Recoverable visual encryption method for privacy area of human body in image collected by unmanned aerial vehicle

Abstract

The present application relates to a recoverable visual encryption method for collecting human privacy areas in images by unmanned aerial vehicles, and a joint deep learning model is constructed by integrating a pre-trained target detection module, a self-adaptive visual encryption module and an enhanced visual decryption module; in the training stage, the parameters of the target detection module are frozen, and the encryption and decryption modules are optimized end-to-end through a joint loss composed of a composite decryption loss function and an encryption intensity loss function; in the application, the target detection module accurately locates the sensitive areas of the human body, the encryption module applies a size-adaptive and intensity-learnable pixelization transformation to generate a natural encrypted image, and under the authorization condition, the decryption module realizes high-fidelity original information recovery through an enhanced U-Net network. Compared with existing traditional image encryption or irreversible desensitization technology, the present application realizes seamless integration with mainstream target detection processes, ensures the integrity and recoverability of information while maintaining high visual concealment, and has the advantages of intelligence, high efficiency and strong practicality.

Description

Technical Field

[0001] This invention relates to a recoverable visual encryption method for human privacy areas in images captured by drones, which is particularly suitable for drone aerial photography and monitoring scenarios. It is used to securely process, controllably access, and restore human privacy areas in images or video frames with high fidelity, ensuring privacy and security while preserving the legal use value of the data. It belongs to the field of information security and artificial intelligence interdisciplinary technology. Background Technology

[0002] With the advancement of the Internet of Things and smart city construction, drones, with their flexibility and maneuverability, have been widely used in public safety inspections, intelligent traffic monitoring, and urban surveying, generating massive amounts of aerial images and video data. However, drone aerial photography inevitably captures personal privacy information such as pedestrian facial features and sensitive body parts. This data faces serious privacy leakage risks during transmission, storage, and sharing, raising serious legal and ethical challenges.

[0003] Existing privacy protection technologies have many shortcomings that make them difficult to adapt to the needs of drone scenarios. The specific technical problems are as follows:

[0004] Irreversibility leads to the loss of data value: Traditional privacy protection methods (such as Gaussian blur, fixed pixelation, and black masking) achieve privacy masking by permanently destroying the original pixel information, making the data unrecoverable. This makes encrypted data unable to meet the needs of legitimate scenarios such as judicial evidence collection and secondary data mining. Users are forced to choose between "privacy-protected incomplete data" and "unprotected original data," lacking a graded and controllable access mechanism.

[0005] 1. Poor adaptability and jarring visual effects: In drone aerial photography, the target scale varies greatly (human figures are large in the foreground and small in the background). Traditional methods use fixed parameters for processing, resulting in excessive blurring of small distant targets (loss of effective scene information) and insufficient occlusion of large foreground targets (failure of privacy protection). At the same time, the crude occlusion method destroys the naturalness and contextual continuity of the image, seriously interfering with the execution of downstream intelligent tasks such as behavior recognition and scene understanding.

[0006] 2. Insufficient real-time performance, making it difficult to adapt to dynamic drone scenarios: Traditional cryptographic schemes (such as AES selective encryption) are designed for text or binary data and are not optimized for visual data processing efficiency. Pixel-level high-intensity operations consume a large amount of computing resources. In drone high-definition video stream (30FPS and above) processing scenarios, stuttering and frame drops are prone to occur, failing to meet real-time requirements.

[0007] 3. Complex key management and high deployment difficulty: Traditional cryptographic schemes require complex key distribution, synchronization and storage mechanisms. In large-scale urban drone monitoring networks, the secure management of massive keys has become a technical bottleneck, limiting its large-scale application.

[0008] 4. Poor compatibility with downstream tasks: The regions processed by traditional encryption schemes exhibit irregular high-entropy random noise, which is not only visually jarring, but also completely unusable for downstream AI visual analysis tasks. Even if the unencrypted regions are clear, the noise blocks will interfere with tasks that rely on global context, such as scene classification and abnormal behavior detection.

