve multi-level scene image feature preservation encryption.

[0062] (2) Controllable image encryption detail processing effect: The present invention selects the size of the encryption sub-block according to the image resolution. Based on this, a three-segment differentiated weight model is proposed, which adopts the tenth power concave function model, the adaptive controllable Sigmoid model, and the bounded exponential convergence model. The L5 scene, L2~L4 scene and L1 scene are set respectively. The plaintext weight is dynamically adjusted by relying on the desensitization relaxation. The uncontrollable step parameter change of the original thumbnail preservation encryption (TPE) algorithm is converted into a continuous curve change, so as to realize the fine and controllable adjustment of the image encryption detail processing effect.

[0063] (3) Multi-level image feature preservation encryption for scene content: The service robot first identifies the inspection and testing scene it is in and determines the corresponding confidentiality level. Based on this scene, panoramic image segmentation and recognition classification algorithms are used to segment, identify and classify the content in the scene into different confidentiality levels. A dynamic recursive method is used to set a gradient desensitization leniency for different levels of confidentiality areas and a clear and visible desensitization leniency for background areas without confidentiality value. This achieves the differentiated effect of privacy encryption in key areas and clear visibility in background areas. Furthermore, the image feature preservation edge processing strategy is applied to optimize the irregular segmentation edges of the image, balancing the jagged defects caused by the segmentation of irregular contours with the edge encryption effect. This completes the image feature preservation encryption of multi-level scene content and improves the recognition and analysis of inspection and testing scenes. Attached Figure Description

[0064] Figure 1 This is a flowchart of the service robot image feature preservation and encryption method in the inspection and detection scenario of the present invention, as described in Embodiment 1 of the present invention.

[0065] Figure 2 This is the original image of the image features captured by the service robot in Embodiment 1 of the present invention;

[0066] Figure 3 This is a encrypted image with preserved image features captured by the service robot in Embodiment 1 of the present invention.

[0067] Figure 4 This is the plaintext image obtained after decryption and preservation of image features captured by the service robot in Embodiment 1 of the present invention;

[0068] Figure 5 This is a comparison curve of the plaintext weights calculated by the Sigmoid model with a controllable rate in the interval [0,1] of Embodiment 1 of this invention and the method of this invention, which varies with the desensitization leniency. In this graph, 5.1 is the L5 scenario, 5.2 is the L2-L4 scenario, 5.3 is the L1 scenario, 5.4 is the plaintext weight curve calculated by the Sigmoid model with a controllable rate in the interval [0,1] (blue line), and 5.5 is the plaintext weight curve calculated by this method (red line).

[0069] Figure 6 The L5 scene image features are preserved and encrypted in Embodiment 1 of this invention;

[0070] Figure 7 The L4 scene image features are preserved and encrypted in Embodiment 1 of this invention;

[0071] Figure 8 The L3 scene image features are preserved and encrypted in Embodiment 1 of this invention;

[0072] Figure 9 The L2 scene image features are preserved and encrypted in Embodiment 1 of this invention;

[0073] Figure 10 The L1 scene image features are preserved and encrypted in Embodiment 1 of this invention;

[0074] Figure 11 This is a schematic diagram of the inspection and detection scene segmentation output of a large number of fragmented masks in Embodiment 1 of the present invention;

[0075] Figure 12 yes Figure 11 A schematic diagram showing the fusion of features from images of the same category into a large region.

[0076] Figure 13 This is a schematic diagram of the mask for the irregular contour of the person category in panoramic image segmentation in Embodiment 1 of the present invention. The right image is the target mask for the person category with irregular contour, and the left image is a magnified schematic diagram of the target mask.

[0077] Figure 14 This is a schematic diagram of the edge processing effect of the character category in Embodiment 1 of the present invention. The right image shows the effect of using the scene content edge processing strategy to encrypt the target mask of the character category only for irregular outlines. The left image is an enlarged schematic diagram of the target mask encryption effect.

[0078] Figure 15This is an example diagram of multi-level image feature preservation encryption for scene content in Embodiment 1 of the present invention. 15.1 and 15.2 correspond to the walls of the background area, 15.3 corresponds to the category of people, 15.4 corresponds to the detection instrument and the detection sample in the sample, 15.5 corresponds to other items, 15.6 corresponds to the tables and chairs of the background area, 15.7 corresponds to the corresponding detection instrument and the detection instrument in the sample, and 15.8 corresponds to the floor of the background area.

[0079] above Figure 2 , Figure 4 , Figure 9 , Figure 10 , Figure 11 , Figure 12 and Figure 13 To protect privacy, the faces of the people in this image have been blurred. The people in this image are workers in the inspection and testing scenario, not any specific individual. Detailed Implementation

[0080] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0081] like Figure 1 As shown in this embodiment, the service robot image feature preservation and encryption method for inspection and testing scenarios includes the following steps:

[0082] (1) Classification of confidentiality levels for inspection and testing scenarios:

[0083] The service robot operation inspection and testing scenarios are divided into five confidentiality levels from L1 to L5, and corresponding hierarchical desensitization leniency value ranges are set. When the service robot enters the corresponding level scenario, the desensitization leniency within that range is randomly selected to avoid outputting the same fixed value, thus matching the randomness of real inspection scenarios.

[0084] Specifically, it includes:

[0085] (1.1) Classification of confidentiality levels for inspection and testing scenarios: Based on the openness of inspection and testing sites, personnel flow, business confidentiality level and experimental risk level, service robot inspection and testing scenarios are uniformly classified into five confidentiality levels from L1 to L5. From fully open public open areas, to office areas, functional laboratories, confidential offices and medium-risk laboratories, and then to high-confidential laboratories and core testing areas, the confidentiality control standards are progressively upgraded to cover all types of inspection and testing operation sites, so as to achieve standardized and refined classification of scenario confidentiality levels, as shown in Table 1 below.

