Cadastral feature automatic extraction and recognition system based on deep learning

Abstract

This invention relates to an automatic cadastral element extraction and recognition system based on deep learning, specifically in the fields of artificial intelligence and deep learning. It restores the topological structure of handwriting through deep restoration technology, ensuring the clarity and integrity of the input data. Leveraging the feature extraction capabilities of a hybrid architecture, it accurately captures the subtle textures and global spatial relationships of handwritten strokes, generating highly discriminative feature representations. Combined with reinforcement learning strategies, the solution transforms cadastral business logic into an intelligent verification mechanism, performing logical reasoning and automatic correction on the recognition results, significantly reducing the coding error rate. Furthermore, with the aid of a meta-learning framework, the solution possesses strong cross-domain adaptability, requiring only a few labeled samples to quickly migrate to different regions. This significantly reduces the cost of manual review while achieving accurate and efficient conversion of cadastral elements from images to structured data.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and deep learning, and more specifically, to a deep learning-based automatic extraction and recognition system for cadastral elements. Background Technology

[0002] Paper-based land registration maps, having been stored for decades, generally exhibit yellowing and brittleness, as well as ink smudging and diffusion. Hand-drawn boundary markers often use special symbols, and land parcel numbers are handwritten by staff, resulting in issues such as cursive writing and simplified characters. When grassroots real estate registration centers undertake the task of digitizing historical maps, they are limited by equipment conditions and often use ordinary scanners with a resolution of no more than 300 dpi to collect images. Furthermore, uneven lighting in the scanning environment leads to dense image noise and loss of details. There is a need to quickly convert millions of historical paper-based land registration maps into structured digital data for database storage, but traditional processing methods are insufficient to handle the above complexities.

[0003] While existing general optical character recognition systems achieve high recognition rates for printed text, they are not trained for handwritten features specific to cadastral scenarios. Their accuracy in recognizing cursive administrative division abbreviations and illegible serial numbers is less than 65%. Map symbol recognition models can handle standard topographic symbols but cannot distinguish cadastral-specific boundary markers. Furthermore, existing convolutional neural network-based map feature extraction methods do not incorporate cadastral semantic constraints. For example, parcel numbers must follow fixed coding rules for province, city, county, township, village, group, and serial number, leading to logical errors in the recognition results. Simultaneously, historical handwritten samples are scarce and lack unified annotation standards. While small-sample learning methods have applications in fields such as medical imaging, they are not yet adapted to the specific needs of cadastral scenarios. In summary, existing methods cannot achieve handwritten feature extraction and end-to-end optimization that integrates cadastral semantic constraints, resulting in insufficient accuracy in recognizing handwritten annotations and special symbols on historical paper cadastral maps. This fails to meet the efficiency and accuracy requirements for batch registration of rural real estate, becoming a core obstacle restricting the digitization of cadastral data. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing an automatic cadastral element extraction and recognition system based on deep learning. The system utilizes a repair module, a feature extraction module, a semantic error correction module, and an adaptive optimization module to solve the problems mentioned in the background.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: specifically including: Repair module: In response to the acquired original paper-based book map scan, the module performs ink smudge repair and paper wrinkle correction on the original paper-based book map scan using a conditional generative adversarial network, and uses the ink connectivity loss function to constrain the stroke topological continuity in order to generate a standardized repair image. Feature extraction module: Receives the standardized repaired image, extracts the local texture features and global topological features of the handwritten strokes through a hybrid architecture of convolutional neural network and transformer, and generates the corresponding stroke topology map feature vector; Semantic error correction module: Receives the feature vector of the stroke topology map, transforms the preset cadastral coding rules into a reward function based on the reinforcement learning strategy, performs logical verification and automatic correction on the preliminary recognition results generated by the pre-trained OCR head, and outputs the error-corrected recognition results that conform to the preset administrative division coding format; Adaptive optimization module: Receives the recognition results after error correction, and based on the model-independent meta-learning framework, uses a small number of labeled samples from the target region to quickly fine-tune the parameters of the repair module, feature extraction module, and semantic error correction module, and outputs a final recognition model adapted to the handwriting style of the target region.

[0006] In a preferred embodiment, the preprocessing of the original paper-based book map scan in the repair module is as follows: First, the original paper-based document scans are normalized in resolution and scaled to a set pixel size. The color images are then converted into single-channel grayscale images. Edge features are extracted from the grayscale images using an edge detection algorithm to generate a binarized edge feature map. This edge feature map is then used as a constraint input to the generator of a conditional generative adversarial network. Subsequently, the generator adopts an encoder-decoder architecture. The encoder part extracts multi-scale features of the grayscale image through multiple residual blocks and downsamples it to a low resolution. The decoder part fuses the features extracted by the encoder part with the upsampling results through skip connections and outputs a preliminary repaired image. A spatial attention mask is introduced at the end of the decoder to enhance the ink smudged area with weights, resulting in a restored image. The ink connectivity loss function is used to calculate the difference in the distribution of connected components between the restored image and the pre-stored standard clear cadastral map template.