[0009] In recent years, related technological explorations have failed to effectively balance the strength of privacy protection, the fidelity of information recovery, real-time processing efficiency, and scenario adaptability, thus failing to meet the comprehensive needs of drone aerial photography scenarios for "precise protection, efficient processing, authorized recovery, and downstream compatibility" of human privacy areas. Therefore, there is an urgent need for a privacy protection technology that combines intelligence, adaptability, recoverability, and high compatibility. Summary of the Invention

[0010] To overcome the shortcomings of existing technologies, this invention provides a recoverable visual encryption method for human privacy regions in images captured by unmanned aerial vehicles (UAVs). It aims to achieve secure, recoverable privacy protection and high-fidelity recovery of specific visual targets within a modern target detection framework. This method overcomes the limitations of traditional destructive anonymization techniques by deeply integrating visual encryption and decryption modules into a neural network and employing an innovative composite loss function for end-to-end training.

[0011] This invention, through modular design, adaptive algorithms, and joint optimization strategies, can not only apply sufficiently strong visual encryption to sensitive targets in images to ensure privacy and security, but also decrypt and recover the original information with high fidelity under authorized conditions. Furthermore, it can be seamlessly integrated into modern target detection processes as a native module, achieving intelligent and adaptive privacy protection.

[0012] A recoverable visual encryption method for human privacy areas in images captured by drones includes the following steps:

[0013] Step 1: Construct a joint deep learning model, which consists of a cascaded object detection module (a pre-trained object detection YOLO11n module), a visual encryption module (learnable), and a visual decryption module (for image restoration);

[0014] The target detection module is a pre-trained target detector used to output the location bounding boxes of human privacy regions. It uses a pre-trained YOLO11n model (or detectors that can output target bounding boxes such as Faster R-CNN, SSD, DETR) to accurately locate human privacy regions in images.

[0015] Visual encryption module: used to apply a recoverable adaptive pixelation visual transformation to the target area, while the original pixels of the non-target area remain unchanged;

[0016] The visual decryption module adopts an enhanced U-Net architecture that integrates channel attention modules and residual modules, including encoders, decoders, and skip connections between layers of the encoder and decoder. The enhanced U-Net architecture that integrates channel attention modules and residual modules is used to perform inverse transformation on the encrypted region to achieve high-fidelity image restoration.

[0017] Step 2: During the training phase, input a dataset of drone aerial images with ground truth bounding box annotations, and determine one or more human privacy target regions that require privacy processing based on the annotation information;

[0018] Specifically, a dataset of drone aerial images with ground truth bounding box annotations is obtained. The annotation information includes the bounding box coordinates (x, y, w, h) and category labels of human privacy regions. Each training image uses the annotation information to identify one or more target regions that require privacy processing.

[0019] Step 3: The visual encryption module applies an adaptive pixelated visual transformation to each target region to generate an encrypted image, while non-target regions retain their original pixels.

[0020] Specifically, the image and the identified target regions are input into the visual encryption module. This module applies an adaptive pixel-based visual transformation to each target region to generate an encrypted image. The granularity of this transformation is dynamically related to the size of the target region, and the visual intensity of the transformation is controlled by model parameters. Non-target regions of the image retain their original pixels.

[0021] Step 4: Input the encrypted image into the visual decryption module and output the reconstructed image;

[0022] Specifically, the encrypted image generated in step three is input into the visual decryption module. This module is a deep convolutional neural network whose task is to learn to perform an inverse transformation on the encrypted region and output a reconstructed image that is very similar to the original image.

[0023] Step 5: Define and calculate the loss functions, including the composite decryption loss function, the encryption strength loss function, and the overall joint loss function;

[0024] Specifically, for the image reconstruction task of the decryption module, a composite loss function is defined and computed to measure the difference between the reconstructed image and the original image. To ensure that the decrypted image can be restored with high quality in color, semantics, and detail, the loss function for this decryption task consists of three weighted parts: pixel-level first-norm loss (L1 Loss) to ensure the accuracy of basic pixel restoration, perceptual loss to ensure high-level semantic and texture consistency, and edge loss focused on contour sharpness restoration.

[0025] Step 6: During the training phase, freeze the parameters of the target detection module and jointly optimize the internal parameters of the visual encryption module and the visual decryption module based on the total joint loss function using the backpropagation algorithm.

[0026] Specifically, during training, the parameters of the object detection module are frozen. Based on the composite loss function calculated in step S5, the learnable parameters of the visual encryption module and all network parameters of the visual decryption module are jointly optimized through the backpropagation algorithm to achieve privacy masking while ensuring decryption recoverability.