[0086] (1.2) Desensitization leniency classification: Based on the five levels of confidentiality in the inspection and testing scenario, a hierarchical desensitization leniency value range is divided accordingly. Following the principle that the higher the confidentiality level, the higher the risk of information leakage, and the lower the desensitization leniency value, exclusive value ranges are matched for levels L1 to L5 in sequence. The security risk level of different scenarios is intuitively distinguished through quantitative data, providing a standardized numerical basis for the risk assessment of service robot inspections, as shown in Table 1.

[0087] Table 1. Classification of Confidentiality Levels for Service Robot Operation Inspection and Testing Scenarios

[0088] .

[0089] (1.3) Random Selection of Desensitization Looseness: When the service robot carries out inspection operations, it first identifies the current inspection and testing scenario, matches the corresponding confidentiality level, and locks a specific desensitization looseness range. To match the randomness of real inspection scenarios, the desensitization looseness value will be randomly selected within the corresponding range to avoid the limitations of fixed values. For the same scenario and the same image, with the help of the randomly selected desensitization looseness during the viewing stage, the plaintext-ciphertext fusion weight generated each time is within the range but different. The preview effect of the image is different each time it is opened, which truly simulates the dynamically changing security risk status under different scenarios. For example, the desensitization looseness randomly selected for 5 scenarios is (0.05, 0.21, 0.52, 0.69, 0.85).

[0090] (2) Random key generation:

[0091] Unique operational parameters of the service robot (service robot serial number, inspection 3D spatial coordinates, fuselage heading angle, and inspection timestamp) are selected as key parameters and normalized. A mapping-coupling iterative algorithm is used to generate an encrypted key stream, ensuring that the key stream generation result is random each time. Specifically, this includes:

[0092] (2.1) Collection of unique key parameters: Select the service robot equipment serial number, inspection three-dimensional space coordinates, fuselage heading angle, and inspection timestamp. Use the above-mentioned operation process parameters with unique characteristics as the original input parameters for key generation.

[0093] (2.2) Key parameter normalization: The four types of original key parameters with different dimensions and scales are normalized and standardized, and all features are uniformly constrained to the interval [0, 1) to eliminate the interference of heterogeneous parameter magnitude differences on the key generation algorithm iteration. Among them, the device serial number is normalized to N. id The formula is:

[0094] (1)

[0095] In the formula R id The table shows the robot's unique hardware serial number, and mod represents the modulo operation.

[0096] Cruise three-dimensional coordinates (H, X) robot Y robot Normalized N pos The formula is:

[0097] (2)

[0098] Where H is the floor number, X is the floor number. robot Y robot Let C be the horizontal and vertical coordinates of the robot's planar inspection. max H is the preset threshold for the maximum diagonal of the inspection map. max It is the threshold of the highest floor of the building. In order to fit the logic of prioritizing positioning and using the floor as a secondary factor, the weight of the plane coordinate is 0.6 and the weight of the detection height is 0.4.

[0099] Normalized heading angle N θ The formula is:

[0100] (3)

[0101] Where θ is the heading angle, and the unit is degrees (°).

[0102] Timestamp normalization: To ensure uniqueness, the date and millisecond combinations are normalized to N. t The formula is:

[0103] (4)

[0104] Where tx = 86400000 represents the total number of milliseconds in one natural day, and T is the millisecond-level timestamp of image acquisition. The formula includes two normalization terms: the first term is the date feature, which is obtained by dividing the timestamp T by the daily milliseconds tx and rounding down to get the cumulative number of days corresponding to the acquisition time. The cumulative number of days is then modulo 100000 and divided by 100000 to map to the 0~1 interval. The second term is the instantaneous acquisition feature of the current day in milliseconds. The timestamp T is modulo tx to get the millisecond offset that has elapsed on that day. This offset is then divided by the total number of milliseconds tx per day and mapped to the 0~1 interval. Finally, N t The average of two normalized features is used to distinguish time points within a single day at the 1ms level; the uniqueness of the timestamp normalized value is ensured by superimposing two layers of time features, guaranteeing the non-repetition and randomness of the key stream.

[0105] (2.3) Random Keystream Generation: The normalized unique key parameters are used as external perturbation inputs to the algorithm. A mapping-coupled iterative algorithm is adopted, and the formula is:

[0106] (5)

[0107] In the formula, n represents the number of iterations, x n-1 y n-1 Let x be the state value of the (n-1)th iteration. n y n , z n Let r1, r2, and r3 be the state values ​​of the nth iteration, and r1, r2, and r3 be the chaotic branch coefficients, with r1 = r2 = r3 = 3.99; w1 and w2 are the external disturbance weights, with w1 = w2 = 0.15; N id N θ N pos N t To normalize the key parameters, an external perturbation is introduced into the iterative system. To avoid iterative convergence failure and ensure that the chaotic sequence continues to have random traversal characteristics, the iterative output values ​​are constrained by an interval, eliminating extreme points that infinitely approach 0 and 1, and limiting the effective value range to [0.01, 0.99]. According to equation (5), continuous iterative iteration can produce a chaotic divergence effect, so that each generated key is unique and has randomness. The solution formula for the key stream seq[n] is:

[0108] (6)

[0109] In the formula, seq[n] is the nth output sequence value in the key stream. The formula sums the three iterative state values, amplifies them by 10000, and then takes the modulo of 1 to output the key stream elements with values ​​in the interval [0,1).

[0110] For example, set the initial values ​​x0=0.23, y0=0.71, select a set of key parameters with the same device serial number, cruise three-dimensional coordinates, and heading angle, and only a 1ms difference in timestamps (such as time 2026-07-29 01:05:45.123000 and time 2026-07-29 01:05:45.124000) to conduct a comparative experiment. The generated results are shown in Table 2. As can be seen from the results, the values ​​of seq[0], seq[1], and seq[2] are different from each other. After continuous loop iteration, a chaotic divergence effect is generated, and the subsequent output key stream sequences are all different. Even if the other inspection parameters (device serial number, cruise three-dimensional coordinates, and heading angle) are exactly the same, there is only a small 1ms difference in timestamps, and the final generated key results are completely different, ensuring that the key stream generation results are random and non-repetitive each time.

