A geological map text interference cleaning and vectorization method and device based on multi-modal large model semantic understanding

Abstract

The application relates to the technical field of image processing, in particular to a geological map text interference cleaning and vectorization method and device based on multi-modal large model semantic understanding, which comprises the following steps: preprocessing a geological raster image to be processed to obtain a standardized image; performing semantic understanding on the standardized image through a geological multi-modal large model to obtain the region category, semantic importance score and text content of each image region; performing a differentiated processing strategy on the text interference region based on the region category and the semantic importance score to obtain a repaired geological image; performing vectorization extraction on the repaired geological image, associating the text content to the attribute field of the corresponding vector element, and generating vector geological data with attributes. The application realizes end-to-end vectorization of accurate cleaning of interference text of a geological map, recovery of geological structure integrity and automatic association of text and attributes, greatly reduces the cost of manual digitization while retaining key geological information.

Description

Technical Field

[0001] This application relates to the field of image processing technology, specifically to a method and apparatus for cleaning and vectorizing textual and image interference in geological maps based on multimodal large model semantic understanding. Background Technology

[0002] Geological maps (such as geological structure maps, mineral prediction maps, and geological hazard distribution maps) are core data carriers for geological research and engineering applications. Their vectorized results are widely used in GIS (Geographic Information System) database construction, resource evaluation, and engineering planning. With the deepening of digital geological surveys, the need for high-precision vectorization of massive historical geological raster maps is becoming increasingly urgent. However, geological maps commonly suffer from overlapping or adhering text annotations with geological lines, patches, patterns, and other elements. How to remove text interference while preserving effective geological information to the greatest extent and ensuring vector geometric accuracy has become a key technical bottleneck in the current intelligent processing of geological maps.

[0003] Existing methods for vectorization and noise reduction of geological maps mainly include: traditional image processing methods, text removal methods based on general object detection, and segmentation methods based on deep learning. Traditional image processing methods use edge detection, morphological filtering, and binarization segmentation to separate lines, patches, text, legends, and other elements from geological raster images, and then generate SHP (Shapefile) vector data using vector tracking algorithms. Text removal methods based on general object detection utilize OCR (Optical Character Recognition) technology or general object detection models to locate text regions in the image, then fill the pixels of these regions with the background color or the mean of the surrounding texture before performing vectorization. Deep learning-based segmentation methods use semantic segmentation models to perform pixel-level classification of geological maps, distinguishing between geological boundaries, text annotations, lithological patterns, etc., and then vectorize the image after text removal.

[0004] While the aforementioned methods improve classification accuracy, significant shortcomings remain. First, text-image overlap leads to severe loss of geometric accuracy. Text annotations on geological maps often overlap or adhere to geological boundaries, fault lines, and lithological boundaries. Traditional OCR or object detection methods can only output rectangular text boxes, leaving blank spaces or using rigid filling after text removal, resulting in boundary breaks and topological errors during vector tracking. The geometric accuracy after vectorization fails to meet engineering application requirements. Second, existing methods lack semantic discrimination capabilities, treating all pixels detected as text as noise and failing to distinguish between noise and effective geological information. Geological map annotations often contain important geological information such as occurrence symbols, stratigraphic codes, and structural numbers; simple removal leads to the loss of key geological elements and fails to assess the importance differences between different text regions. Third, texture restoration capabilities are insufficient. Existing methods often use simple mean filling or copying of surrounding pixels after text removal, failing to perform generative restoration based on the geological semantics of occluded areas (such as fault line extension direction and lithological pattern patterns), leading to interruptions or distortions in vector extraction results. In addition, geological maps from different sources (regional geological maps, detailed mining area maps, and engineering geological maps) have significant differences in cartographic specifications, fonts, colors, and layer overlay relationships. Existing models are usually trained for a single map type, and their performance degrades significantly when transferring between map types, limiting their generalization ability.

[0005] Application content The purpose of this application is to provide a method and apparatus for cleaning and vectorizing textual interference in geological maps based on multimodal large model semantic understanding in order to solve at least one of the above-mentioned technical problems.

[0006] This application achieves the above objectives through the following technical solutions: A method for cleaning and vectorizing textual and image interference in geological maps based on multimodal large-scale model semantic understanding includes the following steps: The geological raster image to be processed is preprocessed to obtain a standardized image; Semantic understanding of the standardized images is performed using a geological multimodal large model to obtain the region category, semantic importance score, and text content of each image region; Based on the region category and the semantic importance score, a differentiated processing strategy is applied to the regions with textual and image interference to obtain the restored geological image; The restored geological image is vectorized and extracted, and the text content is associated with the attribute fields of the corresponding vector elements to generate attributed vector geological data.

[0007] Furthermore, the geological multimodal large model includes: a visual encoding module, a cross-modal alignment module, and a language decoding module; The visual encoding module is used to extract visual features from the standardized image to obtain image features; The cross-modal alignment module is used to map the image features to the feature space of the language decoding module to obtain a joint feature representation; The language decoding module is used to perform semantic decoding on the joint feature representation to obtain the region category, the semantic importance score, and the text sequence vector corresponding to the text content for each image region.

[0008] Furthermore, the geological multimodal large model is obtained through the following steps: Based on the region category prediction probability distribution and the corresponding true category label of each image region, calculate the cross-entropy classification loss; The mean squared error of the importance score is calculated based on the semantic importance score and the corresponding true importance score of each image region. Based on the text sequence vector, calculate the character recognition sequence loss; The joint loss function is obtained by weighted summing of the cross-entropy classification loss, the mean square error of the importance score, and the character recognition sequence loss; The parameters of the geological multimodal large model are adjusted with the goal of minimizing the joint loss function.

[0009] Furthermore, a differentiated processing strategy is implemented for areas with text and image interference, including: For regions whose region category is text or label, compare the semantic importance score with the semantic importance threshold; If the semantic importance score is lower than the semantic importance threshold, then the region is removed. If the semantic importance score is higher than or equal to the importance threshold, and the region is spatially connected to the geological boundary elements, then the image-text separation and repair operation is performed on the spatially connected region. If the semantic importance score is higher than or equal to the importance threshold, and the region is not spatially connected to the geological boundary feature, and the region category belongs to geological indicator labeling, then the region is retained and its text content is used as the attribute field of the corresponding vector feature.