[0007] In a preferred embodiment, the discriminator in the repair module adopts a block discriminant structure to verify the authenticity of the repair result; The discriminator receives the restored image and its corresponding edge feature map as input and outputs a local realism probability map in the form of a two-dimensional matrix. Each element in the two-dimensional matrix represents the clarity of the corresponding local image patch in the input image. During training, the restoration module forces the generator to generate restoration results that are closer to the texture of the real cadastral map by minimizing the cross-entropy loss of the discriminator. Meanwhile, the ink connectivity loss function measures the topological structure of the stroke by counting the number of eight-neighbor connected regions. When isolated noise appears in the restored image, the value of this loss function increases, thereby guiding the generator to maintain the inherent geometric structure and topological continuity of cadastral elements while restoring the ink color, ensuring that the standardized restored image output to the feature extraction module has high availability.

[0008] In a preferred embodiment, the process of extracting local texture features and global topological features of handwritten strokes in the feature extraction module is as follows: First, the standardized repaired image is received and input into the backbone of the convolutional neural network. It then passes through the initial convolutional layer, batch normalization layer, activation function layer, max pooling layer, and four residual blocks in sequence; the output is a local feature map. Next, the local feature map is flattened into a sequence, and sine and cosine position encoding is added to preserve the spatial position information of the strokes. This sequence is then input into the transformer encoder in the hybrid architecture of convolutional neural network and transformer, where a dual spatial and channel attention mechanism is introduced. Spatial attention calculates attention weights by multiplying the query matrix, key matrix, and value matrix, and then uses the calculated attention weights to perform a weighted summation of the value matrix to output spatial weighted features. Channel attention compresses spatially weighted features into channel description vectors through global average pooling, then generates channel weights through a fully connected layer and activation function. The spatially weighted features and channel weights are multiplied element-wise to enhance the channel dimension, and the enhanced global topological feature sequence is output.

[0009] In a preferred embodiment, the specific process of generating the feature vector of the stroke topology graph is as follows: A feature response threshold is set for the enhanced global topological feature sequence to filter out the set of stroke endpoints. The Euclidean distance between any two endpoints is calculated. If the distance is less than the preset endpoint spacing threshold, an edge is added to connect the two endpoints, thereby constructing a stroke topological graph. Next, a two-layer graph convolutional network is used to process the stroke topology graph. The first layer performs a linear transformation on the normalized adjacency matrix and the initial feature matrix of the stroke endpoints, and then inputs the transformation into the activation function to output intermediate features. The normalized adjacency matrix is ​​calculated by adding the adjacency matrix and the identity matrix, and then multiplying the left side by the inverse square root of the degree matrix and the right side by the inverse square root of the degree matrix, respectively. The second layer performs a linear transformation on the normalized adjacency matrix and intermediate features to output higher-order features; finally, global average pooling is performed on the higher-order features to generate stroke topology graph feature vectors.

[0010] In a preferred embodiment, the specific process of performing logical verification and automatic correction in the semantic error correction module is as follows: First, the pre-defined cadastral coding rules are formally modeled and decomposed into three categories: format constraints, regional consistency constraints, and logical rationality constraints. The format constraints stipulate that the code length is 19 characters, consisting of three parts: the first 6 characters are the county-level administrative division code, the middle 6 characters are the cadastral district and cadastral sub-district code, and the last 5 characters are the sequence code, which is a number, and the feature code is 2 uppercase letters; the regional consistency constraint stipulates that the first 6 characters of the target region's code prefix must completely match the preset county-level administrative division code; the logical rationality constraint stipulates that the real estate unit serial number must be greater than or equal to 1 and less than or equal to the current maximum number of real estate units in the cadastral district and sub-district.

[0011] Next, a multi-dimensional reward function is constructed based on the three types of constraints. It consists of a weighted sum of format reward items, regional consistency reward items, and logical rationality reward items. Each reward item returns a positive or negative value based on the evaluation result. The three reward items are multiplied by preset weight coefficients and then summed to generate a total reward value used to evaluate the compliance of the preliminary identification results.

[0012] In a preferred embodiment, the reinforcement learning strategy uses a proximal policy optimization algorithm to train the policy network, taking the feature vector of the stroke topology map and the spliced ​​features of the preliminary recognition result as the input state of the policy network, and the policy network outputs the correction action for each digit of the preliminary recognition result. During training, the objective function is defined as maximizing the expected reward, and a shearing mechanism is introduced to limit the magnitude of policy updates. By calculating the shearing approximation of the probability ratio between the old and new policies and combining it with the advantage function to estimate the relative benefit of actions, the stability of the training process is ensured. The advantage function is calculated using the generalized advantage estimation method and is used to measure the advantage of performing a specified corrective action in a given state compared to the average level. The policy network parameters are updated using the gradient ascent method to maximize the objective function value, thereby learning a policy that can logically correct recognition errors based on stroke topological features.