[0027] Step 7: Inference Application Stage: The image to be processed is input into the model, the target detection module locates the human privacy area, and the visual encryption module generates an encrypted image for public use; in authorized scenarios, the encrypted image is input into the visual decryption module, which outputs a high-fidelity restored image.

[0028] Specifically, during the inference application phase, the user inputs the image to be processed into the trained joint model. The model first uses the object detection module to locate sensitive targets, and then the visual encryption module automatically encrypts these areas and outputs the result. When recovery is needed, authorized users can send the encrypted image to the visual decryption module to obtain the original visual content with high fidelity.

[0029] In step one, the target detection module is any one of YOLO11n, Faster R-CNN, or DETR based on Transformer. During the training phase, all network parameters are frozen to retain its target localization capability.

[0030] The adaptive pixelation visual transformation in step three specifically includes:

[0031] S3.1. Dynamically calculate the pixelated block size k based on the width h and height w of the target region:

[0032] ;

[0033] in d is the preset minimum block size, and d is the preset scale factor. This indicates the floor function;

[0034] S3.2. Perform average pooling on the target region, then upsample to restore it to its original size, creating a pixelated effect. The pixelated target region image is then processed. Compared with the original target region image Through learnable blending parameters The images are then fused to generate the final encrypted region image. :

[0035] ;

[0036] in, The value range is constrained to the (0,1) interval, which is used to control the intensity of the visual transformation.

[0037] The composite decryption loss function in step five The formula is:

[0038] ;

[0039] in ∈[0.5,0.7]、 ∈[0.2,0.3]、 ∈[0.1,0.2], For pixel-level average absolute error loss, For perceptual loss based on pre-trained VGG16 networks, This is the edge loss based on the Sobel operator.

[0040] The encryption strength loss function in step five The formula is:

[0041] ;

[0042] in To encrypt the image, For the original image, It is a binary mask. N represents the total number of pixels in the target region, and ⊙ represents element-wise multiplication.

[0043] The overall joint loss function in step five The formula is:

[0044] ;

[0045] in For target detection loss, ∈[0.05,0.15]、 ∈[0.4,0.6] represents the weight hyperparameter.

[0046] The encoder of the visual decryption module contains four convolutional blocks, each consisting of two 3×3 convolutional layers, a BatchNorm layer, and a ReLU activation function, and uses 2×2 max pooling downsampling; the decoder contains four convolutional blocks and uses 2×2 transposed convolution upsampling; the channel attention module is embedded after the network bottleneck layer and each stage of the decoder.

[0047] A recoverable visual information processing system, comprising a joint deep learning model, with built-in:

[0048] The target detection module is used to identify human privacy regions from images acquired by drones and output their location bounding boxes;

[0049] A visual encryption module, connected to the target detection module, is used to apply the adaptive pixelation visual transformation to the target region based on the bounding box to generate an encrypted image;

[0050] The visual decryption module is used to reverse process encrypted images and output high-fidelity original visual content.

[0051] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0052] 1. High-fidelity recovery, preserving the full value of data: Through a composite decryption loss function (pixel loss + perceptual loss + edge loss) and an enhanced U-Net decryption module, high-fidelity recovery of encrypted areas is achieved. Real-world testing shows that the recovered image achieves a peak signal-to-noise ratio (PSNR) of 33.04 dB and a structural similarity index (SSIM) of 0.9052, exhibiting minimal visual difference from the original image, fully meeting the needs of forensic evidence collection, secondary analysis, and other scenarios. Simultaneously, a tiered access mechanism of "encrypted public - authorized decryption" is provided, balancing privacy protection and data availability.

[0053] 2. Adaptive dynamic adaptation for natural visual effects: The pixel block size is dynamically adjusted based on the target area size, which solves the problem of large differences in target scale in drone aerial photography (avoiding excessive blurring of small distant targets and ensuring sufficient occlusion of large nearby targets); the learnable fusion parameter α dynamically optimizes the encryption strength, so that the encrypted area transitions naturally with the background, avoiding the visual abruptness of traditional methods and ensuring the normal execution of downstream intelligent tasks such as behavior recognition and scene understanding.

[0054] 3. Excellent real-time performance to meet the dynamic needs of UAV scenarios: Adopting the parallel computing characteristics of deep neural networks and the feature sharing design between modules, according to actual tests, on the UAV-borne NVIDIA Jetson Orin NX edge computing platform, the single-frame inference time is only 11.2 ms, and the frame rate reaches 89 FPS, far exceeding the real-time requirements of high-definition video streams (30 FPS), completely solving the problems of high latency and easy stuttering of traditional encryption algorithms.