[0010] Furthermore, the image-text separation and repair operation includes: Based on the region category, the pixels belonging to the text or label category in the spatially connected region are binarized to obtain the text region mask; Obtain the predicted trajectory of the geological boundary obscured by the text region mask; Based on the local image of the spatially contiguous region, the text region mask, and the predicted trajectory, generative texture inpainting is performed to obtain the repaired geological image.

[0011] Furthermore, the generative texture restoration satisfies geological prior constraints, which include: The continuity constraint is used to ensure that the restored geological boundary trajectory remains consistent with the predicted trajectory. Topology consistency constraints are used to ensure that the topological relationship between the repaired region and the surrounding unobstructed regions remains consistent. Lithological periodicity constraints are used to constrain the consistency of periodic texture features in lithological pattern regions.

[0012] Furthermore, the vectorization extraction includes: The restored geological image is converted into a binary edge map, and a graph structure is constructed with edge pixels as nodes and edges defined by eight-neighbor connectivity. Under the path constraint that the node degree is no greater than 2, the maximum weight path is found based on the edge weight and the linear vector features are extracted; The edge weight is determined by the product of the edge strengths between adjacent pixels.

[0013] Furthermore, the text content is associated with the attribute fields of the corresponding vector features, including: Determine the spatial bounding box of each vector element; The text content of image regions whose distance from the spatial bounding box is less than the spatial neighborhood threshold and whose region category belongs to the text or label category is semantically parsed and converted into key-value pairs of attribute names and attribute values, and then written into the attribute fields of the corresponding vector features.

[0014] Furthermore, the method also includes: Based on the attributed vector geological data, geometric integrity, topological accuracy, and attribute retention rate are obtained, and a comprehensive quality score is generated based on the geometric integrity, topological accuracy, and attribute retention rate. If the overall quality score is lower than a preset threshold, the semantic importance threshold, the number of sampling steps for generative texture inpainting, and / or the spatial neighborhood threshold are adaptively adjusted, and the differential processing strategy and the vectorization extraction are re-executed until the overall quality score meets the preset threshold or the maximum number of iterations is reached.

[0015] A device for cleaning and vectorizing textual and image interference in geological maps based on multimodal large model semantic understanding, comprising: The preprocessing unit is used to preprocess the geological raster image to be processed to obtain a standardized image; The semantic understanding unit is used to perform semantic understanding on the standardized image through a geological multimodal large model to obtain the region category, semantic importance score and text content of each image region; A differential processing unit is used to perform a differential processing strategy on regions with text-image interference based on the region category and the semantic importance score, so as to obtain the restored geological image. The vectorization extraction unit is used to extract vectorization data from the restored geological image and associate the text content with the attribute fields of the corresponding vector elements to generate attributed vector geological data.

[0016] The beneficial effects of this application are as follows: First, this application employs semantically driven differential denoising to avoid the loss of key information. This application uses a geological multimodal large model to perform semantic classification and importance scoring of image regions, directly removing low-importance text regions, performing image-text separation and restoration on high-importance regions that are attached to geological boundaries, and retaining high-importance geological indicator labels that are not attached, thus achieving differentiated and refined processing.

[0017] Second, this application integrates generative restoration based on geological priors to ensure the integrity of vector extraction. Based on image-text separation, this application introduces three geological prior constraints: strike continuity, topological consistency, and lithological periodicity. Through generative texture restoration, the occluded areas are reconstructed, ensuring the restoration results conform to geological structural patterns, thereby significantly improving the geometric integrity of subsequent vector extraction.

[0018] Third, this application employs automatic attribute association through a combination of text and image to enhance information value. The text content identified during the semantic understanding stage is automatically converted into attribute name-attribute value key-value pairs after spatial proximity matching and semantic parsing, and written into the attribute fields of the corresponding vector elements. This achieves information enhancement from image to image + attribute, directly supporting subsequent geological analysis and engineering decisions.

[0019] Fourth, this application possesses cross-map generalization capability and strong adaptability. Based on the general semantic understanding capability of the large model and combined with the adaptive quality assessment closed loop, this application can automatically adjust the semantic importance threshold, the number of repair sampling steps, and the spatial neighborhood threshold according to the comprehensive quality score, adapting to multi-source geological maps of different mapping specifications, geological ages, and regions. Its cross-map migration capability is significantly better than existing methods.

[0020] Fifth, this application adopts an end-to-end integrated processing chain to reduce labor costs. From the input of raw geological raster images to the output of attributed vector data, this application covers the entire process of preprocessing, semantic understanding, differential cleaning, generative restoration, vectorization extraction, and attribute association, forming a complete intelligent processing closed loop. Compared to existing segmented operation modes, this significantly reduces manual intervention and is suitable for large-scale engineering applications such as power transmission line selection, mineral resource evaluation, and geological disaster investigation. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating a method for cleaning and vectorizing textual and image interference in geological maps based on multimodal large model semantic understanding, according to one embodiment of this application. Figure 2 This is a schematic diagram of the structure of a geological multimodal large model according to one embodiment of this application; Figure 3 This is a flowchart illustrating the training process of a geological multimodal large model according to one embodiment of this application. Figure 4 A flowchart illustrating a differential processing strategy for regions with textual and graphic interference according to one embodiment of this application; Figure 5 This is a flowchart illustrating the image-text separation and repair operation according to one embodiment of this application. Figure 6 This is a schematic diagram of a geological map text interference cleaning and vectorization device based on multimodal large model semantic understanding, according to one embodiment of this application. Detailed Implementation

[0022] The content of this application will now be discussed with reference to exemplary embodiments. It should be understood that the described embodiments are merely intended to enable those skilled in the art to better understand and thus implement the content of this application, and are not intended to imply any limitation on the scope of this application.

[0023] As used herein, the term "comprising" and its variations are to be interpreted as open-ended terms meaning "including but not limited to". The term "based on" is to be interpreted as "at least partially based on". The terms "one embodiment" and "an embodiment" are to be interpreted as "at least one embodiment".

[0024] As mentioned earlier, existing technologies for handling text-image interference in geological maps generally suffer from several problems: severe loss of geometric accuracy due to text-image adhesion; loss of effective information due to a lack of semantic differentiation capabilities; distortion of vector results due to insufficient texture restoration capabilities; and limited generalization ability across map types. Therefore, there is an urgent need for a text-image interference cleaning and vectorization method capable of semantically understanding geological map content, intelligently distinguishing text from effective information, and achieving generative texture restoration, in order to improve the geometric accuracy and information integrity of vectorized geological map results.