[0013] In a preferred embodiment, the semantic error correction module selects the optimal action sequence from the correction action distribution output by the policy network through greedy decoding, applies the optimal action sequence to the preliminary recognition result, and generates the error-corrected recognition result. Next, the total reward value of the corrected recognition result is calculated. If the total reward value is greater than the preset compliance threshold, the corrected recognition result is confirmed to be valid and output. If the total reward value does not meet the target, the reasoning process of the policy network will be re-executed until a corrected identification result that meets the compliance requirements is generated. The final corrected recognition result conforms to the preset administrative division coding format, and its logical correctness is constrained by the semantics of the stroke topology feature vector. This ensures that the recognition result conforms to the cadastral business rules while avoiding semantic deviations that may be caused by pure rule correction.

[0014] In a preferred embodiment, the specific process of performing rapid fine-tuning in the adaptive optimization module is as follows: First, we enter the meta-training phase and prepare ten source region datasets, each containing a large number of labeled samples. For each source region task, we divide its source region dataset into a support set and a query set. The support set is used to simulate the subsequent rapid adaptation process, and the query set is used for meta-optimization. The inner loop update is performed, the gradient is calculated based on the loss value on the support set, and the initial parameters are updated using the inner loop learning rate to obtain the task adaptation parameters. In this process, the inner loop update follows the gradient descent principle, that is, the task adaptation parameters are equal to the initial parameters minus the product of the inner loop learning rate and the gradient of the support set loss function. Then, the outer loop loss is calculated, and the loss values ​​of each source region task on the query set are summarized. The initialization parameters are updated based on the outer loop learning rate. Through repeated iterations, initialization parameters with strong generalization ability are obtained, so that the model can adapt to new tasks with only a small number of gradient updates.

[0015] In a preferred embodiment, the adaptive optimization module continues to perform the following process: First, a small set of labeled samples from the target region is received, the initialization parameters obtained during the meta-training phase are loaded, gradient descent is performed, the gradient is calculated using the loss value of the target region sample set, and the initialization parameters are fine-tuned based on the inner loop learning rate to obtain the final recognition model parameters adapted to the handwriting style of the target region. During this fine-tuning process, the final recognition model parameters are equal to the loaded initialization parameters minus the product of the inner loop learning rate and the gradient of the loss function of the target region sample set; Next, the weights of all layers in the final recognition model parameters are frozen to ensure that the parameters of the repair module, feature extraction module and semantic error correction module remain stable during the inference process. The final output recognition model can take the original paper-based document scan of the target region as input and output the error-corrected recognition result adapted to the handwriting style of the region, thereby achieving rapid cross-regional migration and high-precision recognition.

[0016] The beneficial effects of this invention are as follows: Deep restoration technology restores the topological structure of handwriting, ensuring the clarity and integrity of input data; relying on the feature extraction capabilities of a hybrid architecture, it can accurately capture the subtle textures and global spatial relationships of handwritten strokes, generating highly discriminative feature representations; combined with reinforcement learning strategies, the solution can transform cadastral business logic into an intelligent verification mechanism, performing logical reasoning and automatic correction on the recognition results, significantly reducing the coding error rate; furthermore, with the help of a meta-learning framework, the solution possesses strong cross-domain adaptability, requiring only a few labeled samples to quickly migrate to different regions, significantly reducing the cost of manual review while achieving accurate and efficient conversion of cadastral elements from images to structured data. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of the system structure of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0020] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0021] Example 1: This example provides the following... Figure 1-2 The system illustrates an automatic cadastral feature extraction and recognition system based on deep learning, specifically comprising: a repair module, a feature extraction module, a semantic error correction module, and an adaptive optimization module connected in sequence; wherein, Repair module: In response to the acquired original paper-based book map scan, the original paper-based book map scan is preprocessed to generate a standardized repair image. The preprocessing includes ink smudge repair and paper wrinkle correction through a conditional generative adversarial network, and the stroke topological continuity is constrained by the ink connectivity loss function. Feature extraction module: Receives the standardized repaired image, extracts the local texture features and global topological features of the handwritten strokes through a hybrid architecture of convolutional neural network and transformer, and generates the corresponding stroke topology map feature vector; Semantic error correction module: Receives the feature vector of the stroke topology map, transforms the preset cadastral coding rules into a reward function based on the reinforcement learning strategy, performs logical verification and automatic correction on the preliminary recognition results generated by the pre-trained OCR head, and outputs the error-corrected recognition results that conform to the preset administrative division coding format; Adaptive optimization module: Receives the recognition results after error correction, and based on the model-independent meta-learning framework, uses a small number of labeled samples from the target region to quickly fine-tune the parameters of the repair module, feature extraction module, and semantic error correction module, and outputs a final recognition model adapted to the handwriting style of the target region.

