Image desensitization method and system based on multi-modal large model

Abstract

The invention provides an image desensitization method and system based on a multi-modal large model, and relates to the technical field of text generation images, and the method comprises the steps: carrying out the preprocessing of an original image, and obtaining a preprocessed image; performing visual reasoning on the preprocessed image by using a preset multi-modal large model to obtain scene understanding information and an image type; performing sensitive information detection on the preprocessed image according to the image type by using the multi-modal large model to obtain a sensitive information detection result; performing risk assessment on the sensitive information detection result by using the scene understanding information to obtain a risk score, and determining a desensitization decision by using the risk score; and finally, desensitizing the preprocessed image by using the desensitization decision and the sensitive information detection result to generate a desensitized image. According to the method, the image understanding ability is enhanced, the limitation of a fixed target is broken through, accurate image desensitization is realized, and the desensitization intelligence is improved.

Description

Technical Field

[0001] This application relates to the field of text-to-image technology, and in particular to an image desensitization method and system based on a multimodal large model. Background Technology

[0002] With the rapid development of artificial intelligence technology and the improvement of digitalization in traffic management, the quantity and complexity of image data generated in traffic accident investigations have increased significantly. This image data includes not only real-world photos of the accident scene, such as road conditions, vehicle damage, and injury records, but also document photos, such as digital documents like identification badges, medical certificates, insurance policies, and written records, as well as various types of materials including surveillance screenshots, vehicle identification information, and evidence provided by third parties. These image materials contain a large amount of sensitive information, such as faces, license plates, company logos, vehicle numbers, names, company information, signatures, and barcodes, posing a significant technical challenge to image anonymization processing.

[0003] Currently, traditional image desensitization techniques mainly include rule-based detection methods, traditional machine learning methods, deep learning object detection methods, and OCR text recognition methods. Rule-based methods rely on features such as color, shape, and texture for detection, which is simple and efficient, but has low recognition accuracy in complex scenes and poor adaptability to diverse information types. Traditional machine learning methods use handcrafted features and classifiers for recognition, significantly improving accuracy, but feature design relies on specialized knowledge, and generalization ability remains limited. Deep learning object detection methods can automatically learn features and have high recognition accuracy, but still require separate model training for different types of sensitive information and lack semantic understanding capabilities, making it impossible to determine whether information needs desensitization. While OCR technology can extract text information, its recognition accuracy in complex backgrounds is low, its localization accuracy is insufficient, and it cannot understand the semantics of text content, making it difficult to make intelligent desensitization judgments.

[0004] Therefore, existing technologies have significant limitations when processing multimodal and multi-type sensitive information: First, they have weak single-modal processing capabilities and are difficult to integrate visual and textual information; second, they lack contextual semantic understanding capabilities and are difficult to make intelligent desensitization decisions; third, fixed target detection limits the flexibility and scalability of the method; and fourth, they have low levels of intelligence and are difficult to cope with complex and ever-changing real-world application scenarios. Summary of the Invention

[0005] Therefore, it is necessary to address the problems of insufficient detection accuracy, lack of understanding ability, limitations in fixed target detection, and low level of intelligence in existing image desensitization technologies. This paper proposes an image desensitization method and system based on a multimodal large model to achieve high precision in image desensitization, while enhancing understanding ability, overcoming the limitations of fixed target detection, and improving the intelligence of image desensitization.

[0006] To achieve the above-mentioned technical effects, the technical solution of the present invention is as follows: An image desensitization method based on a multimodal large model includes: S1. Preprocess the original image to obtain the preprocessed image; S2. Use a preset multimodal large model to perform visual reasoning on the preprocessed image to obtain scene understanding information and image type; S3. Using the multimodal large model, perform sensitive information detection on the preprocessed image according to the image type to obtain sensitive information detection results; S4. Use the scenario understanding information to perform a risk assessment on the sensitive information detection results, obtain a risk score, and use the risk score to determine a desensitization decision; S5. Using the desensitization decision and the sensitive information detection results, the preprocessed image is desensitized to generate a desensitized image.

[0007] Preferably, the preprocessing of the original image to obtain the preprocessed image includes: S11. Perform multimodal feature extraction on the original image to obtain a multidimensional feature vector; S12. Input the multidimensional feature vector into a preset intelligent decision-making model, and have the intelligent decision-making model output a decision vector; S13. The original image is enhanced using the decision vector to obtain the preprocessed image.

[0008] Preferably, the preset intelligent decision-making model includes an input layer, a hidden layer, and a dual-branch output structure layer connected in sequence. The dual-branch output structure layer consists of a policy classification branch output layer and a parameter regression branch output layer. The input layer receives the multidimensional feature vector and transmits it to the hidden layer. The hidden layer maps and compresses the multidimensional feature vector to obtain shared features. The policy classification branch output layer performs policy classification processing on the shared features to obtain a preprocessed policy. The parameter regression branch output layer performs parameter regression processing on the shared features to obtain preprocessed parameters.

[0009] Preferably, the multimodal large model includes a visual reasoning module and a desensitization module connected in sequence. The pre-defined multimodal large model is used to perform visual reasoning on the preprocessed image to obtain scene understanding information and image type, including: S21. Use the visual reasoning module in the multimodal large model to perform scene analysis on the preprocessed image to obtain scene understanding information; S22. Calculate the correlation between the preprocessed image and the preset sensitive content to obtain a content correlation score; S23. Perform Bayesian fusion on the scene understanding information and the content relevance score to obtain the image type.

[0010] Preferably, the multimodal large model includes a visual reasoning module and a desensitization module connected in sequence. The step of using the multimodal large model to detect sensitive information in the preprocessed image based on the image type to obtain a sensitive information detection result includes: S31. Construct a detection prompt template corresponding to the image type; S32. Using the detection prompt template, guide the desensitization module in the multimodal large model to perform coarse-grained sensitive information detection on the preprocessed image to obtain coarse-grained detection results; S33. Input the coarse-grained detection results into a preset lightweight model to refine the bounding box and obtain a refined bounding box; S34. The coarse-grained detection results and the refined bounding box are fused to obtain the sensitive information detection results.

[0011] Preferably, the step of using the scene understanding information and the sensitive information detection results to perform a risk assessment on the preprocessed image to obtain a risk score includes: S41. Based on the sensitive information detection results, calculate the information type risk respectively. Exposure level risk Identifiability risk and contextual risks ; Risk of the information type The calculation formula is as follows:

[0012] in, The type weight is the type weight corresponding to the i-th type of sensitive information. Score the sensitivity of the i-th type of sensitive information; The level of exposure risk The calculation formula is as follows:

[0013] in, Score the visual salience of sensitive information. Sensitivity rating for sensitive information; The identifiable risks The calculation formula is as follows:

[0014] in, Score the resolution quality of areas containing sensitive information. Score the clarity of areas containing sensitive information. Score the integrity of sensitive information; The context risk The calculation formula is as follows:

[0015] in, Score the importance of scenarios involving sensitive information. Rate the use cases of sensitive information. Score the legal compliance of sensitive information; S42. Risk of Information Type Exposure level risk Identifiability risk and contextual risks Weighted fusion yields a risk score. ; The risk score The calculation formula is as follows:

[0016] Where α is the weighting coefficient for information type risk, β is the weighting coefficient for exposure level risk, γ is the weighting coefficient for identifiability risk, and δ is the weighting coefficient for context risk.

[0017] Preferably, determining the desensitization decision using the risk score includes: S43. Map the risk score to the corresponding risk score interval, and set different sets of candidate desensitization strategies for different risk score intervals; S44. Perform an applicability evaluation on the candidate desensitization strategies in the candidate desensitization strategy set, and obtain an applicability score for each candidate desensitization strategy; S45. The applicability score is used to evaluate the utility of each candidate desensitization strategy to obtain the utility function of each candidate desensitization strategy, and the desensitization strategy with the largest utility function is selected as the final desensitization decision.