[0111] Table 2 Random Key Stream Generation

[0112]

[0113] (3) Robot image feature preservation encryption and decryption

[0114] To accommodate input images of different resolutions, the size of the encrypted sub-blocks is selected, and the Thumbnail Preservation Encryption (TPE) algorithm is used to encrypt the image, generating ciphertext for transmission and storage. In the decryption stage, the same key stream is used to decrypt the ciphertext and recover the original image.

[0115] (3.1) Selection of Encryption Sub-block Size: Considering that the size of the sub-block determines the visual effect of the encrypted image, the larger the side length of the sub-block, the more blurred the encryption details. In order to adapt to input images of different resolutions (P m P n When the side length of the selected sub-block is N, in the engineering implementation, considering memory alignment, parallel processing acceleration, and compatibility with existing image encoding and decoding standard macroblocks, the side length of the encrypted sub-block is preferably selected as an integer power of 2. Therefore, N can be selected as:

[0116] (7)

[0117] Where N is the side length of the encrypted sub-block, round(·) represents rounding to the nearest integer, and P m P n These refer to the width and height of the image, respectively.

[0118] (3.2) Robot Image Feature Preservation Encryption: The Thumbnail Preservation Encryption (TPE) algorithm is adopted. The image grid is divided into blocks, and the permutation sequence is generated iteratively and the permutation encryption sub-block pixels are rearranged. The output is a global block permutation encryption image for transmission and storage. While completing the image pixel encryption, the coarse-grained scene features of the image are preserved, such as... Figure 2 The original image is encrypted as follows: Figure 3 As shown, the encryption preserves the coarse-grained scene outline of the robot's inspection images, allowing for online preview without decryption. It can also quickly distinguish between the laboratory and inspection points based on prior visual knowledge, without exposing privacy within the images. The specific encryption steps are as follows:

[0119] Step 1: Image grid block traversal; slide the image row by row and column by column to generate sub-blocks by dividing it into N×N encrypted sub-blocks.

[0120] Step 2: Generate the permutation sequence; use the key stream to generate the permutation sequence.

[0121] Step 3: Sub-block pixel rearrangement permutation; use the permutation sequence to shuffle and rearrange the pixel array generated by the encrypted sub-block and restore the N×N matrix, and rewrite it back to the image.

[0122] Step 4: Keep the robot image features encrypted; continue to execute steps 2 and 3, traverse all encrypted sub-blocks in the image, and finally output the global block permutation encrypted image.

[0123] (3.3) Decryption while preserving robot image features; using the same keystream as encryption to solve for the inverse permutation index, cyclically permuting to recover sub-block pixels, completely decrypting and recovering the original image. For example... Figure 4 As shown, the original image can be completely decrypted and recovered. The specific encryption steps are as follows:

[0124] Step 1: Traverse the image grid in blocks; sequentially traverse all N×N encrypted sub-blocks.

[0125] Step 2: Solve for the inverse permutation index; use the same key stream as the encryption key to generate the permutation sequence and solve for the inverse permutation index.

[0126] Step 3: Pixel permutation and recovery of encrypted sub-blocks; pixels are permuted inversely to restore their original positions.

[0127] Step 4: Decrypt the robot image features; continue executing steps 2 and 3, traversing all complete encrypted sub-blocks within the image, and finally obtain the decrypted image.

[0128] (4) Encryption of multi-level scene image features:

[0129] Considering the requirements for gradient and hierarchical encryption strength in inspection and testing scenarios, and the issues that prioritizing integer powers of 2 in encryption sub-block engineering can lead to large fixed block intervals, rapid parameter jumps, and difficulty in accurately controlling image detail processing, a three-segment differentiated weight model (Equation 8) is proposed based on the five-level confidentiality requirements of inspection and testing scenarios. This model constructs a segmented hierarchical image feature preservation encryption system, employing a tenth-power concave function model, an adaptive controllable Sigmoid model, and a bounded exponential convergence model. Specifically, the tenth-power concave function model is used for L5 scenarios, the adaptive controllable Sigmoid model for L2-L4 scenarios, and the bounded exponential convergence model for L1 scenarios. By dynamically adjusting the plaintext weights at different levels based on the desensitization leniency, the uncontrollable step parameter changes of the original thumbnail preservation encryption (TPE) algorithm are transformed into continuous curve changes. The plaintext weight W is dynamically calculated through the desensitization leniency α. ori (α), as shown in formula (8), the corresponding curve is as follows Figure 5 The red line at 5.5 in the image is then used to implement near-zero plaintext weight encryption in L5 scenes, adaptive balancing in L2-L4 scenes, and bounded convergence in L1 scenes, as shown in the example. Figure 5 As shown, 5.1 is the L5 scenario, 5.2 is the L2-L4 scenario, and 5.3 is the L1 scenario.

[0130] (8)

[0131] in: W ori(α) represents the three-stage differentiated plaintext weights. The first stage uses the tenth-power concave function model for plaintext weight calculation in the L5 scenario; the second stage uses the adaptive controllable sigmoid model for plaintext weight calculation in the L2-L4 scenarios; and the third stage uses the bounded exponential convergence model for plaintext weight calculation in the L1 scenario. α represents the desensitization leniency. V Max V The desensitization leniency thresholds for L5, L2-L4, and L1 scenarios are set to 0.2, 0.8, and W, respectively. L W M Let α be the plaintext weight with threshold values ​​of 0.2 and 0.8 respectively, and k, λ, and Max be the weights. W These are the steepness coefficient, adaptive convexity coefficient, and global scaling coefficient in the second adaptive controllable Sigmoid model, which are [4.9, 5.1], [0.68, 0.72], and [0.6, 0.7], respectively. In this embodiment, they are taken as 5.0, 0.7, and 0.65, respectively. k2, Max END These are the steepness coefficient and the maximum convergence upper limit coefficient in the third bounded exponential convergence model, respectively [5.8, 6.1] and [0.75, 0.83]. In this embodiment, the values ​​are 6.0 and 0.8.