[0022] In this embodiment, it is specifically necessary to explain the preprocessing process of the original paper-based document scan in the repair module as follows: First, the original paper-based document scans are normalized in resolution and scaled to a set pixel size. The color images are then converted into single-channel grayscale images. The Canny edge detection algorithm is used to extract edge features from the grayscale images. False edges are filtered out by setting high and low thresholds (high threshold of 200 and low threshold of 100), generating a binarized edge feature map, where a pixel value of 1 represents an edge point and a pixel value of 0 represents a background point. Next, edge features are extracted from the grayscale image using an edge detection algorithm to generate a binarized edge feature map. This edge feature map is then used as a conditional constraint input to the generator of the conditional generative adversarial network. Subsequently, the generator adopts an encoder-decoder architecture. The encoder part extracts multi-scale features of the grayscale image through multiple residual blocks and downsamples it to a low resolution. The decoder part fuses the features extracted by the encoder part with the upsampling results through skip connections and outputs a preliminary repaired image. A spatial attention mask is introduced at the end of the decoder to enhance the ink smudged area with weights, resulting in the restored image. A weight matrix with values ​​between zero and one is generated by the Sigmoid activation function to enhance the ink smudged area with weights, resulting in the restored image. Finally, the difference in the number distribution of connected components between the restored image and the pre-stored standard clear cadastral map template is calculated using the ink connectivity loss function. The parameters of the generator are iteratively optimized through the backpropagation algorithm until the strokes of the output standardized restored image remain continuous and unbroken. The optimization rounds are set to 300 rounds with a learning rate of 0.0002 until the strokes of the output standardized restored image remain continuous and unbroken, and the error in the number of eight-neighbor connected regions is less than 5%. The discriminator in the repair module adopts a block discriminant structure to verify the authenticity of the repair results; it outputs a 70x70 receptive field to verify the authenticity of the repair results. The discriminator receives the repaired image and the corresponding edge feature map as input, and outputs a local realism probability map in the form of a two-dimensional matrix. Each element in the two-dimensional matrix represents the clarity of the corresponding local image patch in the input image. During training, the repair module minimizes the cross-entropy loss of the discriminator, updates the discriminator parameters using the Adam optimizer, and sets the discriminator training frequency to once every five generator iterations, forcing the generator to produce repair results that are closer to the real cadastral map texture. Meanwhile, the ink connectivity loss function measures the topological structure of strokes by counting the number of eight-neighbor connected regions. When isolated noise or stroke interruptions appear in the restored image, the value of this loss function increases. The weight coefficient of the ink connectivity loss function is set to 0.1, thereby guiding the generator to maintain the inherent geometric structure and topological continuity of cadastral elements while restoring ink color, ensuring that the standardized restored image output to the feature extraction module has high availability.

[0023] In this embodiment, it is specifically necessary to explain the process of extracting local texture features and global topological features of handwritten strokes in the feature extraction module as follows: First, the standardized repaired image is received and input into the backbone of the convolutional neural network, passing through the initial convolutional layer, batch normalization layer, activation function layer and max pooling layer in sequence, as well as four residual blocks; The initial convolutional layer is configured with a 7x7 kernel, 64 output channels, and a stride of 2, and is used to initially extract shallow edge features; the batch normalization layer and activation function layer are used to enhance the non-linear expressive power of the features; the max pooling layer is configured with a 3x3 kernel and a stride of 2, and is used to downsample the feature map and compress the spatial dimension. Four residual blocks are used to extract local texture features at multiple scales, and finally output local feature maps; Next, the local feature map is flattened into a sequence, and sine and cosine position encoding is added to preserve the spatial position information of the strokes. This sequence is then input into the transformer encoder in the hybrid architecture of convolutional neural network and transformer, where a dual spatial and channel attention mechanism is introduced. Spatial attention calculates attention weights by multiplying the query matrix, key matrix, and value matrix. The attention weights are obtained by Softmax normalization of the transpose of the query and key. The query matrix, key matrix, and value matrix are generated by multiplying the input sequence with three different learnable weight matrices. The calculated attention weights are used to perform a weighted summation on the value matrix to output spatially weighted features. During training, the dimensions of the query, key, and value matrices are all set to 64, and the number of multi-head attention heads is 8. By concatenating the eight attention results and performing a linear transformation, spatially weighted features with a dimension of 512 are output. Channel attention compresses spatially weighted features into channel description vectors through global average pooling, and then generates channel weights through a fully connected layer and activation function. The spatially weighted features and channel weights are multiplied element-wise to enhance the channel dimension, and the enhanced global topological feature sequence is output. The number of intermediate neurons in the fully connected layer is set to 256. The channel weights with values ​​between zero and one are generated through the Sigmoid activation function to amplify important channel features in the spatially weighted features and suppress redundant background noise. The specific process of generating the feature vector of the stroke topology graph is as follows: The enhanced global topological feature sequence is set with a feature response threshold of 0.5 (i.e., 50% feature response intensity). The set of stroke endpoints is filtered out, and the Euclidean distance between any two endpoints is calculated. If the distance is less than the preset endpoint spacing threshold of 10 pixels, an edge is added to connect the two endpoints, thereby constructing a stroke topological graph. The feature response threshold is set to 0.5 to filter out pixels with a feature response intensity higher than 50% as candidate endpoints. The preset endpoint spacing threshold is set to 10 pixels to determine whether two endpoints belong to the same stroke structure. Next, a two-layer graph convolutional network is used to process the stroke topology graph. The first layer performs a linear transformation on the normalized adjacency matrix and the initial feature matrix of the stroke endpoints, and then inputs the transformed matrix into an activation function to output intermediate features. The normalized adjacency matrix is ​​calculated by adding the adjacency matrix and the identity matrix, and then multiplying the left and right sides of the inverse square root of the degree matrix, respectively. The number of output channels of the first layer's linear transformation is set to 256, and nonlinearity is introduced through the ReLU activation function. The second layer performs a linear transformation on the normalized adjacency matrix and intermediate features to output higher-order features; the number of output channels for the second-layer linear transformation is set to 256; finally, global average pooling is performed on the higher-order features to generate a stroke topology graph feature vector; the global average pooling operation takes the average of the 256-dimensional features of all nodes in the node dimension to obtain the final 256-dimensional stroke topology graph feature vector.

