A desensitization method and system for crowdsourcing high-precision maps based on deep learning

Abstract

The present invention discloses a method and system for desensitizing map data collected through crowdsourcing based on deep learning. The present invention constructs a "High-precision Map Desensitization Rules and Target Sensitivity Level Table" to classify desensitized targets and stipulates desensitization methods, including shape class, privacy class and text class. For shape-type and privacy-type sensitive targets in high-precision maps, the position of the sensitive target is detected using a sensitive target detection algorithm based on FasterRCNN, and different desensitization methods are used according to the sensitivity level of the sensitive target. For text-type sensitive targets in high-precision maps, the natural scene text detection model I3CL is used to detect all text areas in the map, the text recognition model SVTR is used to identify the text area to obtain the text content of the text area, the sensitive text in the text area is matched by the sensitive word matching module, the sensitive area removal model based on LaMa is used for the sensitive text area, and the sensitive text is erased from the image.

Description

Technical Field

[0001] The present invention belongs to the field of information security, and specifically relates to a method and system for desensitizing high-precision map data collected by crowdsourcing based on deep learning, which is mainly used for desensitizing high-precision map images. Background Art

[0002] High-definition maps (HD maps) are a crucial component of autonomous driving systems and play a vital role in autonomous driving technology. Currently, HD maps are at a critical stage of technological development. However, during the HD map acquisition and production process, HD map desensitization primarily relies on manual desensitization, which is repetitive and costly and impacts HD map production efficiency. Compared to manual desensitization, automated HD map desensitization using deep learning can not only improve production efficiency but also significantly reduce HD map production costs.

[0003] High-precision maps contain a large amount of sensitive information, such as the height and width limits, clearance, load capacity, and slope of important bridges; as well as private information such as faces and license plates, which could be considered public privacy. Direct use of this sensitive information without proper processing can easily lead to leaks.

[0004] High-precision map images are complex and contain a wide variety of sensitive objects. Existing desensitization efforts for HD maps rely on manual desensitization, with related research focusing on desensitizing private information such as faces and license plates. However, sensitive units and facilities, as well as sensitive attribute information in HD maps, contain even more sensitive information and require desensitization. These sensitive objects have distinct characteristics. Faces, license plates, and sensitive units and facilities have distinct physical features, while sensitive text within sensitive attribute information describes these attributes. Therefore, detecting and identifying sensitive objects in HD maps is difficult using a single deep learning algorithm. It requires the coordinated use of multiple algorithms, taking into account the characteristics of the sensitive objects. How to leverage deep learning technology to efficiently and accurately detect and identify sensitive information in HD maps, desensitize this sensitive information, and achieve automated desensitization for HD maps remains an unresolved challenge. Summary of the Invention

[0005] The present invention aims to propose a crowdsourcing-collected high-precision map desensitization method based on deep learning. By constructing a "High-Precision Map Desensitization Rules and Target Sensitivity Level Table" to classify the desensitized targets, the sensitive information in the high-precision map is detected and identified through the collaborative processing of multiple deep learning algorithms. Different processing methods are adopted for sensitive targets of different levels to achieve automated desensitization of high-precision maps.

[0006] The technical solution adopted by the present invention is: constructing a "High-precision Map Desensitization Rules and Target Sensitivity Level Table" to classify desensitized targets and stipulate desensitization methods. Specifically, sensitive targets are divided into high sensitivity and low sensitivity levels, and sensitive targets are divided into three categories, namely shape, privacy, and text. Among them, the privacy category belongs to the low sensitivity level, and the shape and text categories belong to the high sensitivity level. For shape and privacy sensitive targets in high-precision maps, the location of sensitive targets is detected using a sensitive target detection algorithm based on FasterRCNN, and different desensitization methods are used according to the sensitivity level of the sensitive targets. For text-type sensitive targets in high-precision maps, the natural scene text detection model I3CL is used to detect all text areas in the map, and the text recognition model SVTR is used to identify the text area to obtain the text content of the text area. The sensitive text in the text area is matched by the sensitive word matching module. The sensitive area removal model based on LaMa is used for the sensitive text area, and the sensitive text is erased from the image.