[0055] 4. Secure and reliable, flexible deployment: The decryption logic is internalized into the weights of the deep neural network, making it extremely difficult to reverse engineer compared to traditional explicit key schemes; the end-to-end architecture requires no manual intervention and automatically completes the entire process of "detection-encryption-decryption", which can be seamlessly deployed in automated systems such as drone inspection and security monitoring; no complex key management mechanism is required, which lowers the technical threshold for large-scale deployment at the city level.

[0056] 5. Lossless detection accuracy and strong compatibility with downstream tasks: The parameters of the pre-trained target detection module are frozen during the training phase to ensure that the target detection accuracy is not affected after the introduction of privacy protection functions; the encrypted region adopts natural pixel transformation instead of random noise, which preserves the semantic contour information of the region. Downstream AI models can normally use the global context for analysis, which solves the defect of traditional solutions interfering with downstream tasks. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is a schematic diagram of the system data flow during the training phase of the present invention, illustrating the complete process of "input image → target localization → encryption generation → decryption and reconstruction → loss calculation → parameter update";

[0059] Figure 2 This is a schematic diagram of the system data flow during the inference application phase of this invention, illustrating the dual-path process of "encryption and public disclosure" and "authorization and decryption".

[0060] Figure 3 This is a schematic diagram of the enhanced U-Net network structure of the visual decryption module of the present invention, showing the layout of the encoder, decoder, skip connections, attention module and residual module;

[0061] Figure 4The image shows a visual comparison of the effects of this invention (left: encrypted image, middle: original image, right: decrypted image), which intuitively demonstrates the privacy masking effect and the high-fidelity restoration effect.

[0062] Figure 5 for Figure 4 The magnified view clearly shows the detail masking effect of the encrypted area and the quality of detail reconstruction after decryption. Detailed Implementation

[0063] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] Reference Figure 1 A recoverable visual encryption method for human privacy areas in images captured by drones includes the following steps:

[0065] Step 1: Construct a joint deep learning model, which consists of a cascaded object detection module (a pre-trained object detection YOLO11n module), a visual encryption module (learnable), and a visual decryption module (for image restoration);

[0066] Step 2: During the training phase, a dataset of drone aerial images with ground truth bounding box annotations is obtained. The annotation information includes the bounding box coordinates (x, y, w, h) and category labels of human privacy regions. Each training image identifies one or more target regions that require privacy processing based on the annotation information.

[0067] Step 3: Input the image and the identified target regions into the visual encryption module. This module applies an adaptive pixel-based visual transformation to each target region, generating an encrypted image. The granularity of this transformation is dynamically related to the size of the target region, and the visual intensity of the transformation is controlled by model parameters. Non-target regions of the image retain their original pixels.

[0068] Step 4: Input the encrypted image generated in Step 3 into the visual decryption module. This module is a deep convolutional neural network whose task is to learn to perform an inverse transformation on the encrypted region and output a reconstructed image that is very similar to the original image.

[0069] Step 5: For the image reconstruction task of the decryption module, define and compute a composite loss function to measure the difference between the reconstructed image and the original image. To ensure that the decrypted image can be restored with high quality in color, semantics, and detail, the loss function for this decryption task consists of three weighted parts: pixel-level first-norm loss (L1 Loss) to ensure the accuracy of basic pixel restoration, perceptual loss to ensure high-level semantic and texture consistency, and edge loss focused on contour sharpness restoration.

[0070] Step Six: During training, freeze the parameters of the object detection module. Based on the composite loss function calculated in Step S5, use the backpropagation algorithm to jointly optimize the learnable parameters of the visual encryption module and all network parameters of the visual decryption module to achieve privacy masking while ensuring decryption recoverability.

[0071] Step 7: In the inference application phase, the user inputs the image to be processed into the trained joint model. The model first uses the object detection module to locate sensitive targets, and then the visual encryption module automatically encrypts these areas and outputs the result. When recovery is needed, authorized users can send the encrypted image to the visual decryption module to obtain the original visual content with high fidelity.

[0072] Example 1:

[0073] The core of this embodiment is a joint encryption and decryption model. Based on the advanced target detection model architecture of YOLOv11n, this model deeply integrates visual encryption and visual decryption modules through inheritance and extension.