[0024] In this embodiment, it is necessary to specifically explain the process of performing logical verification and automatic correction in the semantic error correction module as follows: First, the pre-defined cadastral coding rules are formally modeled and decomposed into three categories: format constraints, regional consistency constraints, and logical rationality constraints. The format constraints stipulate that the code length is 19 digits, consisting of three parts: the first 6 digits are the county-level (municipal district, county-level city, banner) administrative division code; the middle 6 digits are the cadastral district and cadastral sub-district code; and the last 5 digits are the sequence code, which is a number, and the feature code is two uppercase letters. For example, the first 6 digits of the administrative division code of Yuzhong District, Chongqing Municipality can be represented as "500103", where "50" represents Chongqing Municipality, "01" represents the municipal district, and "03" represents Yuzhong District; the middle 6 digits of the cadastral district and cadastral sub-district code "123456" represent the 12th cadastral district and the 34th sub-district of Yuzhong District; the feature code "GH" represents residential land; and the last 5 digits of the sequence code "00001" represent the first real estate unit within the cadastral sub-district. The regional consistency constraint stipulates that the first 6 digits of the target region's code prefix must completely match the preset county-level (municipal district, county-level city, banner) administrative division code; for example, if the target region is located in Yuzhong District of Chongqing, then the first 6 digits of the identification result must be "500103"; The logical rationality constraint stipulates that the real estate unit serial number (i.e., the last 5 digits of the sequence code) must be greater than or equal to 1 and less than or equal to 99999, and the real estate unit serial number cannot exceed the current maximum number of real estate units in the cadastral area and sub-area; if the current maximum number of real estate units in the cadastral area and sub-area is set to 5000, then the valid range of the real estate unit serial number is 1 to 5000; Next, a multi-dimensional reward function is constructed based on the above three types of constraints to quantify the compliance of the preliminary identification results. The multi-dimensional reward function is composed of a weighted sum of format reward items, regional consistency reward items, and logical rationality reward items. The specific judgment and value rules for each reward item are as follows: the format reward item is used to judge whether the preliminary identification results conform to the encoding format of twelve pure numbers. If it does, a positive value of one is returned; otherwise, a negative value of one is returned. The regional consistency reward item is used to judge whether the first six regional prefixes of the preliminary identification results are completely consistent with the preset administrative division code of the administrative village. If they are consistent, a positive value of one is returned; otherwise, a negative value of one is returned. The logical rationality reward item is used to judge whether the last three digits of the real estate unit number in the preliminary identification result are within a reasonable range. If the real estate unit number is greater than or equal to 1 and less than or equal to the current maximum number of land parcels in the village group, a positive value of 1 is returned; otherwise, a negative value of 0.5 is returned. Finally, the three reward items are multiplied by the preset weight coefficients of 0.4, 0.3 and 0.3 respectively, and then summed to generate the total reward value used to evaluate the compliance of the preliminary identification results. The weight coefficients are set according to the following criteria: format constraints are the most basic, so they are given the highest weight of 0.4; regional consistency and logical rationality are equally important, and each is given a weight of 0.3. The reinforcement learning strategy uses a proximal policy optimization algorithm to train the policy network. The feature vector of the stroke topology map and the spliced ​​features of the preliminary recognition result are used as the input state of the policy network. The policy network outputs the correction action for each digit of the preliminary recognition result. During training, the objective function is defined as maximizing the expected reward, and a shearing mechanism is introduced to limit the magnitude of policy updates. By calculating the shearing approximation of the probability ratio between the old and new policies and combining it with the advantage function to estimate the relative benefit of actions, the stability of the training process is ensured. The advantage function is calculated using the generalized advantage estimation method and is used to measure the advantage of performing a specified corrective action in a specified state compared to the average level. The policy network parameters are updated using the gradient ascent method to maximize the objective function value, thereby learning a policy that can logically correct recognition errors based on stroke topological features. The optimization process of the objective function introduces a shearing mechanism to limit the probability ratio of the new and old policies to between 0.8 and 1.2, in order to prevent the model from oscillating due to excessively large single update amplitude. At the same time, the advantage function is calculated through the generalized advantage estimation method to balance bias and variance, ensuring that the estimation of the policy gradient is both unbiased and has low variance. The shearing coefficient epsilon is set to 0.2, which limits the probability ratio of the new and old policies to the interval [0.8, 1.2]. The discount factor gamma in the generalized advantage estimation is set to 0.99 to balance the importance of immediate rewards and future rewards. The semantic error correction module selects the optimal action sequence from the correction action distribution output by the policy network through greedy decoding, and applies the optimal action sequence to the preliminary recognition result to generate the corrected recognition result. Specifically, for each digit in the preliminary recognition result, the policy network outputs a ten-dimensional probability distribution (corresponding to the digits 0-9), and the greedy decoding selects the digit with the highest probability value as the correction result for that digit. Next, the total reward value of the corrected identification result is calculated. If the total reward value is greater than the preset compliance threshold of 0.8, the corrected identification result is confirmed to be valid and output. The compliance threshold is set to 0.8, that is, when the total reward value of the corrected identification result is greater than 0.8, it is considered to comply with the cadastral coding rules. If the total reward value does not meet the target, the reasoning process of the policy network will be re-executed until a corrected identification result that meets the compliance requirements is generated. The maximum number of retries is set to 5. If the compliance threshold cannot be reached in 5 consecutive reasoning attempts, the preliminary identification result will be output and marked as pending manual review. The final corrected recognition result conforms to the preset administrative division coding format, and its logical correctness is constrained by the semantics of the stroke topology feature vector. This ensures that the recognition result conforms to the cadastral business rules while avoiding semantic deviations that may be caused by pure rule correction.