[0007] The present invention provides a crowdsourcing high-precision map desensitization method based on deep learning, and the specific implementation steps are as follows:

[0008] (1) Construct a high-precision map desensitization rule and target sensitivity level table, and set the category, sensitivity level, and desensitization rule of sensitive targets;

[0009] (2) Use the natural scene text detection model to locate the text area in the HD map. The detection model will output the coordinates of the minimum enclosing curve of the text area in the HD map;

[0010] (3) Use the text recognition network to identify the text region and obtain the text content in each text instance;

[0011] (4) Collect sensitive words that reflect sensitive attribute information to build a sensitive word library, store it in a tree structure, use AC automaton to match sensitive words, detect whether the identified text area contains sensitive information, and record the coordinates of the text area containing sensitive words;

[0012] (5) Desensitization is performed on the detected sensitive text areas using an image completion model;

[0013] (6) Use the FasterRCNN-based sensitive target detection algorithm to determine the category and location of sensitive targets in high-precision maps;

[0014] (7) According to the category of the detected sensitive target, the sensitivity level corresponding to the target is matched in the high-precision map desensitization rules and the target sensitivity level table;

[0015] (8) Use different desensitization methods according to different sensitivity levels.

[0016] Furthermore, in step (1), sensitive targets are divided into privacy, shape, and text categories; the sensitivity levels are divided into high sensitivity and low sensitivity, with privacy targets belonging to the low sensitivity level and shape and text targets belonging to the high sensitivity level; the desensitization rule for the low sensitivity level is to directly encrypt the sensitive area, and the desensitization rule for the high sensitivity level is to erase significant shapes and logos and keep them consistent with the surrounding environment.

[0017] Furthermore, in step (2), the natural scene text detection model I3CL is used to locate the text area in the high-precision map, specifically including:

[0018] High-precision map images are input into a convolutional neural network combined with a feature pyramid network (FPN) to extract multi-scale image features. The feature maps of each scale are then refined through the intra-text collaborative learning module to enhance the information exchange between the gap areas between text instances and the text. The region candidate network (RPN) is then used to select text candidate boxes for text bounding box regression and text instance segmentation. The bounding box regression branch performs the text box boundary regression task and predicts the position of the text box. The inter-text collaborative learning module is introduced into the text instance segmentation branch to learn the mutual dependencies between text instances and generate detailed text contours.

[0019] Furthermore, the processing process of the in-text collaborative learning module is as follows: first, the input is divided into three branches, each branch includes three layers, and each layer contains three parallel convolution layers, namely 1×N horizontal convolution, N×1 vertical convolution and a normal convolution. The features extracted in each layer are fused and then input into the next layer. Finally, the features of the three branches are fused and then passed through a normal convolution layer. After fusion with the original input features, they are passed through a ReLU activation function;

[0020] The inter-text collaborative learning module adopts the TransFormer structure. It first inputs the features into a convolutional layer to reduce the channel dimension. Then, the feature dimension is reduced through adaptive maximum pooling. Next, each feature is flattened into a vector. The resulting feature sequence is input into a TransFormer encoder with three regular encoder layers and four self-attention layers.

[0021] Furthermore, in step (3), the text region is identified using the text recognition network SVTR, which specifically includes:

[0022] Step 3.1: Input a text image of size H×W×3, where H and W are the height and width of the image, and 3 represents the three RGB color channels. The text image is decomposed into blocks of size H / 4×W / 4 with a depth of D0 through the patch embedding module.

[0023] In step 3.2, three stages of feature extraction are performed at different scales, which is ultimately simplified to a 1×W / 4×D3 feature representation. Each stage consists of a global mixing block, a local mixing block, and a merge or combination operation, where D3 is the depth value.

[0024] In step 3.3, the final feature representation is linearly predicted to generate a character sequence, where components of the same character are transcribed as repeated characters and non-text components are transcribed as blank symbols.

[0025] Furthermore, the specific implementation of step (4) includes:

[0026] Step 4.1: Build a Trie tree. Build a Trie tree of all sensitive words. Each node represents a character, and the path from the root node to the leaf node represents a sensitive word. At the same time, each node stores a flag indicating whether the string ending with the node is a sensitive word.