[0025] Based on this, this application proposes a scheme for cleaning and vectorizing textual and image interference in geological maps based on multimodal large-scale model semantic understanding. This scheme can semantically understand the content of geological maps, intelligently distinguish between text and valid information, and achieve textual and image interference cleaning and vectorization through generative texture restoration. Using a multimodal large-scale language model as the core, it achieves accurate cleaning and high-quality vectorized output of textual and image interference in geological maps through geological semantic-guided text region recognition and classification, a differential processing strategy based on semantic importance, and generative texture restoration technology.

[0026] Example 1 Figure 1 This is a flowchart illustrating a method for cleaning and vectorizing textual interference in geological maps based on multimodal large model semantic understanding, according to one embodiment of this application. Figure 1 As shown in the embodiments of this application, a method for cleaning and vectorizing textual and image interference in geological maps based on multimodal large model semantic understanding is provided, including the following steps: Step S102: Preprocess the geological raster image to be processed to obtain a standardized image; Step S104: Semantic understanding of the standardized image is performed using a geological multimodal large model to obtain the region category, semantic importance score, and text content of each image region; Step S106: Based on the region category and semantic importance score, a differentiated processing strategy is performed on the image-text interference region to obtain the restored geological image; Step S108: The restored geological image is vectorized and extracted, and the text content is associated with the attribute fields of the corresponding vector elements to generate attributed vector geological data.

[0027] Next, in conjunction with the appendix Figure 2-5 Please explain each of the above steps in detail.

[0028] In some embodiments, step S102 includes: Step 1: Obtain the geological raster image to be processed and perform geometric correction and coordinate registration; The source formats of the geological raster images to be processed include, but are not limited to, common geological map storage formats such as TIFF (Tag Image File Format), JPEG (Joint Photographic Experts Group), PNG (Portable Network Graphics), and GeoPDF (Geospatial PDF). Since the original maps may originate from scanned paper maps, screenshots of remote sensing images, or electronic maps exported from different software, they often suffer from geometric distortion, missing coordinates, or inconsistent projections. Therefore, it is necessary to first perform geometric correction and coordinate registration on the images to ensure that the subsequent vectorized output has accurate spatial coordinates.

[0029] Specifically, geometric correction aims to eliminate geometric distortions in maps caused by factors such as scanning warp, paper stretching, and imaging aberration. By identifying control points within the map sheet, methods such as affine transformation, polynomial transformation, or local spline interpolation are used to correct the raster image to a standard geographic coordinate system. Coordinate alignment criteria further align the corrected image with a specified spatial reference system, ensuring that each pixel has true geospatial coordinates.

[0030] Step two: Perform scale normalization and color space conversion on the image; After geometric correction, the image is scaled and color space converted to reduce the interference of uneven illumination and color differences on subsequent semantic segmentation tasks.

[0031] Scale normalization aims to unify images from different sources and with different resolutions to a preset standard size or spatial resolution, avoiding model perception bias caused by differences in image scale.

[0032] In terms of color space conversion, the image is converted from the RGB color space to the HSV (Hue-Saturation-Value) color space. Compared to the RGB space, the HSV space decouples the hue information (H, S) from the luminance channel (V), making the luminance component V independent of color attributes. This allows for a more effective characterization of luminance unevenness caused by scanning conditions, lighting angles, or paper aging. By separating the hue and luminance channels, subsequent processing can perform adaptive equalization or gamma correction on the luminance component V. This suppresses lighting unevenness while fully preserving the hue characteristics of key elements such as geological boundaries and lithological patterns, thereby improving the stability of the multimodal large model in discriminating text and image regions. Color space conversion can also use the HSL (Hue-Saturation-Lightness) or Lab (CIELAB) color spaces.

[0033] To achieve accurate identification and semantic discrimination of textual and image interference elements in geological maps, step S104 constructs and deploys a Geological Multimodal Large Model (Geo-MLM). Unlike traditional single-task detection or segmentation models, this model employs a multi-task joint output mechanism, simultaneously completing region category discrimination, semantic importance assessment, and text content recognition. The three tasks share underlying visual features and mutually reinforce each other through a joint loss function, ultimately providing fine-grained semantic basis for subsequent differentiated processing.

[0034] Figure 2 This is a schematic diagram of a geological multimodal large model structure according to one embodiment of this application. Figure 2As shown, in some embodiments, the geological multimodal large model includes: a visual encoding module 210, a cross-modal alignment module 220, and a language decoding module 230; wherein, The visual encoding module 210 extracts visual features from the standardized image to obtain image features; For example, the standardized image is set as H represents the total number of pixels in the vertical direction of the image, and W represents the total number of pixels in the horizontal direction of the image. Visual encoder Using the Vision Transformer (ViT) or CLIP visual branch as the backbone network, normalized images are mapped to image features. : (1) in, Indicates the visual feature dimension (e.g., 768 dimensions); the downsampling factor is 16. Image features The spatial structure of the original image was preserved, and each image region... Corresponding to a visual vector .

[0035] The cross-modal alignment module 220 maps image features to the feature space of the language decoding module to obtain joint feature representations, enabling visual information to be directly understood by the language decoder. The language decoding module 230, based on large language models such as LLaMA, Qwen, or GPT series variants, receives the joint feature representation output by the cross-modal alignment module 220 for each image region. Simultaneously execute three decoding tasks, outputting the region category, semantic importance score, and text sequence vector corresponding to the text content for each image region: (2) in, C represents the total number of region categories. Represents image region The region category prediction probability distribution is a list of C scores, each representing an image region. The probability of belonging to the corresponding region category. ; Represents image region semantic importance score, The higher the value, the more critical the region is to geological interpretation; Represents image region The text sequence vector corresponding to the text content; Represents image region Image features; This indicates that the language decoding module uses a large language model; Representing a large language model The set of trainable parameters.

[0036] Final regional categories : (3) The region categories include: geological boundaries, fault lines, lithological patterns, stratigraphic symbols, attitude symbols, legend text, and auxiliary annotations. For regions identified as text, the corresponding text content is output, such as J, γ, ∠30°, etc. Attitude symbols and stratigraphic symbols have higher semantic importance scores, while legend text and auxiliary annotations have lower scores.