[0132] In this embodiment, the steepness coefficient k, adaptive convexity coefficient λ, and global scaling coefficient Max are used in the plaintext weight calculation formula of the adaptive controllable Sigmoid model for L2-L4 scenarios. W When W is set to 5, 0.7, and 0.65 respectively, L W M With threshold values ​​of 0.2 and 0.8 for α respectively, the plaintext weights obtained by substituting them into the plaintext weight calculation formula of the adaptive controllable Sigmoid model are 0.1319 and 0.5916 respectively.

[0133] Further includes:

[0134] (4.1) For L5 scenarios, plaintext feature information needs to be almost completely masked. A tenth power concave function model is adopted. When the desensitization leniency approaches 0, the plaintext weight approaches 0 synchronously. When it approaches the 0.2 threshold, it slowly increases to complete the near-zero weight encryption of plaintext in the inspection and detection scenario and maximize the image encryption strength.

[0135] (4.2) For L2, L3, and L4 scenarios, the L4 scenario requires enhanced encryption masking to suppress plaintext information leakage. The corresponding weight curve should rise gently and maintain a low plaintext weight overall. The L3 scenario requires consideration of both security performance and image detail clarity. The corresponding weight curve has the largest increase in weight, with a significant bulge and a balanced ratio of plaintext and ciphertext. The L2 scenario requires gradual restoration of image texture and feature details. The corresponding weight curve's rate of increase slows down significantly, with a gradual and slight increase in plaintext weight. The above requirements can be met by using the traditional Sigmoid model to complete the plaintext fusion weight allocation. The formula for calculating the plaintext weight using the traditional Sigmoid model is:

[0136] (9)

[0137] However, the traditional Sigmoid curve has a fixed shape and cannot adjust the degree of bulge in the middle section of the curve autonomously. It is difficult to flexibly adjust the curve's ascent rate for different security levels from L2 to L4, and its adaptability to different scenarios is insufficient. An adaptive and controllable Sigmoid model can be used to achieve flexible adjustment of the curve shape through parameter optimization. The specific implementation steps are as follows:

[0138] Step 1: Midpoint shift of the curve; Based on the traditional Sigmoid, a midpoint shift is introduced to solve the defect that the traditional Sigmoid curve is approximately flat in the [0,1] independent variable interval and is difficult to distinguish between different security levels. This makes the complete S-curve fall within the [0,1] range, achieving a smooth transition from low weight near α=0 to high weight near α=1, ensuring that the L2~L4 multi-level security intervals all have sufficient distinguishability.

[0139] Step 2: Weight increase rate control; a steepness coefficient k is introduced to adjust the weight increase rate in the interval, and the specific expression is shown in formula (10):

[0140] (10)

[0141] The rate of weight increase can be adjusted within different desensitization leniency ranges, such as Figure 5 As shown by the blue line in section 5.4, where k is set to 5.

[0142] Step 3: Adjust the mid-section's concave-convex shape freely; introduce an adaptive convexity coefficient λ to control the amplitude of the bulge in the middle of the curve, freely adjusting the mid-section's concave-convex shape, overcoming the problem that the traditional Sigmoid only has a fixed and monotonous S-shaped contour and the transition shape cannot be adjusted.

[0143] Step 4: Plaintext weight boundaries are controllable. A global scaling factor Max is introduced. W Limiting the maximum value of plaintext weights restricts the output plaintext weights to a specified range, thus making the weight boundaries controllable.

[0144] Compared to the traditional Sigmoid model, the adaptive and controllable Sigmoid model constructed in this paper has the ability to flexibly adjust multiple parameters and dimensions, and can accurately adapt to the differentiated desensitization and encryption requirements of L2-L4 multi-level security scenarios, significantly improving scenario adaptability and engineering practicality.

[0145] (4.3) For the L1 scenario, set the steepness coefficient k2 and the maximum convergence limit Max. END It adopts a bounded exponential convergence model to smoothly increase the plaintext weight and converge to the maximum convergence limit. Based on the ratio of high plaintext weight and low ciphertext weight, it restores the original details of the image to the greatest extent while retaining basic security protection.

[0146] (4.4) Multi-level scene image feature preservation encryption:

[0147] Based on the confidentiality level of the inspection and testing scenario in which the service robot is located, the desensitization leniency is randomly selected within the corresponding level range. The plaintext weights of different levels are calculated using a three-segment differentiated weight model. For example, the results of solving the plaintext and ciphertext weights for the five scenarios randomly selected in (1.3) above (0.05, 0.21, 0.52, 0.69, 0.85) are shown in Table 3. Then, the plaintext and ciphertext images are fused using an image weighted fusion model to achieve encryption while preserving the image features of the L1 to L5 multi-level scenarios.

[0148] The image weighted fusion model is shown in Equation (11):

[0149] (11)

[0150] Among them, I fusion (α) represents the fused image, α represents the desensitization leniency, and W ori (α) is the plaintext weight, I ori For plaintext images, I enc This is a encrypted image.