[0027] Step 4.2, build AC automaton, convert the Trie tree into AC automaton, add a failure pointer to each node in the Trie tree, pointing to another node on the Trie tree, indicating which node to jump to to continue matching when the match fails;

[0028] Step 4.3, matching text. For each recognized text, start traversing from the root node of the AC automaton; if the current character matches successfully, traverse down to the next node; if the match fails, jump to another node according to the failure pointer to continue matching; when the flag bit of a node is true, it means that a sensitive word has been found, and the location of this text area is recorded.

[0029] Furthermore, the detected sensitive text areas are desensitized using the image completion model LaMa, which specifically includes:

[0030] Step 5.1: Generate a mask using the detected sensitive areas and input the original image and mask into the image completion model;

[0031] In step 5.2, the completion model first performs downsampling for feature extraction;

[0032] In step 5.3, the input features are divided into two branches. The local branch uses the traditional convolution layer to extract local information, and the global branch uses fast Fourier convolution to extract features in both the spatial and frequency domains to achieve feature extraction of the global receptive field. Finally, the local information and the global information are cross-fused.

[0033] In step 5.4, the fused features are upsampled and the desensitized image is output.

[0034] Furthermore, in step (6), the specific implementation of using the sensitive target detection algorithm based on FasterRCNN to locate the category and position of sensitive targets in the high-precision map includes: inputting the image to be desensitized into the convolutional neural network backbone combined with the FPN network structure to extract the multi-scale feature pyramid for subsequent sensitive target detection and classification; the regional candidate network RPN combines the multi-scale feature pyramid prediction to generate a large number of rectangular sensitive target candidate boxes; performing target box border regression and classification to determine the category and position of the sensitive target.

[0035] Furthermore, in step (8), different desensitization methods are used according to different sensitivity levels, specifically including: for privacy-related targets with low sensitivity levels, scrambling encryption is used for the privacy area according to the coordinates corresponding to the privacy targets; for shape-related targets with high sensitivity levels, the image completion model LaMa is used to desensitize the sensitive area according to the coordinates corresponding to the shape-related targets;

[0036] The scrambling encryption for low-sensitivity privacy targets specifically includes: direct encryption of the privacy area according to the coordinates corresponding to the privacy target. Its function is to disrupt the original image structure by changing the position and order of pixels in the image, thereby hiding the image content of the privacy area;

[0037] The LaMa desensitization processing of the image completion model for highly sensitive shape-type targets specifically includes: generating a mask image based on the coordinates corresponding to the shape-type target and the original image; inputting the mask image and the original image into the LaMa algorithm, and regenerating sensitive units and facilities with the help of non-sensitive area image features, and the original sensitive area is replaced with an environment similar to the surrounding area.

[0038] The present invention also provides a crowdsourcing high-precision map desensitization system based on deep learning, comprising:

[0039] one or more processors;

[0040] A storage device is used to store one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in the above solution.

[0041] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows:

[0042] The present invention provides a method for desensitizing crowdsourced map data based on deep learning, which is mainly used for desensitizing real-life images. In view of the fact that there are many types of sensitive targets and complex scenes in crowdsourced data, there are currently no corresponding standards for desensitizing crowdsourced data, and the desensitization detection has a low degree of automation and low detection accuracy. Therefore, a "High-Precision Map Desensitization Rules and Target Sensitivity Level Table" is constructed, which stipulates the categories and sensitivity levels of the targets to be desensitized, and divides the sensitive information collected in the image data by crowdsourcing into privacy, shape, and text categories. The privacy category contains privacy information involving public privacy, the shape category contains sensitive units and facilities with obvious appearance characteristics, and the text category contains sensitive attribute information. This paper proposes a deep learning-based desensitization method for crowdsourced data. To address the complex backgrounds and variable scales of text-sensitive objects, a text detection and recognition model based on I3CL and SVTR is designed. The I3CL text detection algorithm is used to detect text in real-life images, while the SVTR text recognition algorithm is used to recognize text in real-life images. An AC automaton is used to construct an index tree for the sensitive vocabulary. The identified sensitive words are then text-matched, sensitive text areas are located, and the Lama algorithm is used to erase these areas to align them with the surrounding environment, thereby achieving intelligent desensitization. A FasterRCNN-based sensitive object detection model and a Lama-based sensitive area removal model are designed for privacy-sensitive and shape-sensitive objects. For privacy-sensitive objects, an encryption algorithm is used to encrypt the privacy area, while for shape-sensitive objects, Lama is used to erase the sensitive area, improving the accuracy and efficiency of crowdsourced data desensitization. Therefore, the present invention can achieve automated detection and non-invasive desensitization of sensitive objects. Furthermore, the present invention can adopt different desensitization methods for different sensitive objects, achieving a balance between performance and security. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 1 is a general schematic diagram of high-precision map desensitization according to an embodiment of the present invention;