[0074] Detailed description of system components

[0075] 1. Target Detection Module:

[0076] This embodiment uses a pre-trained YOLO11n model on a custom dataset related to UAVs as the object detection module. YOLO11n was chosen because it achieves an excellent balance between detection accuracy and inference speed, making it ideal for applications requiring real-time processing. The method of this invention is not limited to YOLO11n and is also applicable to any other detector architecture that can output object bounding boxes, such as Faster R-CNN, SSD, or Transformer-based DETR. Its function is to receive the input image. It outputs a set of detection results, each containing the bounding box coordinates of the target. And its category label c. When jointly training the encryption and decryption module of the present invention, all network weight parameters of the detection module are set to an untrainable state (i.e., "frozen") to fully retain its powerful feature extraction and target localization capabilities, while significantly reducing training computational overhead.

[0077] 2. Visual encryption module:

[0078] This module is responsible for applying restorable visual transformations to specific target regions. Its input is the original image. and a set of target bounding boxes Its core workflow includes:

[0079] Adaptive block size calculation: For each target region defined by bounding box B with height h and width w, the module dynamically calculates a pixelated block (kernel) size k. The calculation formula is as follows:

[0080] ;

[0081] in, d is the preset minimum block size (6 in this embodiment) to prevent the complete loss of target information due to excessively small size; d is the scale division factor that controls the granularity as the size changes (8 in this embodiment). This is the floor function. This mechanism ensures the adaptability of the visual transformation effect.

[0082] Pixelation and learnable hybrid: The module first processes the target region image patch Pixelated image patches are obtained by applying average pooling and nearest neighbor upsampling. Subsequently, through a learnable scalar parameter (An nn.Parameter) is used to linearly fuse the two to obtain the final encrypted image patch. :

[0083] ;

[0084] in It is constrained within the interval during forward propagation. The value is automatically optimized during training, allowing the network to find the optimal balance between encryption strength and recoverability.

[0085] 3. Visual Decryption Module:

[0086] This module is the core of achieving high-fidelity image restoration; its essence is a specially designed deep network for image translation. For example... Figure 3 As shown, this embodiment uses an enhanced U-Net architecture:

[0087] U-Net backbone: The network has a symmetrical encoder-decoder structure and directly passes the shallow high-resolution features of the encoder to the decoder through "skip connections", which is crucial for restoring the fine texture and edges of the image.

[0088] Enhanced design:

[0089] a) Attention Block: Channel attention blocks are added after the bottleneck layer and each stage of the decoder in the network, enabling the network to adaptively learn the importance of feature channels, thereby focusing on the most valuable information for image reconstruction.

[0090] b) Residual Block: The basic convolutional unit of the network is composed of residual blocks, whose outputs are added to the inputs, making the network easier to learn identity mappings and fine corrections, thus enabling the construction of deeper and more powerful recovery networks.

[0091] Detailed explanation of the model training and evaluation process

[0092] The model training process of this invention is managed by a custom trainer, JointTrainer, and the core steps are as follows:

[0093] Forward propagation: For a batch of input raw images and its true target bounding box First, the visual encryption module is called to generate an encrypted image. Then, Input the image into the visual decryption module to obtain the reconstructed image. .

[0094] Composite decryption loss calculation: Calculation and Composite loss within the target region (defined by mask M) Its formula is:

[0095] ;

[0096] in , , These are the weighting hyperparameters for each loss term (set to 0.6, 0.25, and 0.15 in this embodiment). The definitions of each sub-loss term are as follows:

[0097] Loss: Calculate the average absolute error at the pixel level to ensure the consistency of the basic image content.

[0098] ;

[0099] in The total number of pixels within the target area. The original image, For the original image, It is a binary mask (target area is 1, background is 0). This is for element-wise multiplication.

[0100] Perceived loss ( ): Use a pre-trained VGG16 network as a fixed feature extractor Calculate the mean square error of the multi-layer feature map.

[0101] ;

[0102] in This represents the feature map extracted from the i-th layer. The weights for this layer are defined. This loss constrains the recovery result from a high-level semantic perspective, effectively avoiding image blurring.

[0103] Edge loss ( ): Using the Sobel operator Extract the edge gradient map and calculate its L1 distance.