[0151] Table 3. Obtaining Plaintext and Ciphertext Weights Using the Three-Stage Differentiated Weighting Model

[0152]

[0153] As shown in Table 3, the plaintext and ciphertext weights obtained from the three-segment differentiated weighting model indicate that the plaintext weight in the L5 scene is almost zero, maximizing the image encryption strength, with the following effect: Figure 6 As shown, the L4 scenario maintains a low plaintext weight, enhancing the encryption and masking effect, as shown in the image. Figure 7 As shown, the L3 scene has a balanced ratio of plaintext to ciphertext, achieving a balance between security performance and image detail clarity, as demonstrated by... Figure 8 As shown, L2 scenes have a higher plaintext weight, progressively restoring image texture and feature details, with the following effect: Figure 9As shown, in L1 low-security scenarios, a high plaintext weight and low ciphertext weight ratio maximizes the restoration of original image details while retaining basic fallback security protection, achieving the following effect: Figure 10 As shown.

[0154] (5) Encryption of multi-level image features for scene content preservation:

[0155] (5.1) The panoramic image segmentation algorithm is used to segment and the recognition and classification algorithm is used to identify, classify and fuse the images of the inspection and testing scene:

[0156] A panoramic image segmentation algorithm is used to segment and inspect the target in the image, generating a target mask set. A recognition and classification algorithm is then used to extract and classify the target regions within the masks. Finally, fragmented masks of the same category are fused to form a complete region. Specifically, this includes:

[0157] (5.1.1) Panoramic Image Segmentation of the Inspection and Detection Scene; A panoramic image segmentation algorithm is adopted, using panoramic grid point sampling and hierarchical cropping inference mechanisms to segment the target in the inspection and detection scene image and generate a target mask set. The segmentation results are as follows: Figure 11 As shown, different colors represent different identification masks.

[0158] (5.1.2) Image feature recognition and classification: The CLIP recognition and classification algorithm is used to extract the image features of the target area within the mask, and the cosine similarity is calculated with the text features of each category (person category, detection instrument and sample, scene background, other items). The category corresponding to the maximum similarity is selected as the label content of the current fragment mask.

[0159] (5.1.3) Feature fusion of images of the same category; such as Figure 11 As shown, different colors represent masks for different recognition purposes. Image segmentation outputs a large number of fragmented masks. A bitwise OR fusion strategy is used to perform bitwise OR operations on spatially adjacent fragmented masks with the same label, merging them into connected complete regions, such as... Figure 12 As shown, the same color represents the same label content, and Figure 11 In contrast, fragment masks containing adjacent content with the same tag are merged into a relatively large area.

[0160] (5.2) Classification and grading of scene content:

[0161] The fusion results of image features of the same category are classified into scene content, which specifically includes classified areas (personal categories, detection instruments and samples, other items) and background areas (scene background), and are classified according to the classified risk level of the scene content.

[0162] (5.2.1) Personnel Category: refers to the personnel working in the inspection and testing scenario. This category involves core confidential personnel information in the scenario and has the highest encryption level.

[0163] (5.2.2) Testing instruments and samples: covering specialized testing instruments in the scene and samples to be tested, with high encryption requirements.

[0164] (5.2.3) Other items: Unknown items that are not accurately identified by the algorithm or whose semantic features are ambiguous are uniformly classified into this category to avoid the risk of information leakage caused by unknown items. The encryption requirements are high.

[0165] (5.2.4) Scene background: includes infrastructure such as walls, floors, tables and chairs in the scene. Such targets have significant recognition features and high resolution, which can effectively assist in scene localization and scene attribute judgment, and there is no risk of confidentiality and low encryption requirements.

[0166] (5.3) Image feature-preserving edge segmentation:

[0167] The target mask output by panoramic segmentation has an irregular contour, while the Thumbnail Preservation Encryption (TPE) algorithm requires N×N blocks as the smallest encryption unit. The encryption level of the confidential area is high. If the edge blocks are encrypted as a whole, obvious jagged edges will be produced at the target contour. If edge block encryption is abandoned, although the contour is visually smooth, it will cause the confidential pixels at the target edge to be missed, resulting in incomplete encryption coverage. The block type is dynamically determined by the mask pixel coverage ratio. The encrypted sub-blocks of the confidential area are divided into three categories: core blocks, edge blocks, and background blocks. This facilitates the hierarchical configuration of the desensitization leniency parameters in the later stage, and achieves a balance between contour jaggedness suppression and edge encryption effect.

[0168] (5.3.1) Core block determination: Statistically count the percentage of pixels in the encrypted sub-block that belong to the target object. When the percentage is >95%, it is determined to be a core block.

[0169] (5.3.2) Edge block determination: When the proportion of target pixels in the encrypted sub-block is in the range of [30%, 95%], it is determined to be an edge block.

[0170] (5.3.3) Background block determination: When the pixel ratio of the target object in the encrypted sub-block is <30%, it is determined to be a pure background block.

[0171] (5.4) Multi-level image feature preservation encryption oriented towards scene content:

[0172] This method employs a dynamic recursive approach to hierarchically encrypt the content of sensitive areas within the scene. It randomly selects L1 scene desensitization and loosely encrypted background regions, and uses scene content edge processing strategies to encrypt the scene content edges, thus achieving multi-level image feature preservation encryption oriented towards scene content. Specifically, this includes:

[0173] (5.4.1) Scene Content Hierarchical Encryption: To ensure the distinguishability between scene content levels in classified areas, a minimum spacing constraint Δ is set. A dynamic recursive method is adopted, which dynamically adjusts the upper and lower limits to transform the global constraint (L, R] into a local secure sampling interval (Min). i Max i A random desensitization leniency sequence is generated as U1, U2, ..., UK (K is the number of levels), ensuring that there is a minimum gap Ui-Ui-1>Δ between desensitization leniency levels. The background region randomly obtains the L1 scene desensitization leniency to improve scene discriminability. The specific process of the dynamic recursive method is as follows:

[0174] Step 1: Set the safe sampling interval of UI to (Min) i Max i ];

[0175] Step 2: Solve the dynamic lower limit formula; set the boundary condition Min1=L, and the recursive formula is Min i =Ui-1+Δ+ϵ, where ϵ is the floating-point safety margin, which is set to 0.000001 to prevent constraint failure caused by floating-point precision errors and enhance robustness.