[0044] Figure 2 This is a diagram of the I3CL structure in an embodiment of the present invention;

[0045] Figure 3 This is a structural diagram of the collaborative learning module in the text according to an embodiment of the present invention;

[0046] Figure 4 4 is a structural diagram of SVTR in an embodiment of the present invention.

[0047] Figure 5 This is a structural diagram of an AC automaton in an embodiment of the present invention;

[0048] Figure 6 1 is a LaMa structure diagram in an embodiment of the present invention.

[0049] Figure 7 This is the desensitization effect of privacy-sensitive targets in an embodiment of the present invention, where (a) and (b) are desensitization effects of different privacy-sensitive targets respectively.

[0050] Figure 8 This is the desensitization effect of shape-sensitive targets in an embodiment of the present invention, where (a)-(g) are desensitization effects of different shape-sensitive targets respectively.

[0051] Figure 9 This is the desensitization effect of text-sensitive targets in the embodiment of the present invention. DETAILED DESCRIPTION

[0052] The specific implementation process of the present invention will be described in detail below with reference to the accompanying drawings.

[0053] like Figure 1 As shown, an embodiment of the present invention provides a method for desensitizing high-precision maps collected by crowdsourcing based on deep learning, comprising the following steps:

[0054] Step 1: Construct a "High-Precision Map Desensitization Rules and Target Sensitivity Level Table," which specifies the categories, sensitivity levels, and desensitization rules for sensitive targets. Sensitive targets are categorized as privacy, shape, and text. Privacy targets include faces and license plates that implicate public privacy; shape targets include units and facilities related to public safety, as well as civilian facilities that significantly impact production and daily life; and text targets include feature identification areas in the side or distant views of various units and facilities, containing sensitive text. Sensitivity levels are categorized as high or low, with privacy targets belonging to the low sensitivity level and shape and text targets belonging to the high sensitivity level. The desensitization rule for the low sensitivity level involves direct encryption of sensitive areas, while the desensitization rule for the high sensitivity level involves erasing prominent shapes and identification areas to maintain consistency with the surrounding environment.

[0055] Step 2: For the text area in the crowdsourced high-precision map, use the natural scene text detection model I3CL to locate the text area in the image. The overall framework of the I3CL model is as follows: Figure 2 As shown in the figure. First, the image is input into a backbone network with an FPN architecture to generate a multi-scale feature pyramid. Second, the inter-text collaborative learning module is used to refine features at each scale of the feature pyramid. Next, a region proposal network is used to generate text proposals for subsequent processes. Bounding box regression and mask prediction are then performed in two parallel branches. The frame branch further refines and classifies the text proposals. In the mask branch, the intra-text collaborative learning module is used to collaboratively learn across all positive text instances. The detection model outputs the coordinates of the minimum bounding curve of the text region in the HD map.

[0056] Step 2.1: The high-precision map image is input into the convolutional neural network combined with the feature pyramid network (FPN) to extract multi-scale image features.

[0057] Step 2.2, the structure of the collaborative learning module within the text is as follows Figure 3 As shown in the figure, the intra-text collaborative learning module refines the feature maps of each feature map. The processing process is as follows: first, the input is divided into three branches, each branch includes three layers, and each layer contains three parallel convolution layers, namely 1×N horizontal convolution, N×1 vertical convolution and an ordinary convolution. The features extracted in each layer are fused and then input into the next layer. Finally, the features of the last three branches are fused and then passed through an ordinary convolution layer, and then fused with the original input features and passed through a ReLU activation function. This method includes a set of paths with multiple receptive fields, which enhances the information exchange between the gap areas between text instances and between texts.