[0018] Preferably, the final desensitization decision for:

[0019] in, Risk scoring for sensitive information Sensitive information type The location information of sensitive information in the image. For the first j The utility function corresponding to each candidate desensitization strategy; The first jThe utility function corresponding to each candidate desensitization strategy The formula for calculating is as follows:

[0020] in, For risk control weights, As a weight for visual naturalness, As a weight for visual naturalness, For security weights, For the first j Applicability score of each candidate desensitization strategy; The first j Applicability score of each candidate desensitization strategy The calculation formula is as follows:

[0021] in, The expected score for type matching. For risk adaptability expectation score, For contextual consistency This represents the expected quality score.

[0022] The sensitive information detection result includes a sensitive information bounding box. The step of using the desensitization decision and the sensitive information detection result to desensitize the preprocessed image and generate a desensitized image includes: S51. Adaptively expand the sensitive information bounding box to obtain an expanded bounding box; S52. Based on the desensitization decision, generate a mask on the extended bounding box to obtain a mask box; S53. The mask frame and the preprocessed image are fused to obtain a fused image.

[0023] This invention also proposes an image desensitization system based on a multimodal large model, comprising: The image preprocessing module is used to preprocess the original image to obtain the preprocessed image; The image classification module is used to perform visual reasoning on the preprocessed image using a preset multimodal large model to obtain scene understanding information and image type; The sensitive information detection module is used to perform sensitive information detection on the preprocessed image based on the image type using the multimodal large model, and obtain the sensitive information detection result; The risk scoring module is used to assess the risk of the sensitive information detection results using the scenario understanding information, obtain a risk score, and use the risk score to determine the desensitization decision. The image desensitization module is used to desensitize the preprocessed image using the desensitization decision and the sensitive information detection result, and generate a desensitized image.

[0024] Compared with the prior art, the beneficial effects of the technical solution of the present invention are: This invention proposes an image desensitization method and system based on a multimodal large model. First, the original image is preprocessed to obtain a preprocessed image, providing a stable and reliable input foundation for subsequent analysis. Then, a pre-defined multimodal large model is used to perform visual reasoning on the preprocessed image, obtaining scene understanding information and image type, thereby enhancing the global understanding of the image content. Next, the multimodal large model is used to detect sensitive information in the preprocessed image based on the image type, obtaining sensitive information detection results. This allows the detection process to adjust the detection focus for different image types, overcoming the limitation of traditional target detection which can only handle preset targets, and achieving more generalized sensitive information recognition. Then, the scene understanding information is used to perform risk assessment on the sensitive information detection results, obtaining a risk score, and the risk score is used to determine the desensitization decision, upgrading desensitization from simple rule triggering to risk-based intelligent decision-making. Finally, the desensitization decision and the sensitive information detection results are used to desensitize the preprocessed image, generating a desensitized image. Ultimately, an intelligent image desensitization method with high-precision detection capabilities, strong image understanding capabilities, and greater generalization ability is obtained. Attached Figure Description

[0025] Figure 1 This is a flowchart of the image desensitization method based on a multimodal large model in an embodiment of the present invention; Figure 2 This is a diagram illustrating the overall architecture of the image desensitization system based on a multimodal large model in this invention. Figure 3 This is a structural block diagram of the image desensitization system based on a multimodal large model in an embodiment of the present invention. Detailed Implementation

[0026] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent. It is understandable that some well-known content in the drawings may be omitted for those skilled in the art. In order to make the purpose, technical solution and advantages of this application clearer, the following describes this application in further detail with reference to the accompanying drawings and embodiments.

[0027] Example 1 like Figure 1 and Figure 2 As shown, this embodiment proposes an image desensitization method based on a multimodal large model, including: S1. Preprocess the original image to obtain the preprocessed image; In step S1, the system receives the raw image, supporting common formats such as RGB, PNG, and JPEG, with arbitrary image size. The system first performs input validation, checking image integrity, format compatibility, and size appropriateness. Then, it initializes the processing context, including processing parameters, intermediate results, and status information. Initialization also includes memory allocation, algorithm module loading, and parallel processing environment configuration. The subsequent processing flow consists of three sub-stages: multimodal feature extraction → intelligent decision classification → specialized preprocessing execution.

[0028] S11. Perform multimodal feature extraction on the original image to obtain a multidimensional feature vector; In S11, the feature extraction sub-stage generates a 64-dimensional feature vector. This vector is passed to the decision classification sub-stage. The decision classification sub-stage outputs a five-dimensional decision vector. It includes a strategy type and three processing parameters. The preprocessing execution sub-stage selects the appropriate specialized algorithm based on the decision vector to process the original image and output the enhanced image.

[0029] Specifically, extract 64-dimensional multidimensional feature vectors. It contains seven sub-feature vectors: texture features (8-dimensional) edge features (8-dimensional) color characteristics (12-dimensional) Clarity Features (8-dimensional) contrast characteristics (8-dimensional) noise characteristics (8-dimensional) structural features (12-dimensional). Each feature vector is extracted using a specially designed algorithm to ensure the expressive power and discriminative power of the features.

[0030] S12. Input the multidimensional feature vector into a preset intelligent decision-making model, and have the intelligent decision-making model output a decision vector; In S12, the intelligent decision-making model is an intelligent decision-making classification neural network. This network adopts a multi-task learning architecture. The pre-defined intelligent decision-making model includes a sequentially connected input layer (64-dimensional), three hidden layers (128-64-32 neurons), and a dual-branch output structure layer. The dual-branch output structure layer consists of a policy classification branch output layer and a parameter regression branch output layer. The input layer receives the multi-dimensional feature vector and transmits it to the hidden layer. The hidden layer maps and compresses the multi-dimensional feature vector to obtain shared features. The policy classification branch output layer uses a softmax activation function to perform policy classification processing on the shared features, outputting the probability distribution of five preprocessing policies. The parameter regression branch output layer uses a sigmoid activation function to perform parameter regression processing on the shared features, outputting three continuous parameters: enhancement strength, noise level, and contrast adjustment. The loss function uses a weighted multi-task loss, with a policy classification weight of 0.7 and a total parameter regression weight of 0.3.

[0031] S13. The original image is enhanced using the decision vector to obtain the preprocessed image.

[0032] In S13, the preprocessing strategy comprises five specialized preprocessing algorithms, each with a specific processing flow and parameter configuration. Document-specific processing includes four steps: tilt correction, contrast enhancement, noise suppression, and text sharpening; scene-specific processing includes four steps: white balance correction, dynamic range optimization, detail enhancement, and color saturation adjustment; and low-quality enhancement processing includes four steps: super-resolution reconstruction, noise reduction, contrast enhancement, and edge sharpening.

[0033] S2. Use a preset multimodal large model to perform visual reasoning on the preprocessed image to obtain scene understanding information and image type; Specifically, the multimodal large model includes a visual reasoning module and a desensitization module connected in sequence.

[0034] In step S2, the preprocessed image is received, and the image type classification result and scene understanding information are output. The processing adopts a multi-round analysis mechanism: basic scene analysis → content relevance verification → comprehensive classification decision.

[0035] S21. Use the visual reasoning module in the multimodal large model to perform scene analysis on the preprocessed image to obtain scene understanding information; In S21, the basic scene analysis utilizes the fine-grained visual reasoning capabilities of a multimodal large model to analyze the main components, spatial layout relationships, and scene semantic features of the image. The algorithm employs an attention mechanism to focus on key regions and generates scene understanding information through multi-level feature extraction and fusion. This scene understanding information is a deep semantic representation of the scene.

[0036] S22. Calculate the correlation between the preprocessed image and the preset sensitive content to obtain a content correlation score; In step S22, content relevance verification verifies the association between image content and sensitive information by constructing a knowledge graph of sensitive information. The algorithm includes three steps: entity recognition, relation extraction, and relevance scoring, ultimately outputting a relevance score of 0-1. The knowledge graph includes the field of traffic accident investigation and can be used to verify the association between image content and traffic accidents.

[0037] S23. Perform Bayesian fusion on the scene understanding information and the content relevance score to obtain the image type.

[0038] In S23, the comprehensive classification decision adopts a Bayesian fusion method, which integrates the basic scene analysis results, relevance verification scores, and prior knowledge to calculate the posterior probability of each image type. The decision rule adopts the maximum posterior probability principle, and outputs the final classified image type ∈ {real scene, document photo, surveillance screenshot, other types} and confidence score. When the confidence score is lower than the threshold, a manual review mechanism is initiated.