[0025] In this embodiment, it is necessary to specifically explain the process of performing rapid fine-tuning in the adaptive optimization module as follows: First, in the meta-training phase, prepare no fewer than ten source region datasets, each containing a large number of labeled samples. For each source region task, divide its source region dataset into a support set and a query set. The support set is used to simulate the subsequent rapid adaptation process, and the query set is used for meta-optimization. Set the number of samples in the support set to account for 20% of the total number of samples in each source region dataset, and the number of samples in the query set to account for 80% to ensure the diversity and representativeness of the task distribution during the meta-training process. The inner loop update is performed, calculating the gradient based on the loss value on the support set, and updating the initial parameters using the inner loop learning rate to obtain the task-fit parameters. During this process, the inner loop update follows the gradient descent principle, meaning the task-fit parameters equal the initial parameters minus the product of the inner loop learning rate and the gradient of the support set loss function. The inner loop learning rate controls the step size of each parameter update, with a range of 0.001 to 0.1. A typical value for the inner loop learning rate is 0.01, which offers the best balance between convergence speed and stability. The gradient of the support set loss function represents the rate of change of the error of the initial parameters on the support set samples. Subsequently, the outer loop meta-loss is calculated, summing the loss values ​​of each source region task on the query set. The initialization parameters are then updated based on the outer loop learning rate. Through repeated iterations, initialization parameters with strong generalization ability are obtained, allowing the model to adapt to new tasks with only a small number of gradient updates. The outer loop update also follows the gradient descent principle, i.e., subtracting the product of the outer loop learning rate and the gradient of the meta-loss function from the initialization parameters. The outer loop learning rate controls the magnitude of parameter updates during meta-optimization, with a range of 0.0001 to 0.01. A typical value of 0.001 is set for the outer loop learning rate to smoothly integrate the knowledge of multiple source region tasks at the meta-level. The gradient of the meta-loss function represents the rate of change of the overall error after summing the losses of all source region tasks on the query set. The adaptive optimization module continues to execute the following process: First, a small set of labeled samples from the target region is received, the initial parameters obtained during the meta-training phase are loaded, and no more than five gradient descent operations are performed. The gradient is calculated using the loss value of the target region sample set, and the initial parameters are fine-tuned based on the inner loop learning rate to obtain the final recognition model parameters adapted to the handwriting style of the target region. The maximum number of labeled samples from the target region is set to fifty, and the maximum number of fine-tuning steps is set to three, in order to avoid overfitting in the case of small samples. During this fine-tuning process, the final recognition model parameters are equal to the loaded initialization parameters minus the product of the inner loop learning rate and the gradient of the loss function of the target region sample set; Next, the weights of all layers in the final recognition model parameters are frozen to ensure that the parameters of the repair module, feature extraction module, and semantic error correction module remain stable during inference. The freezing operation specifically refers to setting the model parameters to an untrainable state, thereby maintaining a high degree of adaptability to the handwriting style of the target region during deployment. The final output recognition model can take the original paper-based document scan of the target region as input and output the error-corrected recognition result adapted to the handwriting style of the region, thereby achieving rapid cross-regional transfer and high-precision recognition.