[0104] ;

[0105] This loss term specifically penalizes blurred outlines, strongly driving the network to recover sharp object edges.

[0106] Mosaic encryption strength loss To prevent the visual encryption module from degenerating into an identity mapping (i.e., the output image is unchanged from the input image) and to ensure effective pixelation destruction of privacy regions, this invention introduces an encryption strength loss. This loss does not force the image to approximate a single mean mosaic, but rather maximizes the encryption strength of the image. With the original image The mean absolute difference within the masked region forces the network to move as far away from the original pixel distribution as possible while preserving reversible information.

[0107] ;

[0108] The formula normalizes and inversely optimizes the differences, forcing the network to generate sufficient visual changes in the masked area, thereby ensuring effective masking of privacy information.

[0109] Total joint losses This invention employs a multi-task joint optimization strategy, combining the aforementioned visual reconstruction task with the object detection task and encryption strength constraints to construct the final end-to-end total loss function.

[0110] ;

[0111] in, The YOLO-based object detection loss (including bounding box regression loss, classification loss, and DFL loss) is used to ensure that the model can still maintain high-precision object recognition capabilities on encrypted images. The composite decryption loss calculated above is used to supervise the learning of the visual decryption module; This represents a loss in the strength of the mosaic encryption. and To balance the hyperparameters of the task weights (in this implementation example, Set it to 0.1. Set to 0.5).

[0112] Backpropagation and Update: Based on the calculated total joint loss By using the backpropagation algorithm, all gradients of the visual encryption module, visual decryption module, and object detection backbone network are calculated simultaneously, and the model parameters θ are updated synchronously, thereby achieving collaborative optimization of the entire "encryption-detection-decryption" process.

[0113] Model Evaluation and Selection: After each training epoch, the model's performance is evaluated on an independent validation set. Evaluation metrics can include objective image quality assessment standards such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). The model weights that achieve the best overall performance on the validation set are ultimately selected as the final deployed model.

[0114] Detailed explanation of the model training and evaluation process:

[0115] After the model is trained, its application process in a real-world scenario is as follows:

[0116] Detection and encryption. Image to be processed. The input model is processed by the object detection module, which first locates the predicted bounding boxes of sensitive targets. Subsequently, the visual encryption module instantly encrypts these areas, generating an image. For public display, transmission or storage.

[0117] Decrypt on demand. In authorized scenarios, encrypted images will be decrypted. Input into the visual decryption module of the same model.

[0118] High-fidelity restoration. The decryption module outputs the restored image. It is visually almost indistinguishable from the original image, such as... Figure 4 As shown, the magnified view of the area is as follows: Figure 5 As shown.

[0119] Performance analysis and drone deployment feasibility:

[0120] To verify the airborne performance of the method of this invention on a real drone platform, we selected the Allspark2 scientific research drone equipped with an NVIDIA Jetson Orin NX high-performance edge computing module as the test model. This computing platform has a peak AI computing power of 100 TOPS (Sparse), providing deep learning inference capabilities at the embedded edge that surpass those of conventional consumer-grade desktop NPUs. In the control experiment, we used the Apple M3 chip (NeuralEngine computing power 18 TOPS) as the benchmark platform. When executing the best.pt model of this invention, the average inference time on the M3 platform was approximately 14.8 ms. During the drone airborne deployment phase, the model was converted to the TensorRT engine and deployed on the Orin NX platform. Thanks to its dedicated Tensor Core acceleration and INT8 quantization technology, the average inference time per frame of the model on the airborne end was reduced to 11.2 ms (approximately 89 FPS). Compared with the control group M3 chip, no performance bottleneck was observed; instead, it demonstrated the computing power advantage of dedicated edge AI chips in specific neural network tasks. Experimental results show that the method of the present invention can make full use of high-performance onboard computing power to complete complex visual perception tasks with extremely low latency.

[0121] Application process example

[0122] 1. Drones perform urban security patrols and collect aerial video streams in real time;

[0123] 2. The airborne edge computing platform runs the model of this invention, and the target detection module locates the human privacy area in the video frame in real time;

[0124] 3. The visual encryption module performs adaptive pixelation transformation on the privacy area to generate an encrypted video stream, which is transmitted to the ground monitoring center in real time (non-privacy areas remain clear and do not affect scene monitoring).

[0125] 4. In routine monitoring scenarios, staff can view encrypted video streams, effectively protecting their privacy.