[0176] Step 3: Solve the dynamic upper bound formula; to ensure sufficient space for subsequent Ki numbers and satisfy the Δ constraint, the calculation formula is Max. i =R-(Ki)×(Δ+ϵ);

[0177] Step 4: Randomly generate desensitization leniency; randomly generate desensitization leniency within the safe sampling interval.

[0178] Taking the scene content of the confidential area in scene L4 as an example, with interval L=0.2, R=0.4, and the minimum spacing constraint set as Δ=0.05, ϵ=0.000001, the character category, detection instrument and sample, and other items (number of layers K=3) correspond to U1, U2, and U3 respectively, as shown in the following specific generation:

[0179] (1) Random generation of U1: Set the lower limit to Min1=L=0.2 and solve the upper limit to Max1=0.4-(3-1)×(0.05+0.000001)=0.299998. Randomly generate U1 between (0.2, 0.299998].

[0180] (2) Random generation of U2: Set the lower limit Min2=U1+0.05+0.000001 and the upper limit Max2=0.4-(3-2)×0.050001=0.349999. Randomly generate U2 in (U1+0.050001, 0.349999].

[0181] (3) Random generation of U3: Set the lower limit Min3=U2+0.05+0.000001 and the upper limit Max3=0.4-(3-3)×0.050001=0.4. Randomly generate U3 in (U2+0.050001, 0.4].

[0182] Following the above process, the randomly generated U1, U2, and U3 in the L4 scene are 0.239, 0.302, and 0.390, respectively. The desensitization leniency selection ranges for other inspection and detection scenes are shown in Table 4. The minimum spacing constraint between levels is Δ=0.05, and the floating-point safety margin is ϵ=0.000001. The leniency values ​​for each level satisfy the constraint condition that the leniency of the subsequent level - the leniency of the preceding level > Δ, ensuring the scene level differentiation. The L1 scene already has high recognition accuracy and has restored the original image details to the greatest extent possible; the scene background desensitization leniency is set to the maximum value of 1.

[0183] Table 4: Range of Flexibility for Encryption and Desensitization at the Scene Content Level

[0184]

[0185] Note: In the table, U1 represents the desensitization leniency for the category of people, U2 represents the desensitization leniency for the testing instruments and samples, and U3 represents the desensitization leniency for other items.

[0186] (5.4.2) Setting the scene content edge processing strategy:

[0187] The core block is encrypted according to the scene content's desensitization leniency. The desensitization leniency of the edge blocks is encrypted with an offset of 0.029-0.032 above the corresponding core block's base value. If the offset value exceeds the upper limit of the scene's range, the upper limit is used directly. Taking U1, U2, and U3 randomly generated in the L4 scene as an example, with an offset of 0.03, the desensitization leniency of the U1 core block is 0.239, and the edge block value is 0.269; the desensitization leniency of the U3 core block is 0.390, and the value is 0.42, exceeding the upper limit of the scene's range, so the edge block value is 0.40. The background block is encrypted using a randomly obtained L1 scene desensitization leniency. Taking the L4 scene's character category as an example, the final effect is as follows. Figure 13-14 As shown, Figure 13 The image shown is an image with an irregular outline as the target mask for the human category output from panoramic segmentation. The right image is the target mask for the human category with an irregular outline, and the left image is a magnified schematic diagram of this target mask. Figure 14 The image shows the encryption effect of the scene content edge processing strategy. The right image shows the encryption effect of the scene content edge processing strategy on the target mask of the character category only with irregular outlines. The left image is a magnified schematic diagram of the target mask encryption effect. The hierarchical encryption suppresses the jagged outlines and also ensures the edge encryption effect.

[0188] (5.4.3) Multi-level image feature preservation encryption based on scene content: Taking L4 scene as an example, based on the scene content (person category, detection instrument and sample, other items, scene background) hierarchical encryption and scene content edge processing strategies, and combining the above hierarchical encryption and edge processing strategies, multi-level image feature preservation encryption based on scene content is completed, with the effect as follows: Figure 15 As shown. Figure 15 In section 15.3, the corresponding person category; 15.7, the corresponding detection instrument; 15.4, the corresponding detection sample; and 15.5, the corresponding other items (objects not recognized by the algorithm), the encryption masking effect decreases sequentially to ensure the distinction between content levels in the confidential area scene; through and Figure 7 L4 scene image features are preserved and encrypted. Figure 10 The L1 scene image features are preserved through encryption. Comparison shows that the privacy features of key areas—the person category (15.3), the detection instruments and samples (15.7 and 15.4 respectively), and other items (15.5)—are effectively encrypted, resulting in high security. Meanwhile, the background areas (15.1, 15.2, 15.6, and 15.8) retain basic fallback protection, remaining clearly visible and highly readable. The algorithm achieves encryption and confidentiality of key areas while maintaining clear visibility of background areas through differentiated encryption strategies, enabling legitimate users with prior scene knowledge to perform accurate image reasoning and recognition analysis.

[0189] The above embodiments are only used to illustrate the present invention, and the scope of protection of the present invention is not limited to the above embodiments. Those skilled in the art can achieve the purpose of the present invention based on the above disclosure. Any improvements and modifications made based on the concept of the present invention fall within the scope of protection of the present invention, and the specific scope of protection is determined by the claims.