[0026] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0027] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0028] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0029] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0030] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0031] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0032] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A deep learning-based cadastral feature automatic extraction and identification system, characterized in that, Specifically, it includes: Repair module: In response to the acquired original paper-based book map scan, the module performs ink smudge repair and paper wrinkle correction on the original paper-based book map scan using a conditional generative adversarial network, and uses the ink connectivity loss function to constrain the stroke topological continuity in order to generate a standardized repair image. Feature extraction module: Receives the standardized repaired image, extracts the local texture features and global topological features of the handwritten strokes through a hybrid architecture of convolutional neural network and transformer, and generates the corresponding stroke topology map feature vector; Semantic error correction module: Receives the feature vector of the stroke topology map, transforms the preset cadastral coding rules into a reward function based on the reinforcement learning strategy, performs logical verification and automatic correction on the preliminary recognition results generated by the pre-trained OCR head, and outputs the error-corrected recognition results that conform to the preset administrative division coding format; Adaptive optimization module: Receives the recognition results after error correction, and based on the model-independent meta-learning framework, uses a small number of labeled samples from the target region to quickly fine-tune the parameters of the repair module, feature extraction module, and semantic error correction module, and outputs a final recognition model adapted to the handwriting style of the target region. 2.The deep learning-based cadastral feature automatic extraction and identification system according to claim 1, wherein: The preprocessing process for the original paper-based document scan in the repair module is as follows: First, the original paper-based document scans are normalized in resolution and scaled to a set pixel size. The color images are then converted into single-channel grayscale images. Edge features are extracted from the grayscale images using an edge detection algorithm to generate a binarized edge feature map. This edge feature map is then used as a constraint input to the generator of a conditional generative adversarial network. Subsequently, the generator adopts an encoder-decoder architecture. The encoder part extracts multi-scale features of the grayscale image through multiple residual blocks and downsamples it to a low resolution. The decoder part fuses the features extracted by the encoder part with the upsampling results through skip connections and outputs a preliminary repaired image. A spatial attention mask is introduced at the end of the decoder to weight and enhance the ink smudged area, resulting in the restored image. The difference in the distribution of connected component numbers between the restored image and the pre-stored standard clear cadastral map template is calculated using the ink connectivity loss function.

3. The automatic cadastral element extraction and recognition system based on deep learning according to claim 2, characterized in that: The discriminator in the repair module adopts a block discriminant structure to verify the authenticity of the repair results; The discriminator receives the restored image and its corresponding edge feature map as input and outputs a local realism probability map in the form of a two-dimensional matrix. Each element in the two-dimensional matrix represents the clarity of the corresponding local image patch in the input image. During training, the restoration module forces the generator to generate restoration results that are closer to the texture of the real cadastral map by minimizing the cross-entropy loss of the discriminator. Meanwhile, the ink connectivity loss function measures the topological structure of the stroke by counting the number of eight-neighbor connected regions. When isolated noise appears in the restored image, the value of this loss function increases, thereby guiding the generator to maintain the inherent geometric structure and topological continuity of cadastral elements while restoring the ink color, ensuring that the standardized restored image output to the feature extraction module has high availability.

4. The automatic cadastral element extraction and recognition system based on deep learning according to claim 3, characterized in that: In the feature extraction module, the process of extracting local texture features and global topological features of handwritten strokes is as follows: First, the standardized repaired image is received and input into the backbone of the convolutional neural network. It then passes through the initial convolutional layer, batch normalization layer, activation function layer, max pooling layer, and four residual blocks in sequence; the output is a local feature map. Next, the local feature map is flattened into a sequence, and sine and cosine position encoding is added to preserve the spatial position information of the strokes. This sequence is then input into the transformer encoder in the hybrid architecture of convolutional neural network and transformer, where a dual spatial and channel attention mechanism is introduced. Spatial attention calculates attention weights by multiplying the query matrix, key matrix, and value matrix, and then uses the calculated attention weights to perform a weighted summation of the value matrix to output spatial weighted features. Channel attention compresses spatially weighted features into channel description vectors through global average pooling, then generates channel weights through a fully connected layer and activation function. The spatially weighted features and channel weights are multiplied element-wise to enhance the channel dimension, and the enhanced global topological feature sequence is output.

5. The automatic cadastral feature extraction and recognition system based on deep learning according to claim 4, characterized in that: The specific process of generating the feature vector of the stroke topology graph is as follows: A feature response threshold is set for the enhanced global topological feature sequence to filter out the set of stroke endpoints. The Euclidean distance between any two endpoints is calculated. If the distance is less than the preset endpoint spacing threshold, an edge is added to connect the two endpoints, thereby constructing a stroke topological graph. Next, a two-layer graph convolutional network is used to process the stroke topology graph. The first layer performs a linear transformation on the normalized adjacency matrix and the initial feature matrix of the stroke endpoints, and then inputs the transformation into the activation function to output intermediate features. The normalized adjacency matrix is ​​calculated by adding the adjacency matrix and the identity matrix, and then multiplying the left side by the inverse square root of the degree matrix and the right side by the inverse square root of the degree matrix, respectively. The second layer performs a linear transformation on the normalized adjacency matrix and intermediate features to output higher-order features; finally, global average pooling is performed on the higher-order features to generate stroke topology graph feature vectors.