[0039] S3. Using the multimodal large model, perform sensitive information detection on the preprocessed image according to the image type to obtain sensitive information detection results; In step S3, the preprocessed image and the image type described in this step are used to output the sensitive information detection result. The detection process adopts a multimodal joint detection mechanism: constructing the corresponding detection prompt template → multimodal large model coarse detection → small model refinement → confidence fusion.

[0040] S31. Construct a detection prompt template corresponding to the image type; In S31, the prompt template contains four core components: task description component, precision requirement component, output format component, and context constraint component.

[0041] S32. Using the detection prompt template, guide the desensitization module in the multimodal large model to perform coarse-grained sensitive information detection on the preprocessed image to obtain coarse-grained detection results; In S32, the multimodal large model utilizes its built-in visual localization capability to generate coarse detection results. This result contains multi-dimensional information: Bounding box coordinates:

[0042] Detection confidence level:

[0043] Information type:

[0044] Obstruction level:

[0045] For each coarse detection bounding box Candidate regions are extracted, and spatial localization is performed by combining the 2D coordinates and depth information of the bounding box, thus accurately extracting the target region even in the presence of occlusion or background interference. The extracted candidate regions are shown below. It is then fed into the standardization processing module to be scaled up uniformly. The pixels are preserved while maintaining the original aspect ratio to avoid image distortion affecting subsequent detection accuracy. Then, the standardized candidate regions undergo quality enhancement processing. This quality enhancement process employs a multi-dimensional joint optimization strategy to enhance the contrast, noise, and sharpness of the candidate regions. Specifically, contrast enhancement uses an adaptive histogram equalization algorithm, noise suppression uses a bilateral filtering algorithm to balance smoothness and edge preservation, and sharpness enhancement uses an unsharpened mask algorithm to improve high-frequency details, thus providing higher-quality input for fine-grained localization.

[0046] S33. Input the coarse-grained detection results into a preset lightweight model to refine the bounding box and obtain a refined bounding box; In S33, the lightweight model employs a five-layer CNN refined network. The refined network has a total parameter count of only 2.3MB, yet possesses powerful bounding box refinement capabilities. The network structure comprises three convolutional layers and two fully connected layers, with each convolutional layer followed by batch normalization and a ReLU activation function. The core function of the network is to predict the fine-tuned offset of the bounding box.

[0047] The lightweight model first calculates the geometric parameters of the original bounding box, including: width:

[0048] high:

[0049] Center point coordinates:

[0050]

[0051] Then, fine-tune using the offset: New center point x-coordinate:

[0052] New center point y-coordinate:

[0053] New width:

[0054] A new height:

[0055] Finally, the refined bounding box is reconstructed based on the new parameters. .

[0056] S34. The coarse-grained detection results and the refined bounding box are fused to obtain the sensitive information detection results.

[0057] In S34, the confidence scores of the large and small multimodal models are first normalized using the Sigmoid function. The scope is determined, and then the fusion weights are dynamically calculated based on the detection quality. The quality assessment algorithm considers three dimensions: bounding box tightness, target sharpness, and detection consistency.

[0058] The weighting formula is: Multimodal large model weights:

[0059] Small model weights:

[0060] Finally, the sensitive information detection result is obtained through weighted fusion, and the sensitive information detection result includes precise bounding box, information type, confidence level and depth information.

[0061] By working together with multimodal large and small models, the bounding box localization error was reduced from an average of 8.5 pixels to 1.2 pixels, improving accuracy by 86%, while maintaining high processing speed and robustness.

[0062] S4. Use the scenario understanding information to perform a risk assessment on the sensitive information detection results, obtain a risk score, and use the risk score to determine a desensitization decision; In step S4, the sensitive information detection results and the scene understanding information are received, and a desensitization decision is output. The process adopts a multi-level analysis mechanism: risk assessment → decision determination.

[0063] S41. Based on the sensitive information detection results, calculate the information type risk respectively. Exposure level risk Identifiability risk and contextual risks ; In S41, the information type risk The calculation formula is as follows:

[0064] Specifically, by iterating through all detected sensitive information types using the process of `i`, The type weight is the type weight corresponding to the i-th type of sensitive information. The sensitivity score is given for the i-th type of sensitive information.

[0065] Type weights can be configured as: face License plate Identification number ,sign Sensitivity score The results are calculated by analyzing the image features of the target area, including three dimensions: target size, sharpness, and integrity.

[0066] The level of exposure risk The calculation formula is as follows:

[0067] in, The visual saliency score for sensitive information is calculated using the Itti-Koch visual saliency model, which considers three features: color contrast, brightness contrast, and orientation contrast. Sensitivity scores are assigned to sensitive information based on the degree of occlusion. Clarity of areas containing sensitive information The calculation formula is as follows: .

[0068] The identifiable risks The calculation formula is as follows:

[0069] in, The resolution quality score for sensitive information is calculated based on the pixel density of the target region in the original image. The clarity score for sensitive information is evaluated using the Laplace operator. The integrity score for sensitive information is evaluated through edge continuity analysis.

[0070] The context risk The calculation formula is as follows:

[0071] in, The importance score for sensitive information in a scene is determined by analyzing the position and size of the target in the image; the central area and larger targets have higher importance scores. Score the usage scenarios of sensitive information, calculated based on the preset risk levels of application scenarios; Score the legal compliance of sensitive information, determined according to the privacy protection requirements of relevant laws and regulations.

[0072] S42. Combine the risk of information type , the risk of exposure degree , the risk of identifiability and the context risk by weighted fusion to obtain a risk score .

[0073] In S42, the risk score is calculated by the following formula:

[0074] where α is the weighted coefficient of the information type risk, β is the weighted coefficient of the exposure degree risk, γ is the weighted coefficient of the identifiability risk, and δ is the weighted coefficient of the context risk.

[0075] Specifically, the specific values of the following coefficients are optimized through a large amount of experimental data to ensure the accuracy and rationality of risk assessment:

[0076] S43. Map the risk score to the corresponding risk score range, and set different candidate desensitization strategy sets for different risk score ranges; In S43, first map the desensitization strategy to the risk range based on the total risk score Rtotal to limit the candidate desensitization strategy sets within each risk score range. When Rtotal > 0.8, add the Gaussian blur strategy to the candidate strategy set; when 0.6 < Rtotal ≤ 0.8, add the pixelation strategy to the candidate strategy set; when 0.4 < Rtotal ≤ 0.6, add the mosaic strategy to the candidate strategy set; when Rtotal ≤ 0.4, add the mild blur strategy to the candidate strategy set.

[0077] S44. Evaluate the applicability of the candidate desensitization strategies in the candidate desensitization strategy set to obtain the applicability scores of each candidate desensitization strategy; In S44, the applicability score of each strategy is calculated by weighted calculation in four dimensions:

[0078] where is the expected score of type matching degree, is the expected score of risk adaptability, is the context coordination This represents the expected quality score.

[0079] S45. The applicability score is used to evaluate the utility of each candidate desensitization strategy to obtain the utility function of each candidate desensitization strategy, and the desensitization strategy with the largest utility function is selected as the final desensitization decision.

[0080] In S45, the final desensitization decision for:

[0081] in, Risk scoring for sensitive information Sensitive information type The location information of sensitive information in the image. For the first j The utility function corresponding to each candidate desensitization strategy; :

[0082] Utility function weight dynamic adjustment mechanism:

[0083] When the risk level is high, the algorithm automatically increases the risk control weight and decreases the naturalness weight to ensure that safety is prioritized.

[0084] Each strategy has a set of adjustable parameters, and the algorithm is based on... , and These parameters are dynamically optimized.

[0085] For example: The blur radius of the Gaussian blur strategy:

[0086] Block size for pixelation strategy:

[0087] Through multi-dimensional quantitative risk assessment and intelligent strategy selection, the accuracy of desensitization decision-making is improved by 35%, and the misjudgment rate is reduced by 50%, achieving truly intelligent personalized desensitization processing.

[0088] S5. Using the desensitization decision and the sensitive information detection results, the preprocessed image is desensitized to generate a desensitized image.