[0058] In step 2.3, the region proposal network (RPN) selects text candidate boxes for text bounding box regression and text instance segmentation.

[0059] In step 2.4, the bounding box regression branch performs the text box boundary regression task to predict the location of the text box.

[0060] In step 2.5, the inter-text collaborative learning module is introduced into the text instance segmentation branch. The inter-text collaborative learning module adopts the Transformer structure. The features are first input into a convolutional layer to reduce the channel dimension. Then, the feature dimension is reduced through adaptive maximum pooling. Next, each feature is flattened into a vector. The resulting feature sequence is input into a Transformer encoder, which has three regular encoder layers and four self-attention layers. The self-attention layer can capture the long-range dependencies between different text instances in the image by adaptively focusing on specific text instances with similar background context or font appearance. Through this collaborative learning method, the representation ability of the learned features can be improved. The mutual dependencies between text instances are learned to generate detailed text outlines.

[0061] Step 3: The text recognition network SVTR identifies the text area. The structure of SVTR is as follows: Figure 4 As shown, the network converts the input image text into character components through progressive overlapping patch embedding, where each character component is associated with a portion of the text character in the image. It then passes through three hybrid blocks for merging or combining operations, extracting features at different scales. Finally, a linear predictor is used to obtain the text content of each text instance.

[0062] In step 3.1, the input text image is downsampled through the convolution module to obtain a larger receptive field and improve the ability to express local information.

[0063] In step 3.2, three stages of feature extraction are performed at different scales. Each stage consists of a global mixing block, a local mixing block, and a merging or combining operation. The global mixing block can establish long-term dependencies between components from different characters and weaken the influence of non-text components while increasing the importance of text components. The local mixing block uses a window to constrain character components.

[0064] In step 3.3, the input to stage 1 is an image of size H×W×3, where H and W are the height and width of the image, and 3 represents the three RGB color channels. In the patch embedding module, the image is decomposed into small character components, which may only contain part of a character. At the beginning of each stage, the image is converted into a block of size H / 4×W / 4 with a depth of D0. In the fusion module, a series of mixing blocks capture patterns within and between characters. This includes local and global mixing. The local mixing block is responsible for extracting local patterns similar to strokes, and the global mixing block is responsible for capturing long-range dependencies between different characters. The final fusion operation further processes the features to help integrate feature information at different levels.

[0065] In step 3.4, the input of stage 2 is the feature map of H / 4×W / 4×D0 output by the previous stage, which also includes the fusion module and fusion operation. In this stage, the feature map is further refined to H / 8×W / 4×D1, where D1 is usually deeper than D0 to capture more complex features.

[0066] In step 3.5, the input to stage 3 is H / 8 × W / 4 × D1. The feature maps of the final fusion module and fusion operation are further compressed to H / 16 × W / 4 × D2, and then simplified to a feature representation of 1 × W / 4 × D3. The final feature representation is linearly predicted to generate a character sequence, where components with the same character are transcribed as repeated characters and non-text components are transcribed as whitespace.

[0067] Step 4: Collect sensitive words that reflect sensitive attribute information to build a sensitive word library, store it in a tree structure, and use AC automaton to match sensitive words. The structure of AC automaton is as follows: Figure 5 As shown, the AC automaton is a multi-pattern string matching algorithm based on a Trie tree. The AC automaton consists of two main steps: construction and search. During the construction phase, it first creates a Trie tree to store all pattern strings. It then converts the Trie tree into a finite state automaton by constructing a failure pointer (also called a mismatch pointer). During the search phase, it searches for pattern strings within the text by reading the text character by character and transitioning states. This method can quickly detect whether the identified text region contains sensitive information and record the coordinates of the text region containing sensitive words.

[0068] Step 4.1: Build a Trie tree. This tree represents all sensitive words. Each node represents a character, and the path from the root node to a leaf node represents a sensitive word. Each node can also store a flag indicating whether the string ending with that node is a sensitive word.

[0069] Step 4.2, build AC automaton, convert the Trie tree into AC automaton, add a failure pointer to each node in the Trie tree, pointing to another node on the Trie tree, indicating which node should be jumped to continue matching when the matching fails.