[0126] 5. When an abnormal event is discovered and evidence needs to be collected, authorized staff will input the encrypted video frame into the decryption module and output a high-fidelity recovered image for event tracing and judicial evidence collection.

[0127] In summary, this invention, through a complete end-to-end technical solution, achieves intelligent and recoverable privacy protection of human body regions in UAV images, deeply integrated with modern target detection systems. It not only solves many drawbacks of traditional encryption methods, such as slow processing speed, but also achieves significant progress in protection strength, recovery fidelity, operational efficiency, and application flexibility. The technical solution of this invention has demonstrated its superior performance in actual hardware deployment, verifying its ability to meet the stringent requirements of real-time, efficient, and recoverable privacy protection of high-definition images or video frames when UAVs perform tasks such as inspection, security, or surveying. It possesses extremely high practical value and broad application prospects.

[0128] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.

Claims

1. A method for recoverable visual encryption of human privacy areas in images captured by drones, characterized in that: Includes the following steps: Step 1: Construct a joint deep learning model, which consists of a cascaded object detection module, a visual encryption module, and a visual decryption module; the object detection module is a pre-trained object detector used to output the bounding boxes of the human privacy region; the visual decryption module is an enhanced U-Net architecture that integrates a channel attention module and a residual module, including an encoder, a decoder, and skip connections between the encoder and decoder layers. Step 2: During the training phase, input a dataset of drone aerial images with real target bounding boxes labeled, and determine one or more human privacy target regions that need privacy processing based on the labeling information; Step 3: The visual encryption module applies an adaptive pixelated visual transformation to each target region to generate an encrypted image, while non-target regions retain their original pixels. Step 4: Input the encrypted image into the visual decryption module and output the reconstructed image; Step 5: Define and calculate the loss functions, including the composite decryption loss function, the encryption strength loss function, and the overall joint loss function; The composite decryption loss function in step five The formula is: ； in ∈[0.5,0.7]、 ∈[0.2,0.3]、 ∈[0.1,0.2], For pixel-level average absolute error loss, For the perceptual loss based on the pre-trained VGG16 network, The edge loss is based on the Sobel operator; The encryption strength loss function in step five The formula is: ； in To encrypt the image, For the original image, It is a binary mask. N represents the total number of pixels in the target region, and ⊙ represents element-wise multiplication; The overall joint loss function in step five The formula is: ； in For target detection loss, ∈[0.05,0.15]、 ∈[0.4,0.6] represents the weight hyperparameters; Step 6: During the training phase, freeze the parameters of the target detection module and jointly optimize the internal parameters of the visual encryption module and the visual decryption module based on the total joint loss function using the backpropagation algorithm. Step 7: Inference Application Stage: The image to be processed is input into the model, the target detection module locates the human privacy area, and the visual encryption module generates an encrypted image for public use; in authorized scenarios, the encrypted image is input into the visual decryption module, which outputs a high-fidelity restored image.

2. The method for recoverable visual encryption of human privacy areas in images acquired by a drone according to claim 1, characterized in that: In step one, the target detection module can be any one of YOLO11n, Faster R-CNN, or DETR based on Transformer. During the training phase, all network parameters are frozen to retain its target localization capability.

3. The method for recoverable visual encryption of human privacy areas in images acquired by a drone according to claim 1, characterized in that: The adaptive pixelation visual transformation in step three specifically includes: S3.

1. Dynamically calculate the pixelated block size k based on the width h and height w of the target region: ； in d is the preset minimum block size, and d is the preset scale factor. This indicates the floor function; S3.

2. Perform average pooling on the target region, then upsample to restore it to its original size, creating a pixelated effect. The pixelated target region image is then processed. Compared with the original target region image Through learnable blending parameters The images are then fused to generate the final encrypted region image. : ； in, The value range is constrained to the (0,1) interval, which is used to control the intensity of the visual transformation.

4. The method for recoverable visual encryption of human privacy areas in images acquired by a drone according to claim 1, characterized in that: The encoder of the visual decryption module contains four convolutional blocks, each consisting of two 3×3 convolutional layers, a BatchNorm layer, and a ReLU activation function, and uses 2×2 max pooling downsampling; the decoder contains four convolutional blocks and uses 2×2 transposed convolution upsampling; the channel attention module is embedded after the network bottleneck layer and each stage of the decoder.

Images