[0089] In S5, this step achieves high-quality de-identification through a precise combination of adaptive expansion, multi-style processing, and pixel-level execution. Adaptive expansion dynamically calculates the expansion coefficient and direction of the bounding box based on information type, risk level, and contextual features. Expansion considers three factors: target integrity, edge security, and visual naturalness, ensuring complete coverage while maintaining visual appeal. In multi-style processing, each de-identification style controls its visual effect through a set of parameters. The algorithm automatically optimizes these parameters based on risk assessment results and visual quality requirements, achieving a balance between risk control and visual quality. Pixel-level execution ensures the accuracy and quality of the de-identification process through sub-pixel precision coordinate transformation and accurate pixel value calculation. The algorithm includes three sub-modules: anti-aliasing, edge smoothing, and color coordination, comprehensively improving the de-identification effect.

[0090] Specifically, the process employs an adaptive processing mechanism: boundary expansion → mask generation → gradient processing → color coordination. The boundary expansion stage outputs the expanded bounding box, the mask generation stage outputs the pixel-level mask, the gradient processing stage outputs the gradient blending result, and the color coordination stage outputs the final desensitized image.

[0091] S51. Adaptively expand the sensitive information bounding box to obtain an expanded bounding box; In S51, the adaptive expansion adopts a differential configuration algorithm for basic expansion coefficients: Each information type has a specific basic expansion factor. Face is License plate Identity number is Signature This configuration is determined based on the geometric characteristics of the information itself and protection requirements, ensuring full coverage while avoiding over-expansion.

[0092] Context-aware dynamic adjustment algorithm: The final formula for calculating the expansion coefficient is:

[0093] Context complexity Depth variation is calculated by analyzing the texture complexity and information density of the area surrounding the target. Determination of depth information changes based on 3D bounding boxes.

[0094] S52. Based on the desensitization decision, generate a mask on the extended bounding box to obtain a mask box; In S52, the core technology of the adaptive masking generation algorithm is to dynamically adjust the Gaussian kernel parameters based on the characteristics of the target region. The standard deviation calculation formula is:

[0095]

[0096] in For desensitization intensity, Aspect ratio, This is a depth adjustment factor. The formula for calculating the Gaussian kernel size is:

[0097] Adjusting to an odd number ensures symmetry.

[0098] The block size adaptive mechanism of the intelligent pixelation masking algorithm: base block size The final block size is calculated using the following formula:

[0099] This algorithm supports arrive Dynamic block size adjustment for pixels ensures that larger block sizes are used for higher risk levels. Color retention strategy uses retention parameters. Control, when Use the average value within the block to fill when the value is within the block; otherwise, use the median value.

[0100] The dynamic mosaic masking algorithm's multi-pattern generation mechanism maintains four basic pattern modes: checkerboard, diagonal, random, and circular. Pattern selection is calculated using complexity levels.

[0101] Ensure that different patterns are used for different levels of complexity. Each pattern has a dedicated generation algorithm and color filling strategy.

[0102] S53. The mask frame and the preprocessed image are fused to obtain a fused image.

[0103] In S53, the smooth transition technique of the edge gradient processing algorithm adopts a gradient mask generation algorithm: Create a gradient mask matrix of the same size as the masking area. For each pixel Calculate the minimum distance from it to the four boundaries:

[0104] when Apply the gradient function when appropriate, otherwise set it to... .

[0105] Mathematical implementations of three gradient functions: Smooth step function:

[0106] Easily in / escape function:

[0107] Linear functions:

[0108] Gradient width Available Adjust within the pixel range; recommended value is... Pixel.

[0109] The intelligent matching technique of the adaptive color coordination algorithm uses the K-means clustering algorithm for environmental color extraction: The system expands around the masking bounding box. Pixels as analysis area Convert all pixels in the region to RGB color vectors, and then apply the K-means clustering algorithm (...). Maximum number of iterations random seed Extract three main color tones .

[0110] Color similarity detection and blending algorithm: For each pixel in the masked region, calculate its Euclidean distance to each dominant color:

[0111] when (default When mixing colors, the process is as follows:

[0112] Regional color coordination mechanism: The system creates an independent similarity mask for each primary color tone. This ensures that different areas are color-adjusted based on the most recent primary color tone, achieving a precise and natural color coordination effect.

[0113] Ultimately, the mask integrity reached 99.2%, visual naturalness was 94.5%, edge smoothness was 96.8%, and the processing speed of a single area was less than 0.1 seconds, achieving a high-quality and efficient natural desensitization effect.

[0114] In addition, this embodiment also includes a step of performing a quality check on the desensitized image.

[0115] This step receives both the anonymized image and the original image, and outputs a quality verification report. The verification process employs a comprehensive evaluation mechanism: integrity check → accuracy verification → visual consistency → privacy compliance.

[0116] The integrity check stage performs secondary sensitive information detection on the desensitized image to verify whether any sensitive information is missing, and outputs the omission detection results. The secondary detection uses an independent detection model and a more stringent detection threshold to ensure the reliability of the check. The accuracy verification stage outputs accuracy evaluation results through three dimensions: bounding box accuracy verification, desensitized area integrity verification, and visual effect quality verification, to comprehensively evaluate the accuracy of the desensitization process. The visual consistency stage uses three indicators: structural similarity (SSIM), peak signal-to-noise ratio (PSNR), and visual quality score (VQS) to output a consistency score, quantitatively evaluating the consistency between the desensitized image and the original image in non-sensitive areas.

[0117] The privacy compliance phase, based on relevant laws, regulations, and industry standards, checks whether the de-identification process meets privacy protection requirements and outputs a compliance assessment. The algorithm includes three steps: sensitive information coverage check, irreversibility verification, and compliance scoring. Finally, a comprehensive quality report is generated, which includes scores for various indicators and an overall quality rating.

[0118] Through the above seven-stage precision processing flow, this embodiment achieves the following technological breakthroughs: Accuracy performance metrics: Face recognition accuracy: 95.2% (15-25 percentage points improvement compared to traditional methods); License plate recognition accuracy: 94.8% (20-30 percentage points improvement compared to traditional methods); Document information recognition accuracy: 93.5% (20-35 percentage points improvement compared to traditional methods); Bounding box localization error: Average 1.2 pixels (86% accuracy improvement). Quality performance indicators: Integrity check pass rate: 98.7% (sensitive information coverage); Accuracy verification pass rate: 96.4% (accuracy of bounding boxes and desensitized areas); Visual consistency score: 94.1% (consistency with the original image in non-sensitive areas); Privacy compliance score: 99.2% (complies with relevant laws and regulations). Efficiency performance indicators: Overall processing time: 3.2 seconds / sheet ( (Images); System throughput: 1125 images / hour; Concurrency processing capacity: 50 concurrent connections; Peak memory usage: 2GB; GPU utilization: over 85% The technological innovation of this process lies in the fact that, through a seven-stage pipeline design and precise algorithm coordination, it achieves end-to-end intelligent processing from the original image to a high-quality desensitized image, solving the key technical problems of traditional methods in terms of accuracy, efficiency, and adaptability.

[0119] This embodiment achieves significant breakthroughs in several key technical indicators through innovative technical architecture design, and has significant technical advantages and practical application value compared with traditional technologies.