[0037] Figure 3 This is a flowchart illustrating the training process of a geological multimodal large model according to one embodiment of this application. Figure 3 As shown, in some embodiments, the training and parameter fine-tuning process of a geological multimodal large model includes: Step S302: Calculate the cross-entropy classification loss based on the region category prediction probability distribution and the corresponding true category label of each image region; Step S304: Calculate the mean squared error of the importance score based on the semantic importance score of each image region and the corresponding true importance score; Step S306: Calculate the text recognition sequence loss based on the text sequence vector; Step S308: The cross-entropy classification loss, the mean square error of importance score, and the character recognition sequence loss are weighted and summed to obtain the joint loss function; Step S310: Adjust the parameters of the geological multimodal large model with the goal of minimizing the joint loss function.

[0038] While maintaining the general semantic understanding capabilities of large models, the geological multimodal large model incorporates professional geological knowledge to improve the recognition accuracy and generalization ability of multi-source geological maps.

[0039] For example, a training dataset containing over 100,000 multi-source geological maps is constructed, with each map accompanied by pixel-level region category annotations, text transcription annotations, and semantic importance score annotations. The cross-entropy classification loss is calculated using the following formula. : (4) in, Represents the cross-entropy classification loss; Indicates all spatial regions Summation is used to iterate through every pixel or every image block in an image. Representing a spatial region The actual labels for region category j are manually labeled. When the current spatial region belongs to region category j, ∈{0,1}. =1, otherwise =0; Represents image region The probability of belonging to region category j; Indicates the total number of region categories; Calculate the cross-entropy classification loss During the process, through calculation Mapping probability values ​​to a loss scale, when When it approaches 1, When the value is close to 0, the loss is close to 0; conversely, when the value is close to 0, the When approaching 0, As it approaches negative infinity, the loss becomes enormous.

[0040] The mean squared error of the importance score is calculated using the following formula. : (5) in, The mean squared error of the importance score; This indicates the total number of regions involved in the calculation, such as the total number of all labeled image regions. Represents image region semantic importance score, The higher the importance to geological interpretation, The higher the score, such as when the area category is an attitude symbol or stratigraphic code. When the value is close to 1, and the region category is legend text or auxiliary labeling. Approaching 0; Represents image region The true importance score is manually marked by geological experts. For example, the true importance score of occurrence symbols and stratigraphic codes is set to 0.8~1.0, and the true importance score of legend text and auxiliary annotations is set to 0.0~0.2. The specific settings can be flexibly selected according to the actual situation. The mean squared error of importance scores The deviation between the semantic importance score predicted by the model and the true importance score is measured; as the accuracy of the semantic importance score predicted by the model improves, As a result, it decreases, and ideally approaches 0.

[0041] For text sequence vectors in text regions, CTC (Connectionist Temporal Classification) loss or standard cross-entropy loss is used to handle the alignment problem of variable-length text sequences, and the text recognition sequence loss is calculated. .

[0042] The joint loss function is calculated using the following formula. : (6) in, This represents the first weight matrix; This represents the second weight matrix.

[0043] For each weight matrix in the large-scale geological multimodal model, a parameter-efficient fine-tuning (PEFT) strategy is adopted to maintain the generality of the large model while enhancing the accuracy of geological semantic recognition. For example, low-rank adaptation (LoRA) parameter updates are used, and the specific update method is as follows: (7) in, This represents the updated weight matrix. ; This represents the pre-trained weight matrix, which is frozen during training. Denotes the first low-rank matrix. ; Describes the second low-rank matrix. The number of rows, d, represents the output dimension of the weight matrix; r represents the input dimension of the weight matrix; the number of columns, k, represents the width of the matrix; and the rank... Their numerical values ​​directly determine the amount of GPU memory and computation time required for training a large geological multimodal model.

[0044] As an alternative, LoRA can be used in conjunction with other PEFT methods such as Adapter Tuning, Prefix Tuning, or PromptTuning.

[0045] Step S106, based on the region category, semantic importance score, and spatial location relationship output in step S104, performs hierarchical and categorized differential processing on the image and text interference region to replace the problem of key information loss or geometric breakage caused by traditional denoising methods.

[0046] Figure 4 This is a flowchart illustrating a differential processing strategy for regions with textual and image interference, according to one embodiment of this application. Figure 4As shown, in some embodiments, a differentiated processing strategy is implemented for regions with text and image interference, including: Step S402: For regions whose region category is text or label category, compare the semantic importance score with the semantic importance threshold; For example, the specific value of θ should be adjustable according to the drawing specifications and application scenario. In this embodiment, θ is set to 0.3.

[0047] Step S404: If the semantic importance score is lower than the semantic importance threshold, then remove the region. like If the value is less than θ, the region is determined to be noise that can be removed. The noise is then removed directly and quickly repaired by means of background texture sampling, mean filling, or copying of surrounding pixels, thereby reducing unnecessary computational overhead.

[0048] like If the value is ≥θ, then it is further determined whether the area is spatially connected to geological boundary elements. Based on the spatial connection and area category, attribute preservation or image-text separation repair is performed.

[0049] Step S406: If the semantic importance score is higher than or equal to the importance threshold, and the area is spatially connected to the geological boundary elements, then perform image-text separation and repair operations on the spatially connected area. Spatial adhesion can be determined using methods such as mask dilation and overlap, distance thresholding, connected component merging, contour intersection detection, and pixel-level mask multiplication. These adhesion determination methods are all conventional techniques in image processing and fall within the scope of existing technology. In the specific implementation of this application, any known method or combination thereof can be flexibly selected to complete the adhesion determination based on the image resolution and cartographic specifications. This application will not elaborate on these methods in detail.

[0050] Step S408: If the semantic importance score is higher than or equal to the importance threshold, and the area is not spatially connected to the geological boundary feature, and the area category belongs to geological indicator label, then the area is retained and its text content is used as the attribute field of the corresponding vector feature.

[0051] If the region category belongs to a geologically indicative label such as occurrence symbol or stratigraphic code, then the region will be preserved but not vectorized.

[0052] Specifically, the pixels in this area are preserved at the image level and do not participate in the subsequent geometric vectorization process. Instead, their text content (such as ∠30°, J, γ, etc.) is stored as attribute fields of the corresponding vector elements, achieving separation of image and text information without loss of semantics.