6. The automatic cadastral feature extraction and recognition system based on deep learning according to claim 5, characterized in that: The specific process of performing logical verification and automatic correction in the semantic error correction module is as follows: First, the pre-defined cadastral coding rules are formally modeled and decomposed into three categories: format constraints, regional consistency constraints, and logical rationality constraints. The format constraints stipulate that the code length is 19 characters, consisting of three parts: the first 6 characters are the county-level administrative division code, the middle 6 characters are the cadastral district and cadastral sub-district code, and the last 5 characters are the sequence code, which is a number, and the feature code is 2 uppercase letters; the regional consistency constraint stipulates that the first 6 characters of the target region's code prefix must completely match the preset county-level administrative division code; the logical rationality constraint stipulates that the real estate unit serial number must be greater than or equal to 1 and less than or equal to the current maximum number of real estate units in the cadastral district and sub-district. Next, a multi-dimensional reward function is constructed based on the three types of constraints. It consists of a weighted sum of format reward items, regional consistency reward items, and logical rationality reward items. Each reward item returns a positive or negative value based on the evaluation result. The three reward items are multiplied by preset weight coefficients and then summed to generate a total reward value used to evaluate the compliance of the preliminary identification results.

7. The automatic cadastral feature extraction and recognition system based on deep learning according to claim 6, characterized in that: The reinforcement learning strategy uses a proximal policy optimization algorithm to train the policy network. The feature vector of the stroke topology map and the spliced ​​features of the preliminary recognition result are used as the input state of the policy network. The policy network outputs the correction action for each digit of the preliminary recognition result. During training, the objective function is defined as maximizing the expected reward, and a shearing mechanism is introduced to limit the magnitude of policy updates. By calculating the shearing approximation of the probability ratio between the old and new policies and combining it with the advantage function to estimate the relative benefit of actions, the stability of the training process is ensured. The advantage function is calculated using the generalized advantage estimation method and is used to measure the advantage of performing a specified corrective action in a given state compared to the average level. The policy network parameters are updated using the gradient ascent method to maximize the objective function value, thereby learning a policy that can logically correct recognition errors based on stroke topological features.

8. The automatic cadastral feature extraction and recognition system based on deep learning according to claim 7, characterized in that: The semantic error correction module selects the optimal action sequence from the correction action distribution output by the policy network through greedy decoding, applies the optimal action sequence to the preliminary recognition result, and generates the error-corrected recognition result. Next, the total reward value of the corrected recognition result is calculated. If the total reward value is greater than the preset compliance threshold, the corrected recognition result is confirmed to be valid and output. If the total reward value does not meet the target, the reasoning process of the policy network will be re-executed until a corrected identification result that meets the compliance requirements is generated. The final corrected recognition result conforms to the preset administrative division coding format, and its logical correctness is constrained by the semantics of the stroke topology feature vector. This ensures that the recognition result conforms to the cadastral business rules while avoiding semantic deviations that may be caused by pure rule correction.

9. The automatic cadastral feature extraction and recognition system based on deep learning according to claim 8, characterized in that: The specific process of performing rapid fine-tuning in the adaptive optimization module is as follows: First, we enter the meta-training phase and prepare ten source region datasets, each containing a large number of labeled samples. For each source region task, we divide its source region dataset into a support set and a query set. The support set is used to simulate the subsequent rapid adaptation process, and the query set is used for meta-optimization. The inner loop update is performed, the gradient is calculated based on the loss value on the support set, and the initial parameters are updated using the inner loop learning rate to obtain the task adaptation parameters. In this process, the inner loop update follows the gradient descent principle, that is, the task adaptation parameters are equal to the initial parameters minus the product of the inner loop learning rate and the gradient of the support set loss function. Then, the outer loop loss is calculated, and the loss values ​​of each source region task on the query set are summarized. The initialization parameters are updated based on the outer loop learning rate. Through repeated iterations, initialization parameters with strong generalization ability are obtained, so that the model can adapt to new tasks with only a small number of gradient updates.

10. The automatic cadastral feature extraction and identification system based on deep learning according to claim 9, characterized in that: The adaptive optimization module continues to perform the following process: First, a small set of labeled samples from the target region is received, the initialization parameters obtained during the meta-training phase are loaded, gradient descent is performed, the gradient is calculated using the loss value of the target region sample set, and the initialization parameters are fine-tuned based on the inner loop learning rate to obtain the final recognition model parameters adapted to the handwriting style of the target region. During this fine-tuning process, the final recognition model parameters are equal to the loaded initialization parameters minus the product of the inner loop learning rate and the gradient of the loss function of the target region sample set; Next, the weights of all layers in the final recognition model parameters are frozen to ensure that the parameters of the repair module, feature extraction module and semantic error correction module remain stable during the inference process. The final output recognition model can take the original paper-based document scan of the target region as input and output the error-corrected recognition result adapted to the handwriting style of the region, thereby achieving rapid cross-regional migration and high-precision recognition.

Images