[0120] (1) Significant improvement in accuracy performance Breakthrough in pixel-level precise localization technology: Through a joint detection architecture guided by a multimodal large model and refined by a small model, pixel-level precise localization of sensitive information is achieved. Compared to the average localization error of 8-15 pixels in traditional methods, this invention controls the error to within 1.2 pixels, improving accuracy by more than 86%. Specific technical indicators are as follows: Face recognition accuracy of 95.2%: This represents a 10-25 percentage point improvement in overall recognition accuracy compared to traditional methods' 70-85% under complex scenes, partial occlusion, and varying lighting conditions. License plate recognition accuracy is 94.8%: It supports multiple types of license plates, including standard license plates, temporary license plates, and foreign license plates, which is 14-29 percentage points higher than the 65-80% accuracy of traditional methods. Document information recognition accuracy is 93.5%: It includes various document contents such as identification numbers, signatures, and organization information, which is an improvement of 18-35 percentage points compared to the 60-75% accuracy of traditional methods. Bounding box localization error of 1.2 pixels: Sub-pixel accuracy target localization is achieved through a five-stage joint detection and refinement algorithm. (2) Leapfrog development in the level of intelligence Breakthrough in multi-dimensional intelligent decision-making technology: Through a multi-modal fusion intelligent decision-making mechanism, a technological leap from rule-driven to intelligence-driven decision-making has been achieved. Key technological advantages include: Context-based intelligent judgment: The system possesses deep semantic understanding capabilities, enabling it to analyze the correlation between information content and its contextual environment, and make intelligent de-identification judgments, reducing the false judgment rate by 50%. Zero-shot learning capability: Leveraging the powerful generalization ability of multimodal large models, new types of sensitive information can be identified without retraining, significantly improving adaptability and scalability. Adaptive parameter adjustment: The system can automatically optimize processing parameters and strategy selection based on image features, risk level, and application scenario to achieve personalized intelligent processing. Reduced human intervention: Through end-to-end intelligent processing, the need for human intervention is reduced by more than 90%, significantly improving processing efficiency. (3) Significant optimization of processing efficiency Technical advantages of the unified multimodal processing framework: Through innovative architecture design and algorithm optimization, it achieves a significant improvement in processing efficiency. End-to-end processing time: 3.2 seconds per image. For a 1024×1024 pixel image, the process from input to output takes only 3.2 seconds, an improvement of over 60% compared to the traditional method's 8-15 seconds. System throughput of 1125 images / hour: Supports large-scale batch processing, meeting the high throughput requirements of practical applications. 50-channel concurrent processing capability: Supports concurrent processing of multiple image streams through parallel computing and resource optimization. Resource utilization optimization: GPU utilization exceeds 85%, peak memory usage is only 2GB, and resource allocation is efficient and reasonable. (4) Comprehensive enhancement of adaptability and generalization ability Breakthrough in multi-scenario adaptability: Through a unified multimodal processing framework and intelligent algorithm design, perfect adaptation to various complex scenarios is achieved. Multi-image type support: It can process various types of images such as on-site photos, document photos, surveillance screenshots, and case files, without the need to design special algorithms for each type. Multi-information type recognition: Supports the recognition and processing of various sensitive information such as faces, license plates, ID numbers, signatures, corporate logos, barcodes, and QR codes. Adaptation to complex scenes: Maintains high-precision recognition capability even in challenging scenarios such as densely packed multiple subjects, partial occlusion, varying lighting, and complex backgrounds. Rapid Adaptation to New Scenarios: When new types of sensitive information or application scenarios emerge, the system can adapt quickly without the need for large-scale retraining. (5) Comprehensive assurance of quality control Breakthrough in multi-dimensional quality control technology: A six-layer quality control mechanism ensures the reliability and compliance of desensitization results. Completeness 98.7%: Through a secondary detection mechanism, complete coverage of sensitive information is ensured, with an omission rate of less than 1.3%. 96.4% accuracy: Multiple verification mechanisms ensure high accuracy in bounding box positioning and desensitized areas. Visual consistency 94.1%: Through intelligent color coordination and edge gradient processing, the desensitized areas are ensured to blend naturally with the surrounding environment. Privacy compliance 99.2%: Through legal and regulatory adaptation mechanisms, we ensure that the anonymization process complies with relevant privacy protection requirements. (6) Significantly enhanced practical application value Significant implications for the industrialization of this technology: The technological breakthrough of this invention provides important support for the practical application and industrialization of image desensitization technology. Lowering the application threshold: Through intelligent and automated processing, the difficulty and operational threshold of image desensitization technology are reduced. Enhance application effectiveness: Through high-precision and high-quality desensitization processing, improve practical application effectiveness and user experience. Expanding application scope: Through strong adaptability and generalization capabilities, the application scope and market space of the technology can be expanded. Promoting Industrial Development: Through technological breakthroughs and performance improvements, we will promote the industrialization and standardization of image desensitization technology. In summary, this embodiment, through innovative technical architecture design and sophisticated algorithm implementation, has achieved technological breakthroughs in multiple dimensions such as accuracy, intelligence, efficiency, and adaptability, providing important technical support and practical guidance for the development and application of image desensitization technology.

[0121] Example 2 Taking the desensitization of traffic accident investigation images as an example, this embodiment further illustrates the image desensitization method based on a multimodal large model.

[0122] In traffic accident investigations, the resulting image data is highly complex and diverse, including various types such as on-site photos, case file photos, and surveillance screenshots. These materials contain a large amount of sensitive information of different types, posing significant technical challenges to anonymization processing.

[0123] (1) Image preprocessing implementation method Input: Original image Processing flow: Multimodal feature extraction comprehensively describes the multimodal features of an image through 64-dimensional feature vectors, providing accurate data support for intelligent decision-making.

[0124] Texture feature extraction (8-dimensional): The system converts the image into a grayscale image and then applies the Local Binary Pattern (LBP) algorithm to extract texture features. The LBP operator generates an 8-bit binary pattern by comparing the grayscale values ​​of the center pixel with the surrounding 8 pixels, and finally calculates the frequency of different patterns as the texture feature vector.

[0125] Edge feature extraction (8-dimensional): The system uses the Canny edge detection algorithm to extract edge information of the image, and then calculates the statistical features of the edge, including mean, standard deviation, edge density, maximum and minimum values, quantiles and edge gradient energy, to comprehensively describe the edge structure features of the image.

[0126] Color feature extraction (12-dimensional): The system converts the image from the RGB color space to the HSV color space, and then calculates the statistical features of the three channels of hue, saturation and lightness, including the mean and standard deviation, for a total of 6 basic features and 6 statistical features, which fully describe the color distribution characteristics of the image.

[0127] Sharpness Feature Extraction (8-dimensional): The system uses the Laplacian operator to calculate the second derivative of the image and evaluate the image's sharpness level. Features include the variance of the Laplacian response, the mean and quantiles of the absolute values, as well as the proportion of high-response pixels and gradient energy, used to quantify the image's focus quality.

[0128] Contrast Feature Extraction (8-dimensional): The system calculates the overall contrast of the image based on the Michelson contrast formula and further analyzes the statistical features related to contrast, including contrast, gray-level variance, dynamic range, mean normalization, high and low pixel distribution ratio, and relative contrast of different regions.

[0129] Noise feature extraction (8 dimensions): The system uses a 3×3 high-frequency filter to extract the high-frequency components of the image, and then analyzes the statistical characteristics of the noise level, including the noise standard deviation, variance, absolute mean and extreme values, as well as the noise density and the proportion of abnormal noise.

[0130] Structural Feature Extraction (12-Dimensional): The system uses the Sobel operator to calculate the horizontal and vertical gradients of the image, and then calculates the gradient magnitude and direction. Finally, it statistically analyzes the histogram distribution of the gradient directions, weights the gradient magnitude information, and describes the structural features and directionality of the image.

[0131] (2) Implementation of intelligent decision classification Intelligent decision-making based on neural network classifiers: Neural network classifier architecture design: The system constructs a specialized multi-task neural network classifier with one input layer, three hidden layers, and two output branches, capable of simultaneously performing policy classification and parameter regression prediction.

[0132] Network architecture design: The first layer is a fully connected layer with 128 neurons, using the ReLU activation function, and the input dimension is a 64-dimensional feature vector. This layer is followed by a 30% Dropout layer to prevent overfitting. The second layer contains 64 neurons, also using the ReLU activation function, followed by a 20% Dropout layer. The third layer contains 32 neurons, using the ReLU activation function, and serves as the final layer for feature extraction.

[0133] The dual-branch output structure consists of two branches: the first branch is the policy classification branch, containing 5 neurons that use the softmax activation function to output 5 different preprocessing policy types; the second branch is the parameter regression branch, containing 3 independent output neurons, each using the sigmoid activation function to output enhancement intensity, noise level, and contrast adjustment parameters, respectively.

[0134] Loss function optimization configuration: The system uses the Adam optimizer with a learning rate of 0.001. The loss function adopts a multi-task weighted strategy. For the classification strategy, cross-entropy loss is used with a weight of 0.7; for the three regression tasks, mean squared error loss is used with a weight of 0.1 for each.

[0135] Intelligent decision-making reasoning algorithm: Upon receiving the 64-dimensional feature vector, the neural network classifier performs forward propagation computation. First, feature reshaping ensures the input format is correct, then the network outputs multiple prediction results. The system obtains the policy classification result through the argmax function, and directly outputs the specific values ​​for enhancement strength, noise reduction level, and contrast adjustment.