[0053] For regions with semantic importance scores higher than or equal to the threshold θ and spatially connected to geological boundary elements, image-text separation and restoration operations are required. While removing text pixels, the occluded geological boundary is reconstructed based on prior geological knowledge, ensuring that the restored image maintains continuity and consistency with the surrounding unoccluded areas in terms of geometric shape and texture features, providing high-quality, interference-free input for subsequent vectorization extraction.

[0054] Figure 5 This is a flowchart illustrating the image-text separation and repair operation according to one embodiment of this application. Figure 5 As shown, in some embodiments, the image-text separation and repair operation includes: Step S502: Based on the region category, binarize the pixels belonging to the text or label category in the spatially connected region to obtain the text region mask; For example, setting the text region mask to .in Represents image region The pixels identified as needing processing. =0 indicates a non-text area (including geological boundaries, background textures, etc.).

[0055] Step S504: Obtain the predicted trajectory of the geological boundary obscured by the text region mask. ; Within the text masking area, the spatial orientation and geometric shape of the masked geological boundary are inferred, providing structured geometric prior constraints for generative restoration and preventing the restoration results from deviating from the true geological laws.

[0056] Based on the geological structure context of the unmasked area surrounding the mask, a large geological multimodal model or other model with trajectory prediction capabilities outputs the implicit curve parameters of the masked boundary. The predicted trajectory of the boundary within the mask is then determined. Normalized parameter curve: (8) When the input parameter t is given, the output is the two-dimensional coordinates (x, y) of that position on the image. in, This represents the normalized position along the trajectory, independent of the actual pixel length of the mask, facilitating unified model processing; the output coordinates (x, y) ∈ It is a two-dimensional real number, supporting sub-pixel precision; The control point sequence of the model output trajectory A continuous trajectory is generated by interpolating m-order Bézier curves: (9) (10) in, This represents the k-th control point; Represent the Bernstein basis functions; To ensure a smooth geometric transition between the repaired boundary and the unmasked area outside the mask, at the mask boundary... and Apply directional derivative constraints at: , (11) in, This represents the tangent direction vector extracted from the unoccluded area around the mask, ensuring that the repair trajectory achieves geometric continuity at the boundary. continuous).

[0057] Step S506: Generative texture inpainting is performed based on the local image of the spatially contiguous region, the text region mask, and the predicted trajectory to obtain the repaired geological image; wherein, the generative texture inpainting satisfies the geological prior constraints.

[0058] A Geo-Inpainting Network is constructed, based on a U-Net or diffusion model, and incorporating geological prior constraints to achieve controlled, geologically compliant generative restoration. The input to the Geo-Inpainting Network consists of a partial image of a geological map obscured by text and a mask of the text region. and predicted trajectory The output is a restored local image, in which the occluded geological boundaries are reconstructed in a continuous and smooth manner.

[0059] For example, let the image to be repaired be I, and the missing pixels in the mask region be denoted as: I masked =I⊙M text (12) Among them, I masked ⊙ represents pixels within the mask area that are obscured by text; ⊙ represents the Hadamard product.

[0060] Image I to be repaired and text region mask Pixel-by-pixel multiplication extracts pixels obscured by text within the masked area. The goal of the restoration is to generate a restored geological image. ,make In the non-masked area, it is completely consistent with the image to be repaired (I), and in the masked area, it conforms to the geological structure.

[0061] A diffusion-based restoration method is employed. The forward diffusion process gradually adds Gaussian noise to the clear image. (13) (14) in, Indicates diffusion time step Random noise under, T represents the number of sampling steps; This represents the retention factor for the original image. Indicates the noise addition factor; Represents the original, clear image; This represents random Gaussian noise; Indicates the preset noise scheduling parameters, s ; Represents the identity matrix. This process will restore the original sharp image. The gradual perturbation is pure random noise .

[0062] The reverse generation process uses text region masks. and predicted trajectory As a conditional constraint: (15) in, Represents a multivariate normal distribution; This represents the conditional mean predicted by the neural network, used to recover a clear restored image from noise. ; Represents the trainable weight parameters of the geological generation and repair network; Let represent the noise variance at step t, which is a preset hyperparameter and is related to forward diffusion; That is, given the noisy image I at step t... t Text area mask M text and predicted trajectory Under the conditions, the first Step Image It follows a normal distribution. It follows a matrix with mean μ and covariance matrix of... Gaussian distribution; Compared with the traditional unconditional diffusion model In comparison, this application will... and The injection conditions distribution enabled controllable restoration guided by geological laws.

[0063] Geological prior constraints include: strike continuity constraints, which are used to ensure that the trajectory of the restored geological boundary remains consistent with the predicted trajectory; topological consistency constraints, which are used to ensure that the topological relationship between the restored area and the surrounding unobstructed areas remains consistent; and lithological periodic constraints, which are used to ensure that the periodic texture characteristics of the lithological pattern area are consistent.

[0064] For example, to ensure that the repair results conform to the geological structure, the following loss term is introduced during the training phase: Towards continuous loss : Constraint restoration based on the restored geological image The actual trajectory extracted With predicted trajectory Maintain consistency.

[0065] (16) Topology consistency loss The local topological relationship between the constrained repair area and the surrounding unobstructed area remains unchanged.

[0066] (17) in, Indicates the mask area; This represents a reference structure image extracted from the periphery of the mask; Represents the restored geological image In the image area Local topological descriptors; Indicates the reference structure image in the image region Local topological descriptors; Lithological periodic loss (For lithological patterned areas): The constrained repair results conform to the periodic texture pattern.

[0067] Let the spatial frequency of the pattern be... Period T = 2π / ω, template function ;but: (18) in, Indicates the area of ​​the pattern; Indicates periodic translation; Represents a two-dimensional integer lattice space; Represents image region Repaired geological images ; The total repair loss function is calculated using the following formula. : (19) in, Represents the total repair loss function; This represents the mean square error within the masked area; , , This represents the weighting coefficients, used to characterize the relative importance of the three types of geological prior constraints. These coefficients can be adapted and adjusted for different map types (structural maps, mineral maps, and engineering geological maps). Total repair loss function. This value is used to measure the overall difference between the restored image and the ideal restoration result. The smaller the value, the more the restoration result conforms to the geological structure.