[0070] Step 4.3: Matching Text. For each recognized text, traverse the text starting from the root node of the AC automaton. If the current character matches successfully, traverse down to the next node. If the match fails, jump to another node based on the failure pointer to continue matching. When the flag bit of a node is true, it indicates that a sensitive word has been found, and the location of this text area is recorded.

[0071] In step 5, the image completion model LaMa is applied to the detected sensitive text areas. The LaMa network input consists of an original RGB color image (3 channels) and a binary mask image (1 channel). The mask image marks the area to be inpainted, typically in white, and the rest in black. The mask is multiplied with the original image, a step known as "masking" or "covering," resulting in an image with the missing area. The mask and the masked image are stacked together to form a 4-channel input tensor, allowing the network to utilize both image information and mask position information. The LaMa network incorporates a fast Fourier convolution module, which has a global receptive field and can capture long-range dependencies in the image, facilitating the inpainting of large missing areas. The input first passes through a downsampling layer to reduce the spatial dimensionality for subsequent processing. The data flow is then split into two branches: a local branch and a global branch. The local branch uses traditional convolutional layers to extract local features. The global branch utilizes a fast Fourier convolution module to extract global context information. In this branch, the image data is converted to the frequency domain, convolution is performed, and then the inverse transformation is returned to the spatial domain. Finally, the information from the local and global branches is fused and upsampled, gradually restoring the original image's resolution to the original image's original resolution. A mask is generated using the detected sensitive regions. The original image and mask are then fed into the image completion model, which uses the non-sensitive regions in the image as known information to regenerate the sensitive regions, replacing the original sensitive regions.

[0072] Step 5.1: Generate a mask using the detected sensitive areas and input the original image and mask into the image completion model.

[0073] In step 5.2, the network first performs downsampling for feature extraction.

[0074] In step 5.3, the input features are divided into two branches. The local branch is responsible for extracting local information, and the global branch uses fast Fourier convolution to extract features in both the spatial domain (image domain) and the frequency domain to achieve feature extraction of the global receptive field. Finally, the local information and the global information are cross-fused.

[0075] In step 5.4, the fused features are upsampled and the desensitized image is output.

[0076] Step 6: For sensitive units and facilities in the crowdsourced high-precision map, use the target detection algorithm FasterRCNN to locate the position of sensitive targets in the image, and the detection model will output the coordinates of the sensitive units and facilities.

[0077] In step 6.1, the image to be desensitized is fed into a convolutional neural network backbone combined with an FPN network structure to extract a multi-scale feature pyramid for subsequent sensitive object detection and classification. A multi-scale feature pyramid is a combination of convolutional feature maps of different scales. It can fuse semantic features from different stages of the convolutional neural network to generate convolutional feature maps of different scales to accommodate sensitive objects of different sizes in the image.

[0078] Step 6.2, in the first stage, the region proposal network RPN combines multi-scale feature pyramid prediction to generate a large number of rectangular text candidate boxes.

[0079] Step 6.3, the second stage, performs target frame regression and classification to determine the category and location of sensitive targets.

[0080] Step 7: According to the category of the detected sensitive target, match the sensitivity level corresponding to the target in the "High-precision Map Desensitization Rules and Target Sensitivity Level Table" and adopt different desensitization methods according to different sensitivity levels.

[0081] Step 8: For privacy-related targets, which are of low sensitivity, scrambling encryption is used on the privacy area according to the coordinates corresponding to the privacy target. Its function is to disrupt the original image structure by changing the position and order of pixels in the image, thereby hiding the image content of the privacy area.

[0082] Step 9: For shape-type targets, which are at a high sensitivity level, the LaMa algorithm is used to complete the sensitive areas according to the coordinates corresponding to the shape-type targets.

[0083] Step 9.1: Generate a mask image based on the coordinates corresponding to the shape class target and the original image.

[0084] In step 9.2, the mask image and the original image are input into the LaMa algorithm. With the help of the image features of the non-sensitive area, the sensitive units and facilities are regenerated, and the original sensitive area is replaced with an environment similar to the surrounding area.