[0136] Decision output analysis: The strategy type is output as an integer, ranging from 0 to 4, corresponding to document specialization, scene specialization, low-quality enhancement, high-quality optimization, and special image processing, respectively. Enhancement intensity is output as a floating-point number from 0 to 1, representing the processing intensity. Noise reduction level is output as an integer from 0 to 3, representing the noise reduction level. Contrast adjustment is output as a floating-point number from 0.5 to 2.0, representing the contrast adjustment coefficient.

[0137] Specialized preprocessing execution: The system implements an intelligent preprocessing algorithm distribution mechanism, automatically selecting the optimal specialized processing flow based on the strategy type output by the decision classifier. The distribution algorithm supports five different processing strategies: document specialized processing (strategy type 0), scene specialized processing (strategy type 1), low-quality enhancement processing (strategy type 2), high-quality optimization processing (strategy type 3), and special image processing (strategy type 4).

[0138] The document-specific preprocessing algorithm, tailored to the characteristics of document images, implements four core processing steps: Step 1: Document Skew Correction Algorithm The system first converts the input RGB image to a grayscale image, and then uses the Canny edge detection algorithm to extract the document's edge information. The Canny algorithm uses a low threshold of 50 and a high threshold of 150, and employs a 3×3 aperture size for edge detection.

[0139] The system then applies the Hough transform to detect straight lines in the image, with an angular resolution of 1 degree and a distance resolution of 1 pixel. When a straight line with more than 100 voting points is detected, the system extracts the parameters of these lines and processes only the top 10 most significant lines to reduce computational complexity.

[0140] For each detected straight line, the system calculates its tilt angle relative to the horizontal line and filters out outlier angles with an absolute value exceeding 45 degrees. The retained set of valid angles is used to calculate the optimal tilt angle, employing a median algorithm instead of the average to improve robustness to outliers.

[0141] After obtaining the optimal tilt angle, the system calculates the geometric center of the image, constructs a 2×3 rotation matrix, and applies affine transformation to perform rotation correction on the image to ensure that the document content is restored to a horizontal position.

[0142] Step 2: Adaptive Contrast Enhancement The system enhances the contrast of the corrected document image based on the contrast adjustment parameters output by the decision classifier. A linear transformation formula is used, where the alpha parameter controls the contrast gain and the beta parameter controls the brightness shift. The alpha value typically ranges from 0.5 to 2.0, effectively improving document readability.

[0143] Step 3: Intelligent Noise Suppression When the decision classifier determines that noise reduction is needed, the system applies a median filter for noise suppression. The filter kernel size is dynamically adjusted based on the noise level, calculated as 3 + noise level × 2, supporting four kernel sizes: 3×3, 5×5, 7×7, and 9×9. Median filtering effectively removes impulse noise such as salt-and-pepper noise while maintaining the sharpness of document edges.

[0144] Technical effects: Through specialized document preprocessing, the tilt correction accuracy of the document image reaches over 95%, the contrast is improved by 40%, and the noise removal rate reaches 85%, resulting in a significant improvement in overall image quality. Output: Preprocessed image.

[0145] (3) Implementation of a joint detection architecture guided by a multimodal large model and refined by a small model The core innovation of this embodiment lies in the realization of a joint detection architecture that guides large-scale multimodal detection and refines small-scale detection. This architecture achieves a complete transformation from the original image to high-precision detection results through a five-stage precision processing flow.

[0146] Phase A: Coarse Detection Algorithm for Multimodal Large Models This stage receives the preprocessed image and outputs coarse detection results using the visual localization capabilities of a multimodal large model: Step A1: Build a specialized detection prompt template For traffic accident investigation scenarios, we have designed the following detection prompt template: Detection task: {detection_type} Testing requirements: Pixel-level precision positioning, with an error of no more than 2 pixels. Supports 3D spatial awareness and outputs depth information. Considering occlusion relationships and complex backgrounds Output bounding box coordinates [x1, y1, x2, y2, depth] and confidence score. Output format: JSON structured data Step A2: Multimodal large model inference processing The multimodal large model M_multimodal performs the visual localization task by taking the raw image I_raw as input and the specialized cue P_detection as input:

[0147] in, It includes the bounding box coordinates, detection confidence, information type, and occlusion degree from Example 1.

[0148] Technical effects: Leveraging the strong semantic understanding capabilities of multimodal large models, preliminary detection in complex scenarios is achieved, with an average detection accuracy of 85% and a bounding box error of approximately 8.5 pixels.

[0149] Phase B: Candidate Region Extraction and Preprocessing This stage uses the coarse detection result R_coarse to standardize and enhance the quality of each candidate region.

[0150] Algorithm implementation steps: Step B1: Candidate Region Extraction For each coarse detection result Execute region extraction:

[0151] Step B2: Image Normalization Processing Extracted candidate regions Standardized to Pixels:

[0152] Step B3: Quality Enhancement Treatment Enhance the quality of standardized areas:

[0153] Quality enhancement parameter configuration: Contrast enhancement factor: 1.2-1.5 (adaptive adjustment) Noise reduction strength: (Based on noise level) Sharpening level: 0.3-0.7 (based on image sharpness) Technical effects: Specialized preprocessing improved the quality of candidate regions by 45%, laying the foundation for subsequent small model refinement.

[0154] Phase C: Lightweight Small Model Refinement In this stage, a specially trained lightweight CNN model is used to adjust the bounding boxes with pixel-level precision.

[0155] Small model architecture design: CNN_Small architecture: Input layer: RGB image -- Convolutional layer 1: indivual Convolution kernel, ReLU activation, BatchNorm Pooling layer 1: Max pooling -- Convolutional layer 2: indivual Convolution kernel, ReLU activation, BatchNorm Pooling layer 2: Max pooling -- Convolutional layer 3: indivual Convolution kernel, ReLU activation, BatchNorm -- Fully connected layer 1: One neuron, Dropout ReLU activation -- Output layer: one neuron Linear activation Bounding box refining algorithm: Step C1: Offset Prediction Input the enhanced candidate region R_enhanced into the small model:

[0156] in, This indicates the offset of the bounding box.

[0157] Step C2: Coordinate Transformation Calculation Based on the original bounding box And offset, calculate the refined bounding box:

[0158] The meaning and definition of each parameter are given in Example 1.

[0159] Step C3: Boundary Constraint Processing Ensure the refined bounding box is within the image area:

[0160] Technical effects: Through refinement with a small model, the bounding box error was reduced from an average of 8.5 pixels to 1.2 pixels, improving accuracy by 86%.

[0161] Phase D: Confidence Calibration and Fusion This stage involves weighted fusion of the confidence scores of the large and small multimodal models.

[0162] Confidence calibration algorithm: Step D1: Confidence Normalization Normalize the two confidence levels to the range [0,1]:

[0163]

[0164] Step D2: Dynamic Weight Calculation The fusion weights are dynamically calculated based on detection quality.

[0165]

[0166] Step D3: Weighted fusion calculation

[0167] Technical effects: By using confidence fusion, detection stability is improved by 25% and the false positive rate is reduced by 40%.

[0168] Phase E: Final detection result output Output the final detection result containing complete information:

[0169] (4) Implementation of Multi-Dimensional Risk Assessment Algorithm This embodiment implements a unique multi-dimensional risk assessment algorithm, providing a quantitative basis for intelligent de-identification decisions.

[0170] The core formula for risk assessment is as follows:

[0171] Weighting coefficient configuration: (Information type weight) (Exposure weight) (Identifiability weight) (Context weight) Then, based on the total risk score Rtotal, a risk interval mapping is performed on the desensitization strategy to define the set of candidate desensitization strategies within each risk score interval. When Rtotal > 0.8, the Gaussian blur strategy is added to the candidate strategy set; when 0.6 < Rtotal ≤ 0.8, the pixelation strategy is further added to the candidate strategy set; when 0.4 < Rtotal ≤ 0.6, the mosaic strategy is further added to the candidate strategy set; when Rtotal ≤ 0.4, the mild blur strategy is further added to the candidate strategy set.