[0068] (20) in, It is used to measure the difference between the pixel value of the repaired area and the real value. When there is a real label, it supervises the pixel fidelity. In unsupervised scenarios, it can be replaced by adversarial loss or perceptual loss. This represents the original, complete image that is not occluded. This represents the Frobenius norm of the matrix.

[0069] Towards continuous loss This forces the actual boundary in the repair results to conform to the geometric trend predicted by the large model. Topological consistency loss. Ensure that the repaired boundary maintains the correct topological relationships with the surrounding area, such as intersection, connection, or parallelism. Lithological periodic loss. By using periodic self-similarity constraints, the restored lithological patterns maintain a periodic pattern consistent with the surrounding unobstructed areas.

[0070] Step S108 converts the repaired geological image into editable and analyzable vector geological data, realizing the geometric extraction from geological raster image to vector features. At the same time, the text content identified and retained in step S104 is automatically associated with the attribute fields of the corresponding vector features, and finally outputs standard vector data format with attributes.

[0071] Vectorization extraction transforms the restored geological image into vector geometric elements composed of nodes and edges, accurately extracting the geometric trajectories of linear elements such as geological boundaries and fault lines, providing quantifiable spatial data for subsequent GIS analysis and engineering applications; specifically: Edge detection (using the Canny operator) is performed on the restored geological image, converting it into a binary edge map. Each edge pixel in the binary edge map is treated as a graph node, and the connections between nodes are defined using eight-neighbor connectivity to construct the graph structure. : (twenty one) in, Represents the set of edge pixel nodes. This represents the set of eight neighboring connected edges.

[0072] The line segment extraction problem is transformed into a maximum weight path cover problem on a graph. Under the path constraint that the node degree is no greater than 2 (ensuring the extracted results are continuous line segments rather than branching networks), the goal is to find the subset of edges that maximizes the total edge weight. : (twenty two) The degree of each node in path st is ≤2 (path constraint); Where 'e' represents an edge connecting two adjacent pixel nodes, which is a single edge; This represents the set of edges for all extracted line segment paths; The edge weight is determined by the product of the edge strengths between adjacent pixels. (twenty three) Among them, edge weight This is used to characterize the joint probability estimate of two adjacent pixels belonging to the same geological boundary. The higher the weight, the more likely the edge is to belong to the real geological boundary. and This represents two adjacent pixels in a binary edge map; edge_strength(⋅) represents the confidence level that the pixel belongs to a geological boundary, which can be determined by the output of the edge detection algorithm or the probability map directly generated by the repair network.

[0073] A depth-first search (DFS) method is used to traverse all connected components, extract the longest path in each component, and output as a linear vector feature. : (twenty four) in, This represents the coordinates of the vertex of the polyline.

[0074] The highly important geological markers (such as occurrence symbols and stratigraphic codes) identified and retained in step S104 are automatically mapped to the attribute fields of the corresponding vector elements, realizing the integrated extraction of geometry and attributes and avoiding manual secondary data entry.

[0075] Text sequence vectors identified from image regions Perform semantic parsing and convert it into key-value pairs of attribute names and attribute values. : (25) For example, the text ∠30° is parsed to generate a key-value pair (attitude, ∠30°), and the text J is parsed to (stratigraphic designation, J).

[0076] For each extracted vector feature Define its spatial bounding box Place objects whose distance from the bounding box is less than the spatial neighborhood threshold. Furthermore, the text content of an image region whose region category belongs to the text or annotation category is associated with the attribute fields of that vector feature: (26) in, Represents a set of attributes; Represents the image region p to its spatial bounding box. Minimum Euclidean distance; spatial neighborhood threshold The setting is adaptive according to the map sheet scale, with a typical value of 0.5% of the map sheet width; This represents a key-value pair.

[0077] The extracted linear / area vector features and their associated attribute sets are output as standard Shapefile (SHP) format data. Each feature contains geometric information (vector features). ) and attribute set And retain the real spatial coordinate reference system registered in step S102 to ensure that vector data can be directly loaded into the GIS platform for spatial analysis and engineering applications.

[0078] In some embodiments, after vectorization output is completed, a multi-dimensional quality assessment system covering geometric accuracy, topological relationships, and semantic integrity is constructed to calculate a comprehensive quality score to determine whether the output meets the standards. If the comprehensive quality does not meet the preset standards, the key parameters of the front-end processing are automatically adjusted, and differentiated image processing and vectorization extraction are re-executed, forming a closed-loop mechanism of evaluation-feedback-optimization, thereby improving the method's generalization adaptability and output stability to stratigraphic maps from different sources and of different map types. Specifically, this includes: Construct a multi-dimensional quality assessment index: quantify the deviation between the automatically extracted results and expert reference annotations from three dimensions: geometric completeness, topological accuracy, and attribute retention rate, so as to provide a quantifiable decision basis for subsequent adaptive parameter optimization.

[0079] For each reference boundary in the validation set, calculate the proportion of the arc length of the overlapping segment between the automatically extracted boundary and the reference boundary to the total arc length of the reference boundary, and then average this proportion across all reference boundaries. The geometric completeness is then calculated using the following formula. : (27) in, Indicates geometric completeness; Indicates the arc length of the curve; This represents the corresponding boundary line automatically extracted from the k-th line; This represents the reference boundary line of the kth expert annotation; Indicates the total number of limits.

[0080] Evaluate the consistency between automatically extracted vector features and expert annotations regarding spatial topological relationships (such as intersection, connection, parallelism, etc.). Define a set of topological relationships. Connect, tangent, parallel, unrelated The topology accuracy is calculated using the following formula. : (28) N pairs =M*(M-1) / 2 (29) in, Indicates topology accuracy; This represents the total number of feature pairs, that is, the number of all unordered pairs (i,j) in the extracted vector feature set that require comparison of topological relationships; This indicates an indicator function that outputs 1 if the condition within the parentheses is true, and 0 otherwise. This represents the topological relationship between automatically extracted feature i and feature j; This represents the topological relationship between feature i and feature j manually labeled by geological experts on the validation dataset; M represents the total number of vector features.

[0081] Topological relationships The specific determination can be based on geometric features such as the distance between feature endpoints, the minimum Euclidean distance, and the directional angle, which are common knowledge in GIS spatial topology analysis and will not be elaborated in detail in this application.