[0085] Figure 7 、 Figure 8 and Figure 9 The system demonstrates the desensitization effects of privacy-sensitive, shape-sensitive, and text-sensitive objects, with the original image shown on the left and the desensitized image on the right. The desensitization effects on faces and license plates are shown for privacy-sensitive objects. The desensitization effects on shape-sensitive objects include bridge height and weight limit signs, chimneys, condensation towers, high-voltage power towers, and meteorological towers. The desensitization effects on text-sensitive objects include archives and other important facilities containing sensitive text.

[0086] The embodiment of the present invention also provides a crowdsourcing high-precision map desensitization system based on deep learning, including:

[0087] one or more processors;

[0088] A storage device is used to store one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in the above solution.

[0089] The above content is a further detailed description of the present invention in conjunction with the preferred embodiment, and the specific implementation of the present invention should not be considered to be limited to these descriptions. Those skilled in the art should understand that various modifications can be made to the details without departing from the scope of the appended claims, and all should be considered to fall within the scope of protection of the present invention.

Claims

1. A crowdsourcing high-precision map desensitization method based on deep learning, characterized by: The following steps are involved: Step (1) constructs a high-precision map desensitization rule and target sensitivity level table, and sets the category, sensitivity level, and desensitization rule of the sensitive target; In step (1), sensitive targets are divided into privacy, shape, and text categories; the sensitivity levels are divided into high sensitivity and low sensitivity, with privacy targets belonging to the low sensitivity level and shape and text targets belonging to the high sensitivity level; the desensitization rule for the low sensitivity level is to directly encrypt the sensitive area, and the desensitization rule for the high sensitivity level is to erase the significant shapes and logos and keep them consistent with the surrounding environment; Step (2): Use the natural scene text detection model to locate the text area in the high-precision map. The natural scene text detection model will output the coordinates of the minimum enclosing curve of the text area in the high-precision map; Step (3): Use the text recognition network to identify the text region and obtain the text content in each text instance; Step (4) collects sensitive words that reflect sensitive attribute information to build a sensitive word library, and stores it in a tree structure. Use AC automaton to match sensitive words, detect whether the identified text area contains sensitive information, and record the coordinates of the text area containing sensitive words; Step (5): Desensitizing the detected sensitive text areas using an image completion model; Step (6), use the sensitive target detection algorithm based on FasterRCNN to determine the category and location of sensitive targets in the high-precision map; Step (7), according to the category of the detected sensitive target, match the sensitivity level corresponding to the sensitive target in the high-precision map desensitization rules and target sensitivity level table; In step (8), different desensitization methods are used according to different sensitivity levels.

2. The method for desensitizing high-precision maps collected through crowdsourcing based on deep learning according to claim 1 is characterized by: In step (2), the natural scene text detection model I3CL is used to locate the text area in the high-precision map, specifically including: High-precision map images are input into a convolutional neural network combined with a feature pyramid network (FPN) to extract multi-scale image features. The intra-text collaborative learning module is then used to refine the feature maps at each scale to enhance the information exchange between the gap areas between text instances and the text. The region candidate network (RPN) is then used to select text candidate boxes for text bounding box regression and text instance segmentation. The bounding box regression branch performs the text box boundary regression task and predicts the position of the text box. The inter-text collaborative learning module is introduced into the text instance segmentation branch to learn the interdependence between text instances and generate detailed text contours.

3. The method for desensitizing high-precision maps collected through crowdsourcing based on deep learning according to claim 2 is characterized by: The processing process of the in-text collaborative learning module is as follows: first, the input is divided into three branches, each branch consists of three layers, and each layer contains three parallel convolutional layers, namely 1×N horizontal convolution, N×1 vertical convolution and a normal convolution. The features extracted in each layer are fused and then input into the next layer. Finally, the features of the three branches are fused and then passed through a normal convolution layer. After fusion with the original input features, they are activated by a ReLU function. The inter-text collaborative learning module adopts the TransFormer structure. It first inputs the features into a convolutional layer to reduce the channel dimension; then, it reduces the feature dimension through adaptive maximum pooling. Next, each feature is flattened into a vector; the resulting feature sequence is input into a TransFormer encoder with three regular encoder layers and four self-attention layers.