[0172] This strategy selection mechanism ensures an exact match between the desensitization intensity and the information sensitivity, avoiding the problems of over-desensitization or under-desensitization.

[0173] Specifically, the detailed calculation of the four risk components is shown in Embodiment 1.

[0174] Using the applicability score, the utility evaluation of each candidate desensitization strategy is carried out to obtain the utility function of each candidate desensitization strategy, and the desensitization strategy with the maximum utility function is selected as the final desensitization decision.

[0175] The final desensitization decision is shown in Embodiment 1.

[0176] The weights of the utility function are as follows: (Risk weight) (Naturalness weight) (Security weight) Technical effects: Through quantitative risk assessment, the accuracy rate of the desensitization decision is increased by 35%, and the misjudgment rate is reduced by 50%.

[0177] In this embodiment, an innovative multi-level adaptive mask generation technology is developed to achieve a natural desensitization effect with pixel-level accuracy.

[0178] (5) Implementation manner of the multi-level adaptive mask generation technology In this embodiment, an innovative multi-level adaptive mask generation technology is developed to achieve a natural desensitization effect with pixel-level accuracy. An innovative phased mask generation process is implemented: boundary expansion (E) → mask generation (M) → gradient processing (G) → color coordination (C) → final image (I_final) Phase 1: Adaptive expansion of the 3D bounding box To ensure complete coverage of sensitive information and handle incomplete edge situations, it is necessary to calculate the expansion coefficient based on the information type and context features.

[0179] First, set the basic expansion coefficients for different information types: Facial information: The basic expansion factor is 1.15. License plate information: The basic expansion coefficient is 1.10. Identification number: The basic expansion factor is 1.18. Signature information: The basic expansion factor is 1.12. Other types: The default base expansion factor is 1.10. Then, dynamically adjust based on contextual features: Context complexity adjustment: The higher the complexity, the more scalable the factor increases by 5%. Depth perception adjustment: The greater the change in depth, the more the expansion coefficient increases by 2%. The final formula for calculating the expansion factor is:

[0180] Phase 2: Pixel-level mask generation Three core masking algorithms are implemented to accurately mask sensitive information according to the selected strategy.

[0181] Adaptive Gaussian masking algorithm: This algorithm dynamically adjusts the Gaussian blur parameters based on region features to achieve a natural masking effect. The algorithm execution flow is as follows: Dynamic parameter calculation: The system first calculates the width and height of the bounding box, and then calculates the standard deviation of the Gaussian kernel based on the region size and desensitization intensity. The horizontal standard deviation σx is 15% of the minimum region size multiplied by the desensitization intensity, while the vertical standard deviation σy is further adjusted to take into account the aspect ratio variation.

[0182] Adaptive kernel size: The Gaussian kernel size is automatically calculated based on the standard deviation to ensure the continuity and naturalness of the blur effect. The kernel size is taken as 6 times the maximum standard deviation and adjusted to an odd number to ensure symmetry.

[0183] Two-dimensional Gaussian kernel generation: The system generates separate one-dimensional Gaussian kernels in the horizontal and vertical directions, and then obtains the two-dimensional Gaussian kernel through matrix multiplication.

[0184] Masking application: The generated 2D Gaussian kernel is applied to the sensitive area to achieve a smooth and natural blur effect.

[0185] Intelligent pixelation masking algorithm: The algorithm divides the sensitive region into small blocks and fills them with a uniform color, while also supporting adjustment of color retention. The algorithm execution flow is as follows: Adaptive block size calculation: The base block size is 4 pixels, which is dynamically adjusted according to the risk level. The final block size is 4 + risk level × 8 pixels, ensuring that higher-risk areas use larger block sizes.

[0186] Block processing loop: The system divides the region into non-overlapping blocks and processes each block independently. For each block ranging from 4×4 to 12×12 pixels, the system selects a processing method based on color preservation.

[0187] Color preservation strategy: When the color preservation level is greater than 0.5, the system uses the average value of pixels within the block to fill in the color and maintain the main color features; when the color preservation level is low, the median value is used to reduce the impact of outliers and improve visual consistency.

[0188] Result synthesis: All processed blocks are recombined to form a complete pixelated mask area.

[0189] Dynamic mosaic masking algorithm: This algorithm generates diverse mosaic patterns based on image complexity, enhancing the visual effect and irreversibility of the masking. The algorithm execution flow is as follows: Pattern selection mechanism: The system maintains four mosaic pattern modes: checkerboard pattern, diagonal pattern, random pattern, and circular pattern. A modulo operation is performed based on the complexity level and the total number of patterns to select the appropriate pattern, ensuring that different patterns are used for different levels of complexity.

[0190] Chessboard Pattern Generation: When a chessboard pattern is selected, the system uses an 8×8 pixel block size. For each block, the filling method is determined based on its position on the chessboard (odd or even of the row index + column index). Odd-numbered blocks use randomly generated RGB colors, while even-numbered blocks use the average color of the original area.

[0191] Diagonal Pattern Generation: When a diagonal pattern is selected, the system uses a 6x6 pixel block size. The fill strategy is determined by the sum of the block's row and column indices, creating an alternating diagonal visual effect. Blocks with the same diagonal position use a uniform fill method.

[0192] Color Fill Strategy: For all patterns, the system first attempts to fill with the average color of the original blocks to maintain harmony with the surrounding environment. When the original blocks are empty or invalid, a neutral gray (RGB value 128, 128, 128) is used for filling.

[0193] Result Integration: All processed blocks are recombined into a complete mosaic mask area to ensure pattern continuity and visual consistency.

[0194] Phase 3: Edge Gradient Processing To ensure a natural blend between the masked area and the original image, the system supports three edge gradient processing curves: Smoothstep gradient: This gradient method uses a cubic polynomial function to achieve a smooth transition, providing the most natural visual effect at the boundaries. The gradient function is... ,in It is a normalized distance parameter.

[0195] Ease in and out gradient: This gradient changes slowly at the beginning and end, and quickly in the middle, creating a smooth edge transition effect.

[0196] Linear gradient: This gradient method uses a simple linear relationship, and the further away from the boundary, the higher the transparency of the original image.

[0197] Smooth step gradient implementation: First, create a gradient mask matrix of the same size as the masking area. For each pixel in the mask, calculate its minimum distance to the four boundaries (top, bottom, left, and right). When the distance is less than the gradient width, apply a smooth step function; when the distance exceeds the gradient width, set it to completely opaque (value 1.0).

[0198] Finally, the system applies a gradient mask to the blend of the masked area and the original image, achieving a natural transition from a complete mask to the original image. The gradient width can be set to 2-10 pixels, but 5 pixels is typically used for the best visual effect.

[0199] Phase 4: Adaptive Color Coordination To achieve visual harmony between the masked area and the surrounding environment, the system extracts and analyzes the dominant color tone of the surrounding environment based on the K-means clustering algorithm and performs intelligent color adjustment.

[0200] Ambient color extraction algorithm: Surrounding region definition: The system extends the mask bounding box by 20 pixels as the analysis region to ensure sufficient environmental color information is collected. Simultaneously, boundary checks ensure that the extracted region does not exceed the original image area.

[0201] K-means clustering analysis: The system converts all pixels in the surrounding area into three-dimensional color vectors, and then uses the K-means clustering algorithm (3 cluster centers, 42 random seeds, 10 iterations) to extract three main hues. The clustering process automatically groups pixels with similar colors to form cluster centers representing the dominant hue of the environment.

[0202] Dominant color acquisition: After clustering is completed, the system obtains the color values ​​of the three cluster centers as the dominant color of the environment. These colors represent the most frequently occurring color combinations in the surrounding environment.

[0203] Intelligent color adjustment algorithm: Color similarity detection: For each pixel in the processed masked area, the system calculates its Euclidean distance to each dominant color. When the distance is less than a set threshold (default 30 color units), the pixel is considered similar to the dominant color.

[0204] Color blending strategy: The system uses a 30% ambient color blending ratio to adjust the colors of similar pixels. The new pixel color calculation formula is as follows: .

[0205] Regional adjustment: The system creates an independent similarity mask for each primary color tone, ensuring that color adjustments in different regions are based on the closest ambient primary color tone, achieving a more accurate and natural color coordination effect.