[0082] This measure assesses the proportion of key geological labels (such as attitude symbols and stratigraphic codes) successfully associated with their corresponding vector feature attribute fields. It involves counting the total number of actual key geological labels in the map sheet and calculating the percentage of those successfully associated. The attribute retention rate is then calculated using the following formula. : (30) in, Indicates the attribute retention rate; This indicates the total number of key geological markers that actually exist in the map sheet; Indicates the first Whether a label has been successfully associated with the attribute table of a vector feature, if the association is successful. .

[0083] By integrating three-dimensional metrics, a single scalar is generated to determine whether the overall vectorized output quality meets application requirements. The comprehensive quality score is calculated using the following formula: (31) (32) in, Indicates the overall quality score; , , These represent the quality weight coefficients corresponding to the respective indicators.

[0084] Adaptive iterative optimization: When the overall quality score fails to meet the standard, key processing parameters are automatically adjusted and the core process is re-executed, achieving quality improvement and parameter self-calibration without human intervention.

[0085] when (The overall quality score threshold) triggers adaptive parameter updates. The adjustable parameter set Θ includes: semantic importance threshold. Number of sampling steps in the diffusion model and spatial neighborhood threshold .

[0086] The parameter adaptive update employs a gradient-based heuristic update: (33) in, This indicates the updated parameters; This indicates the parameters before the update; Indicates the learning rate; Indicates the overall quality score of the target; The gradient of the overall quality score with respect to the parameter set is approximated by the difference of historical evaluation results: (34) The update direction is automatically determined by the difference results: if increasing a parameter improves Q, then continue adjusting in that direction; otherwise, decrease it in the opposite direction. Repeat the differentiation strategy, image-text separation and repair, vectorization extraction and attribute association, and recalculate the overall quality score. until satisfied Or reach the preset maximum number of iterations .

[0087] In summary, this application has at least the following technical effects: First, this application achieves pixel-level semantic interpretation of elements in geological maps by constructing a large multimodal geological model and performing multi-task joint learning. It simultaneously outputs regional categories, text sequence vectors, and semantic importance scores, effectively overcoming the one-size-fits-all defect of traditional methods that simply treat all text pixels as noise. This provides a refined semantic decision-making basis for subsequent differentiated processing.

[0088] Second, this application implements graded treatment of areas with text-image interference based on semantic importance scores: low-importance legend text is directly cleaned and removed, while high-importance geological indicators such as occurrence symbols and stratigraphic codes are preserved and converted into vector attributes. For areas where text and geological boundaries are spatially connected, refined text-image separation and trajectory prediction repair are initiated. While thoroughly eliminating text-image interference, the loss of key geological information and the breakage of geometric topology are avoided.

[0089] Third, this application uses the predicted trajectory and boundary continuity parameters of the occluded geological boundary as geometric priors to inject into the generative repair network. Combined with multi-dimensional geological constraint losses such as orientation continuity, topological consistency and lithological periodicity, it guides the diffusion model or U-Net to reconstruct the texture of the occluded area in accordance with geological laws, which significantly improves the geometric continuity and structural fidelity of the repaired boundary and lays the image foundation for subsequent high-precision vectorization.

[0090] Fourth, this application establishes a text-image attribute association mechanism based on spatial proximity, which automatically maps geological semantic information such as stratigraphic codes and attitude angles identified from text regions into attribute fields of corresponding vector elements, generating standard SHP vector data with attributes. This realizes the direct conversion from raster images to integrated geometric and attribute vector results, significantly reducing the cost of manual post-entry data entry and directly supporting subsequent GIS spatial analysis and engineering applications.

[0091] Fifth, this application constructs a multi-dimensional quality evaluation index system covering geometric completeness, topological accuracy, and attribute retention rate. The output quality is automatically determined by the comprehensive quality score, and key parameters such as semantic importance threshold, repair sampling steps, and spatial neighborhood threshold are adaptively adjusted when the standard is not met. The processing-extraction-evaluation closed loop is executed iteratively, enabling the method to automatically adapt to multi-source stratigraphic maps of different cartographic specifications and sources, significantly improving the generalization ability and output stability of cross-map migration.

[0092] Example 2 Figure 6 This is a schematic diagram of a geological map interference cleaning and vectorization device based on multimodal large model semantic understanding, according to one embodiment of this application. Figure 6 As shown, this application provides a device for cleaning and vectorizing geological map textual interference based on multimodal large model semantic understanding, including: Preprocessing unit 610 is used to preprocess the geological raster image to be processed to obtain a standardized image; The semantic understanding unit 620 is used to perform semantic understanding on standardized images through a geological multimodal large model to obtain the region category, semantic importance score and text content of each image region; The differential processing unit 630 is used to perform differential processing strategies on regions with textual and image interference based on region category and semantic importance score, so as to obtain the restored geological image. The vector extraction unit 640 is used to extract vector data from the restored geological image and associate the text content with the attribute fields of the corresponding vector elements to generate attributed vector geological data.

[0093] In some embodiments, the differentiation processing unit 630 is further configured to: compare the semantic importance score with the semantic importance threshold for regions whose region category is text or label; if the semantic importance score is lower than the semantic importance threshold, remove the region; if the semantic importance score is higher than or equal to the importance threshold, and the region is spatially connected to the geological boundary feature, perform image-text separation and repair operations on the spatially connected region; if the semantic importance score is higher than or equal to the importance threshold, and the region is not spatially connected to the geological boundary feature, and the region category is geological indicator label, retain the region and use its text content as the attribute field of the corresponding vector feature.

[0094] In some embodiments, the vectorization extraction unit 640 is further configured to: convert the restored geological image into a binary edge map, and construct a graph structure with edge pixels as nodes and edges defined by eight-neighbor connectivity; under path constraints where the node degree is no greater than 2, find the maximum weight path based on the edge weight and extract linear vector features; wherein the edge weight is determined by the product of the edge strengths between adjacent pixels.

[0095] In some embodiments, the vectorization extraction unit 640 is further configured to: determine the spatial bounding box of each vector feature; and after semantic parsing, convert the text content of image regions whose distance from the spatial bounding box is less than the spatial neighborhood threshold and whose region category belongs to the text or label category into key-value pairs of attribute names and attribute values, and write them into the attribute field of the corresponding vector feature.