4. The method for desensitizing high-precision maps collected through crowdsourcing based on deep learning according to claim 1 is characterized in that: In step (3), the text region is identified using the text recognition network SVTR, which specifically includes: Step 3.1: Input a text image of size H×W×3, where H and W are the height and width of the image, and 3 represents the three RGB color channels. The text image is decomposed into Blocks of size , depth ; Step 3.2, three stages of feature extraction are performed at different scales, which is finally simplified to Each stage consists of a global mixing block, a local mixing block, and a merging or combining operation. is the depth value; In step 3.3, the final feature representation is linearly predicted to generate a character sequence, where components of the same character are transcribed as repeated characters and non-text components are transcribed as blank symbols.

5. The method for desensitizing high-precision maps collected through crowdsourcing based on deep learning according to claim 1 is characterized in that: The specific implementation of step (4) includes: Step 4.1: Build a Trie tree. Build a Trie tree of all sensitive words. Each node represents a character, and the path from the root node to the leaf node represents a sensitive word. At the same time, each node stores a flag indicating whether the string ending with the node is a sensitive word. Step 4.2, build AC automaton, convert the Trie tree into AC automaton, add a failure pointer to each node in the Trie tree, pointing to another node on the Trie tree, indicating which node to jump to to continue matching when the match fails; Step 4.3, matching text. For each recognized text, start traversing from the root node of the AC automaton; if the current character matches successfully, traverse down to the next node; if the match fails, jump to another node according to the failure pointer to continue matching; when the flag bit of a node is true, it means that a sensitive word has been found, and the location of this text area is recorded.

6. The method for desensitizing high-precision maps collected through crowdsourcing based on deep learning according to claim 1 is characterized by: The image completion model LaMa is used to perform desensitization processing on the detected sensitive text areas, including: Step 5.1: Generate a mask using the detected sensitive areas and input the original image and mask into the image completion model; In step 5.2, the image completion model first performs downsampling for feature extraction; In step 5.3, the input features are divided into two branches. The local branch uses traditional convolutional layers to extract local information, while the global branch uses fast Fourier convolution to extract features in both the spatial and frequency domains to achieve feature extraction of the global receptive field. Finally, the local information and the global information are cross-fused. In step 5.4, the fused features are upsampled and the desensitized image is output.

7. The method for desensitizing high-precision maps collected through crowdsourcing based on deep learning according to claim 1 is characterized by: In step (6), the specific implementation of using the sensitive target detection algorithm based on FasterRCNN to locate the category and position of sensitive targets in the high-precision map includes: inputting the image to be desensitized into the convolutional neural network backbone combined with the FPN network structure to extract the multi-scale feature pyramid for subsequent sensitive target detection and classification; the region candidate network RPN combines the multi-scale feature pyramid prediction to generate a large number of rectangular sensitive target candidate boxes; performing target box border regression and classification to determine the category and position of the sensitive target.

8. The method for desensitizing high-precision maps collected through crowdsourcing based on deep learning according to claim 1 is characterized by: In step (8), different desensitization methods are used according to different sensitivity levels, including: for privacy-related targets with low sensitivity levels, scrambling encryption is used for the privacy area according to the coordinates corresponding to the privacy target; for shape-related targets with high sensitivity levels, the image completion model LaMa is used to desensitize the sensitive area according to the coordinates corresponding to the shape-related targets; The scrambling encryption for low-sensitivity privacy targets specifically includes: direct encryption of the privacy area according to the coordinates corresponding to the privacy target. Its function is to disrupt the original image structure by changing the position and order of pixels in the image, thereby hiding the image content of the privacy area; The LaMa desensitization processing of the image completion model for highly sensitive shape-type targets specifically includes: generating a mask image based on the coordinates corresponding to the shape-type target and the original image; inputting the mask image and the original image into the LaMa algorithm, and regenerating sensitive units and facilities with the help of non-sensitive area image features, and the original sensitive area is replaced with an environment similar to the surrounding area.

9. A crowdsourcing high-precision map desensitization system based on deep learning, characterized by: include: one or more processors; A storage device for storing one or more programs, which, when executed by the one or more processors, causes the one or more processors to implement the method according to any one of claims 1 to 8.

Images