[0206] Technical Results: Mask Integrity: 99.2%; Visual Naturalness: 94.5%; Edge Smoothness: 96.8%; Processing Speed: Single Area <0.1 seconds Through the complete technical implementation described above, this embodiment achieves the following performance indicators: Detection accuracy indicators: Face recognition accuracy: 95.2%; License plate recognition accuracy: 94.8%; Document information recognition accuracy: 93.5%; Bounding box positioning error: 1.2 pixels.

[0207] Desensitization quality indicators: Integrity check pass rate: 98.7%; Accuracy verification pass rate: 96.4%; Visual consistency score: 94.1%.

[0208] System performance metrics: Overall processing time: 3.2 seconds / image; System throughput: 1125 images / hour; Concurrency processing capacity: 50 concurrent connections; Peak memory usage: 2GB; GPU utilization: over 85%.

[0209] Example 3 like Figure 3 As shown, this embodiment also proposes an image desensitization system based on a multimodal large model, including: The image preprocessing module is used to preprocess the original image to obtain the preprocessed image; The image classification module is used to perform visual reasoning on the preprocessed image using a preset multimodal large model to obtain scene understanding information and image type; The sensitive information detection module is used to perform sensitive information detection on the preprocessed image based on the image type using the multimodal large model, and obtain the sensitive information detection result; The risk scoring module is used to assess the risk of the sensitive information detection results using the scenario understanding information, obtain a risk score, and use the risk score to determine the desensitization decision. The image desensitization module is used to desensitize the preprocessed image using the desensitization decision and the sensitive information detection result, and generate a desensitized image.

[0210] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. For those skilled in the art, other variations or modifications can be made based on the above description. It is impossible to exhaustively list all the implementation methods here. All obvious variations or modifications derived from the technical solutions of the present invention are still within the protection scope of the present invention.

Claims

1. An image desensitization method based on a multimodal large model, characterized in that, include: S1. Preprocess the original image to obtain the preprocessed image; S2. Use a preset multimodal large model to perform visual reasoning on the preprocessed image to obtain scene understanding information and image type; S3. Using the multimodal large model, perform sensitive information detection on the preprocessed image according to the image type to obtain sensitive information detection results; S4. Use the scenario understanding information to perform a risk assessment on the sensitive information detection results, obtain a risk score, and use the risk score to determine a desensitization decision; S5. Using the desensitization decision and the sensitive information detection results, the preprocessed image is desensitized to generate a desensitized image.

2. The image desensitization method based on a multimodal large model according to claim 1, characterized in that, The preprocessing of the original image to obtain the preprocessed image includes: S11. Perform multimodal feature extraction on the original image to obtain a multidimensional feature vector; S12. Input the multidimensional feature vector into a preset intelligent decision-making model, and have the intelligent decision-making model output a decision vector; S13. The original image is enhanced using the decision vector to obtain the preprocessed image.

3. The image desensitization method based on a multimodal large model according to claim 2, characterized in that, The preset intelligent decision-making model includes an input layer, a hidden layer, and an output structure layer connected in sequence. The input layer receives the multidimensional feature vector and transmits it to the hidden layer. The hidden layer maps and compresses the multidimensional feature vector to obtain shared features. The output structure layer performs policy classification processing on the shared features to obtain a decision vector.

4. The image desensitization method based on a multimodal large model according to claim 1, characterized in that, The multimodal large model includes a visual reasoning module and a desensitization module connected in sequence. It uses the preset multimodal large model to perform visual reasoning on the preprocessed image to obtain scene understanding information and image type, including: S21. Use the visual reasoning module in the multimodal large model to perform scene analysis on the preprocessed image to obtain scene understanding information; S22. Calculate the correlation between the preprocessed image and the preset sensitive content to obtain a content correlation score; S23. Perform Bayesian fusion on the scene understanding information and the content relevance score to obtain the image type.

5. The image desensitization method based on a multimodal large model according to claim 1, characterized in that, The multimodal large model includes a visual reasoning module and a desensitization module connected in sequence. The step of using the multimodal large model to detect sensitive information in the preprocessed image based on the image type, and obtaining the sensitive information detection result, includes: S31. Construct a detection prompt template corresponding to the image type; S32. Using the detection prompt template, guide the desensitization module in the multimodal large model to perform coarse-grained sensitive information detection on the preprocessed image to obtain coarse-grained detection results; S33. Input the coarse-grained detection results into a preset lightweight model to refine the bounding box and obtain a refined bounding box; S34. The coarse-grained detection results and the refined bounding box are fused to obtain the sensitive information detection results.

6. The image desensitization method based on a multimodal large model according to claim 1, characterized in that, The risk assessment of the preprocessed image using the scene understanding information and the sensitive information detection results to obtain a risk score includes: S41. Based on the sensitive information detection results, calculate the information type risk respectively. Exposure level risk Identifiability risk and contextual risks ; Risk of the information type The calculation formula is as follows: in, The type weight is the type weight corresponding to the i-th type of sensitive information. Score the sensitivity of the i-th type of sensitive information; The level of exposure risk The calculation formula is as follows: in, Score the visual salience of sensitive information. Sensitivity rating for sensitive information; The identifiable risks The calculation formula is as follows: in, Score the resolution quality of areas containing sensitive information. Score the clarity of areas containing sensitive information. Score the integrity of sensitive information; The context risk The calculation formula is as follows: in, Score the importance of scenarios involving sensitive information. Rate the use cases of sensitive information. Score the legal compliance of sensitive information; S42. Risk of Information Type Exposure level risk Identifiability risk and contextual risks Weighted fusion yields a risk score. ; The risk score The calculation formula is as follows: Where α is the weighting coefficient for information type risk, β is the weighting coefficient for exposure level risk, γ is the weighting coefficient for identifiability risk, and δ is the weighting coefficient for context risk.

7. The image desensitization method based on a multimodal large model according to claim 6, characterized in that, The process of determining desensitization decisions using the risk score includes: S43. Map the risk score to the corresponding risk score interval, and set different sets of candidate desensitization strategies for different risk score intervals; S44. Perform an applicability evaluation on the candidate desensitization strategies in the candidate desensitization strategy set, and obtain an applicability score for each candidate desensitization strategy; S45. The applicability score is used to evaluate the utility of each candidate desensitization strategy to obtain the utility function of each candidate desensitization strategy, and the desensitization strategy with the largest utility function is selected as the final desensitization decision.

8. The image desensitization method based on a multimodal large model according to claim 7, characterized in that, The final desensitization decision for: in, Risk scoring for sensitive information Sensitive information type The location information of sensitive information in the image. For the first j The utility function corresponding to each candidate desensitization strategy; The first j The utility function corresponding to each candidate desensitization strategy The formula for calculating is as follows: in, For risk control weights, As a weight for visual naturalness, As a weight for visual naturalness, For security weights, For the first j Applicability score of each candidate desensitization strategy; The first j Applicability score of each candidate desensitization strategy The calculation formula is as follows: in, The expected score for type matching. For risk adaptability expectation score, For contextual consistency This represents the expected quality score.

9. The image desensitization method based on a multimodal large model according to claim 1, characterized in that, The sensitive information detection result includes a sensitive information bounding box. The step of using the desensitization decision and the sensitive information detection result to desensitize the preprocessed image and generate a desensitized image includes: S51. Adaptively expand the sensitive information bounding box to obtain an expanded bounding box; S52. Based on the desensitization decision, generate a mask on the extended bounding box to obtain a mask box; S53. The mask frame and the preprocessed image are fused to obtain a fused image.

10. An image desensitization system based on a multimodal large model, characterized in that, include: The image preprocessing module is used to preprocess the original image to obtain the preprocessed image; The image classification module is used to perform visual reasoning on the preprocessed image using a preset multimodal large model to obtain scene understanding information and image type; The sensitive information detection module is used to perform sensitive information detection on the preprocessed image based on the image type using the multimodal large model, and obtain the sensitive information detection result; The risk scoring module is used to assess the risk of the sensitive information detection results using the scenario understanding information, obtain a risk score, and use the risk score to determine the desensitization decision. The image desensitization module is used to desensitize the preprocessed image using the desensitization decision and the sensitive information detection result, and generate a desensitized image.

Images