Claims

1. A method for preserving and encrypting image features of a service robot in an inspection and detection scenario, characterized in that, Includes the following steps: (1) Classification of confidentiality levels for inspection and testing scenarios: Service robot operation inspection and testing scenarios are divided into five confidentiality levels from L1 to L5, and corresponding desensitization relaxation value ranges are set for each level. (2) Random key generation: Select unique operation process parameters of the robot as key parameters and generate a random key stream; (3) Robot image feature preservation encryption and decryption: The Thumbnail Preservation Encryption (TPE) algorithm is used to encrypt the features of images captured by the service robot, generating ciphertext images for transmission and storage. The original image, i.e., the plaintext image, is then decrypted using the same key stream used for encryption. (4) Encryption of multi-level scene image features: Based on the confidentiality level of the inspection and testing scenario in which the service robot is located, the desensitization leniency is determined within the corresponding level range. A three-stage differentiated plaintext weight model is adopted, and the plaintext weights of different levels are dynamically calculated based on the desensitization leniency. Then, an image weighted fusion model is used to fuse plaintext and ciphertext images to achieve encryption while maintaining the image features of multi-level scenarios. The three-stage differentiated plaintext weight model includes a tenth power concave function model, an adaptive controllable Sigmoid model, and a bounded exponential convergence model. The L5 scenario uses the tenth power concave function model, the L2-L4 scenarios use the adaptive controllable Sigmoid model, and the L1 scenario uses the bounded exponential convergence model. (5) Encryption of multi-level image features for scene content preservation: The service robot first identifies the inspection and testing scenario it is in and determines the corresponding confidentiality level. Based on the confidentiality level scenario, it uses a panoramic image segmentation algorithm to segment and a recognition and classification algorithm to identify, classify and fuse the scenario content, divide the scenario content into confidential areas and background areas, divide it into different levels according to the confidentiality risk level of the scenario content, and then perform image feature preservation edge division. A dynamic recursive method is used to perform hierarchical encryption of classified areas, setting a low desensitization leniency for classified areas and a high desensitization leniency for background areas; A scene content edge processing strategy is adopted to encrypt the scene content edges, thereby achieving multi-level image feature preservation encryption oriented towards scene content.

2. The service robot image feature preservation and encryption method for inspection and testing scenarios according to claim 1, characterized in that, In step (1), the desensitization leniency range for scenario L1 is (0.8, 1], for scenario L2 it is (0.6, 0.8], for scenario L3 it is (0.4, 0.6], for scenario L4 it is (0.2, 0.4], and for scenario L5 it is (0, 0.2]. Among them, scenario L1 is a public open area, scenario L2 is an office area, scenario L3 is a functional laboratory, scenario L4 is a classified office and a medium-risk laboratory, and scenario L5 is a high-secrecy laboratory and a core testing area.

3. The service robot image feature preservation and encryption method for inspection and testing scenarios according to claim 1, characterized in that, Step (2) involves selecting unique robot operation process parameters as key parameters and generating a random key stream, including: (2.1) Unique key parameter acquisition: Select the service robot equipment serial number, inspection three-dimensional space coordinates, fuselage heading angle and inspection timestamp as unique operation process parameters of the service robot, and use the unique operation process parameters as the original input parameters for key generation; (2.2) Key parameter normalization: The four types of original key parameters, namely service robot equipment serial number, inspection three-dimensional space coordinates, fuselage heading angle and inspection timestamp, are uniformly constrained to the interval [0, 1) and normalized to eliminate the interference of heterogeneous parameter magnitude differences on the key generation algorithm iteration; (2.3) Random key stream generation: The normalized unique key parameters are used as the external perturbation input of the algorithm. The mapping-coupled iterative algorithm is adopted. Relying on the chaotic divergence effect generated by the algorithm, a random key stream with non-repeating results in each iteration is generated through multiple rounds of iterative calculation.

4. The service robot image feature preservation and encryption method for inspection and testing scenarios according to claim 1, characterized in that, In step (3), the Thumbnail Preservation Encryption (TPE) algorithm is used to encrypt the image features captured by the service robot, generating a ciphertext image for transmission and storage. During the viewing phase, the original image, i.e., the plaintext image, is recovered using the same key stream as the encryption. Specifically, this includes: (3.1) Selection of encrypted sub-block size: The side length N of the encrypted sub-block is selected as: （1） Where N is the side length of the encrypted sub-block, round(·) means rounding to the nearest integer, and Pm and Pn refer to the width and height of the image, respectively; (3.2) The Thumbnail Preservation Encryption TPE algorithm is adopted to divide the image grid into blocks. Using the key stream, the permutation sequence is generated and the permutation encryption sub-block pixels are rearranged in a loop. The full-domain block permutation encryption image is output for transmission and storage. While completing the image pixel encryption, the coarse-grained scene features of the image are preserved. (3.3) Decryption while preserving robot image features: Solve the inverse permutation index using the same key stream as encryption, cyclically permutate to recover sub-block pixels, fully decrypt and recover the original image.

5. The service robot image feature preservation and encryption method for inspection and testing scenarios according to claim 1, characterized in that, The three-stage differentiated plaintext weight model described in step (4) is shown in formula (8): （8）； Among them: W ori (α) represents the three-stage differentiated plaintext weights. The first stage uses the tenth-power concave function model for plaintext weight calculation in the L5 scenario; the second stage uses the adaptive controllable sigmoid model for plaintext weight calculation in the L2-L4 scenarios; and the third stage uses the bounded exponential convergence model for plaintext weight calculation in the L1 scenario. α represents the desensitization leniency. V Max V The desensitization leniency thresholds for L5, L2-L4, and L1 scenarios are set to 0.2, 0.8, and W, respectively. L W M Let α be the plaintext weight with threshold values ​​of 0.2 and 0.8 respectively, and k, λ, and Max be the weights. W These are the steepness coefficient, adaptive convexity coefficient, and global scaling coefficient in the second adaptive controllable Sigmoid model, which are [4.9, 5.1], [0.68, 0.72], and [0.6, 0.7], respectively; k2, Max END These are the steepness coefficient and the maximum convergence upper limit coefficient in the third bounded exponential convergence model, respectively [5.8, 6.1] and [0.75, 0.83].