[0096] In some embodiments, the apparatus further includes: an adaptive optimization unit 650, configured to acquire geometric integrity, topological accuracy and attribute retention rate based on attributed vector geological data, and generate a comprehensive quality score based on geometric integrity, topological accuracy and attribute retention rate; If the overall quality score is lower than the preset threshold, the semantic importance threshold, the number of sampling steps for generative texture inpainting, and / or the spatial neighborhood threshold will be adaptively adjusted, and the differential processing strategy and vectorization extraction will be re-executed until the overall quality score meets the preset threshold or the maximum number of iterations is reached.

[0097] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the device described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0098] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the application concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

[0099] It should be understood that the order of the steps in the application content and embodiments of this application does not necessarily imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

Claims

1. A method for cleaning and vectorizing textual and image interference in geological maps based on multimodal large-scale model semantic understanding, characterized in that, Includes the following steps: The geological raster image to be processed is preprocessed to obtain a standardized image; Semantic understanding of the standardized images is performed using a geological multimodal large model to obtain the region category, semantic importance score, and text content of each image region; Based on the region category and the semantic importance score, a differentiated processing strategy is applied to the regions with textual and image interference to obtain the restored geological image; The restored geological image is vectorized and extracted, and the text content is associated with the attribute fields of the corresponding vector elements to generate attributed vector geological data.

2. The method for cleaning and vectorizing textual and image interference in geological maps based on multimodal large model semantic understanding as described in claim 1, characterized in that, The geological multimodal large model includes: a visual encoding module, a cross-modal alignment module, and a language decoding module; The visual encoding module is used to extract visual features from the standardized image to obtain image features; The cross-modal alignment module is used to map the image features to the feature space of the language decoding module to obtain a joint feature representation; The language decoding module is used to perform semantic decoding on the joint feature representation to obtain the region category, the semantic importance score, and the text sequence vector corresponding to the text content for each image region.

3. The method for cleaning and vectorizing geological map textual interference based on multimodal large model semantic understanding according to claim 2, characterized in that, The geological multimodal large model was obtained through the following steps: Based on the region category prediction probability distribution and the corresponding true category label of each image region, calculate the cross-entropy classification loss; The mean squared error of the importance score is calculated based on the semantic importance score and the corresponding true importance score of each image region. Based on the text sequence vector, calculate the character recognition sequence loss; The joint loss function is obtained by weighted summing of the cross-entropy classification loss, the mean square error of the importance score, and the character recognition sequence loss; The parameters of the geological multimodal large model are adjusted with the goal of minimizing the joint loss function.

4. The method for cleaning and vectorizing geological map textual interference based on multimodal large model semantic understanding according to claim 1, characterized in that, Implement differentiated processing strategies for areas with text and image interference, including: For regions whose region category is text or label, compare the semantic importance score with the semantic importance threshold; If the semantic importance score is lower than the semantic importance threshold, then the region is removed. If the semantic importance score is higher than or equal to the importance threshold, and the region is spatially connected to the geological boundary elements, then the image-text separation and repair operation is performed on the spatially connected region. If the semantic importance score is higher than or equal to the importance threshold, and the region is not spatially connected to the geological boundary feature, and the region category belongs to geological indicator labeling, then the region is retained and its text content is used as the attribute field of the corresponding vector feature.

5. The method for cleaning and vectorizing textual interference in geological maps based on multimodal large model semantic understanding according to claim 4, characterized in that, The image-text separation and repair operation includes: Based on the region category, the pixels belonging to the text or label category in the spatially connected region are binarized to obtain the text region mask; Obtain the predicted trajectory of the geological boundary obscured by the text region mask; Based on the local image of the spatially contiguous region, the text region mask, and the predicted trajectory, generative texture inpainting is performed to obtain the repaired geological image.

6. The method for cleaning and vectorizing textual interference in geological maps based on multimodal large model semantic understanding according to claim 5, characterized in that, The generative texture restoration satisfies geological prior constraints, which include: The continuity constraint is used to ensure that the restored geological boundary trajectory remains consistent with the predicted trajectory. Topology consistency constraints are used to ensure that the topological relationship between the repaired region and the surrounding unobstructed regions remains consistent. Lithological periodicity constraints are used to constrain the consistency of periodic texture features in lithological pattern regions.

7. The method for cleaning and vectorizing textual and image interference in geological maps based on multimodal large model semantic understanding according to claim 1, characterized in that, The vectorization extraction includes: The restored geological image is converted into a binary edge map, and a graph structure is constructed with edge pixels as nodes and edges defined by eight-neighbor connectivity. Under the path constraint that the node degree is no greater than 2, the maximum weight path is found based on the edge weight and the linear vector features are extracted; The edge weight is determined by the product of the edge strengths between adjacent pixels.

8. The method for cleaning and vectorizing textual and image interference in geological maps based on multimodal large model semantic understanding according to claim 1, characterized in that, Associating the text content with the attribute fields of the corresponding vector features includes: Determine the spatial bounding box of each vector element; The text content of image regions whose distance from the spatial bounding box is less than the spatial neighborhood threshold and whose region category belongs to the text or label category is semantically parsed and converted into key-value pairs of attribute names and attribute values, and then written into the attribute fields of the corresponding vector features.

9. The method for cleaning and vectorizing textual interference in geological maps based on multimodal large model semantic understanding according to claim 1, characterized in that, The method further includes: Based on the attributed vector geological data, geometric integrity, topological accuracy, and attribute retention rate are obtained, and a comprehensive quality score is generated based on the geometric integrity, topological accuracy, and attribute retention rate. If the overall quality score is lower than a preset threshold, the semantic importance threshold, the number of sampling steps for generative texture inpainting, and / or the spatial neighborhood threshold are adaptively adjusted, and the differential processing strategy and the vectorization extraction are re-executed until the overall quality score meets the preset threshold or the maximum number of iterations is reached.

10. A device for cleaning and vectorizing textual interference in geological maps based on multimodal large model semantic understanding, characterized in that, include: The preprocessing unit is used to preprocess the geological raster image to be processed to obtain a standardized image; The semantic understanding unit is used to perform semantic understanding on the standardized image through a geological multimodal large model to obtain the region category, semantic importance score and text content of each image region; A differential processing unit is used to perform a differential processing strategy on regions with text-image interference based on the region category and the semantic importance score, so as to obtain the restored geological image. The vectorization extraction unit is used to extract vectorization data from the restored geological image and associate the text content with the attribute fields of the corresponding vector elements to generate attributed vector geological data.

Images