6. The service robot image feature preservation and encryption method for inspection and testing scenarios according to claim 5, characterized in that, Step (4) of preserving and encrypting the multi-level scene image features specifically includes: (4.1) For the L5 scenario, a tenth power concave function model is used to obtain the plaintext weights and maximize the image encryption strength; (4.2) For L2-L4 scenarios, an adaptive controllable Sigmoid model is used to obtain plaintext weights. Based on the Sigmoid model, a steepness coefficient k is introduced to adjust the rate of increase of weights in intervals, an adaptive convexity coefficient λ is introduced to control the bulge in the middle of the curve, and a global scaling coefficient Max is introduced. W Constrain the range of plaintext weights to achieve controllable plaintext and ciphertext weights; (4.3) For the L1 scenario, set the steepness coefficient k2 and the maximum convergence upper limit coefficient Max. END The plaintext weights are obtained by using a bounded exponential convergence model, which allows the plaintext weights to rise steadily and converge to the maximum convergence limit. (4.4) Multi-level scene image feature preservation encryption: Randomly select the desensitization leniency in the corresponding interval of different level scenes, use a three-segment differentiated plaintext weight model to calculate the plaintext weight of different levels, and then use an image weighted fusion model to fuse plaintext images and ciphertext images to achieve multi-level scene image feature preservation encryption from L1 to L5.

7. The service robot image feature preservation and encryption method for inspection and testing scenarios according to claim 1, characterized in that, Step (5) employs a panoramic image segmentation algorithm for segmentation and a recognition and classification algorithm for identification, classification, and fusion. This includes using a panoramic image segmentation algorithm to segment and detect scene image targets, generating a target mask set, using a recognition and classification algorithm to extract target regions within the mask and identify and classify them, and fusing image features of the same category into one region. Specifically, this includes: (5.1.1) Inspection and detection scene image segmentation: The panoramic image segmentation algorithm is adopted. Through panoramic grid point sampling and hierarchical cropping inference mechanism, the target in the inspection and detection scene image is segmented in the panoramic view to generate a target mask set; (5.1.2) Image feature recognition and classification: The image features of the target area within the mask are extracted using a recognition and classification algorithm. The cosine similarity is calculated between the image features and the text features of each category of scene content. The category corresponding to the maximum similarity is selected as the label content of the current fragment mask. (5.1.3) Fusion of features of the same category: For the large number of fragmented masks output by image segmentation, a fusion strategy of bitwise OR of the masks is used to merge the fragments with adjacent identical label content into a complete large region.

8. The method for preserving and encrypting image features of service robots in inspection and testing scenarios according to claim 1, characterized in that, The classified area mentioned in step (5) includes personnel categories, testing instruments and samples, and other items. The background area includes scene background, which is divided into different levels according to the classified risk level of the scene content, from high to low as follows: (5.2.1) Person categories, (5.2.2) Testing instruments and samples, (5.2.3) Other items, (5.2.4) Scene background.

9. The service robot image feature preservation and encryption method for inspection and testing scenarios according to claim 8, characterized in that, In step (5), a dynamic recursive method is used to perform hierarchical encryption of the classified areas. A low desensitization leniency is set for the classified areas, and a high desensitization leniency is set for the background areas. The selection range of the desensitization leniency for the scene content hierarchical encryption is as follows: For scenario L1, the desensitization leniency for the person category U1 is (0.8, 0.899998], the desensitization leniency for the detection instrument and sample U2 is (U1+0.050001, 0.949999], the desensitization leniency for other items U3 is (U2+0.050001, 1], and the desensitization leniency for the scene background is 1; For scenario L2, the desensitization leniency for person category U1 is (0.6, 0.699998], the desensitization leniency for detection instrument and sample U2 is (U1+0.050001, 0.749999], the desensitization leniency for other items U3 is (U2+0.050001, 0.8], and the desensitization leniency for scenario background is (0.8, 1.0]. For scenario L3, the desensitization leniency for person category U1 is (0.4, 0.499998], the desensitization leniency for detection instrument and sample U2 is (U1+0.050001, 0.549999], the desensitization leniency for other items U3 is (U2+0.050001, 0.6], and the desensitization leniency for scenario background is (0.8, 1.0]; For scenario L4, the desensitization leniency for person category U1 is (0.2, 0.299998], the desensitization leniency for detection instrument and sample U2 is (U1+0.050001, 0.349999], the desensitization leniency for other items U3 is (U2+0.050001, 0.4], and the desensitization leniency for scenario background is (0.8, 1.0]; For scenario L5, the desensitization leniency for the character category U1 is (0, 0.099998), the desensitization leniency for the detection instrument and sample U2 is (U1+0.050001, 0.149999), the desensitization leniency for other items U3 is (U2+0.050001, 0.2), and the desensitization leniency for the scene background is (0.8, 1.0).

10. The service robot image feature preservation and encryption method for inspection and testing scenarios according to claim 1, characterized in that, The image feature preservation edge segmentation in step (5) includes: dynamically determining the block type using the mask pixel coverage ratio, dividing the encrypted sub-blocks of the sensitive area into three categories: core block, edge block, and background block, which facilitates the subsequent hierarchical configuration of desensitization leniency parameters and achieves a balance between contour jaggedness suppression and edge encryption effect, wherein: (5.3.1) Core block determination: Calculate the percentage of pixels in the encrypted sub-block that belong to the target object. When the percentage is >95%, it is determined to be a core block. (5.3.2) Edge block determination: When the proportion of target pixels in the encrypted sub-block is in the range of [30%, 95%], it is determined to be an edge block; (5.3.3) Background block determination: When the pixel ratio of the target object in the encrypted sub-block is <30%, it is determined to be a pure background block; The scene content edge encryption strategy used in step (5) includes: core blocks are encrypted according to the scene content desensitization leniency, edge blocks are encrypted with an offset of 0.029-0.032 above the corresponding core block baseline value, and background blocks are encrypted by randomly obtaining the L1 scene desensitization leniency. The hierarchical encryption achieves a balance between contour jaggedness suppression and edge encryption effect.

Images