Geological environment information intelligent extraction and identification method based on GCR model

Abstract

The invention relates to the technical field of intelligent geological information processing, and discloses an intelligent geological environment information extraction and recognition method based on a GCR model, and the method comprises the steps: building a feature index set through multi-source data standardization, integrating a geological feature enhancement module in an OCR frame for multi-stage training, and achieving the recognition of stratum features through the combination of morphology and a color clustering algorithm. And dynamic parameter optimization is implemented based on difference analysis to form a complete technical closed loop from data preprocessing to model self-optimization. Model overfitting is effectively inhibited through a dynamic training control mechanism, multi-modal fusion analysis of geologic symbols, colors and spatial features is realized in combination with multi-head attention, differential parameter optimization is implemented based on the importance of the geologic features, and while the recognition precision of complex geologic elements is improved, the recognition efficiency of the geologic elements is improved. The adaptive capacity of the model to non-uniformly distributed data is remarkably enhanced, and the whole-process improvement from feature learning to training optimization is formed.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent processing of geological information, and specifically to an intelligent extraction and recognition method for geological environment information based on the GCR model. Background Technique

[0002] With the in-depth promotion of the new urbanization strategy, geological environment elements have become the basic decision-making basis for urban planning and construction. As the core carrier for recording stratum structures and lithological characteristics, the digital analysis efficiency of geological maps directly affects the safety assessment accuracy of urban underground space development. Currently, the industry generally uses a method combining OCR technology with traditional image processing for geological element extraction. However, in the face of complex geological symbol systems and multi-source heterogeneous data characteristics, the existing technical system exposes obvious adaptability bottlenecks.

[0003] There are systematic defects in the existing technical solutions when realizing intelligent analysis of geological elements: the traditional OCR model is limited by the single-modal analysis mechanism of text features and is difficult to effectively capture the correlation features between geological symbols and spatial attributes, resulting in a coupling error between symbol recognition rate and stratum boundary extraction accuracy; although the deep learning-based solutions have made breakthroughs in the depth of feature learning, the parameter optimization strategy in the training process lacks specific design for geological data and is prone to overfitting due to uneven sample distribution; in addition, the existing methods are not robust enough to the noise interference of scanned documents and are prone to feature misjudgment when processing historical paper maps. These problems jointly restrict the digital conversion efficiency of geological environment information and are difficult to meet the needs of multi-dimensional data fusion analysis for urban geological safety assessment.

[0004] Taking the metamorphic rock area in the northeastern margin of the Yangtze Platform as an example, the complex stratum contact relationships and multi-phase tectonic traces in this area pose higher requirements for feature extraction algorithms. When the existing technologies process such geological elements with strong spatial correlation, they often lack a cross-modal feature fusion mechanism, resulting in a decoupling error between lithological symbol recognition and stratum boundary division. This technical defect not only affects the generation efficiency of geological survey results but may also cause the lack of key data for urban underground space development risk assessment. Therefore, developing an intelligent analysis technology for geological information with adaptive feature fusion ability has become the key path to breaking through the current digital bottleneck of geological surveys. Summary of the Invention

[0005] Aiming at the deficiencies of the existing technologies, the present invention provides an intelligent extraction and recognition method for geological environment information based on the GCR model, which solves the problems of multi-modal feature decoupling, strong noise sensitivity, and model overfitting existing in the digitalization process of traditional geological maps.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a method for intelligent extraction and identification of geological environment information based on the GCR model, comprising the following steps: S1. Data standardization: Perform format conversion and feature extraction on multi-source geological data to construct a feature index set including symbols, colors and geometric features; S2. Model construction and training: Based on the feature index set, a geological feature enhancement module is integrated into the OCR framework, and an optimized model is generated through multi-stage enhancement training; S3. Stratum feature identification: applying the optimization model, combining morphological operations and color clustering algorithms to output stratigraphic boundaries and lithologic attributes; S4. Dynamic model optimization: Based on the difference analysis between the recognition results and the verification indicators, the model parameters are updated through feature fusion and early stopping strategy.

[0007] Preferably, the data standardization process in step S1 includes: Convert scanned geological maps into standard image formats and perform denoising; Verify the coordinate system and metadata of GIS files; Establish a mapping relationship library between geological symbols and lithologic properties.

[0008] Preferably, in the model construction and training of step S2: The geological feature enhancement module includes a convolution optimization layer and a geological rule verification unit; The multi-stage enhancement training consists of three stages of gradually increasing the complexity of image noise.

[0009] Preferably, the geological rule verification unit is configured as follows: When the sequence of layers violates the geological time scale, a correction mechanism is triggered; Detect symbolic logic conflicts and perform rule-based error correction.

[0010] Preferably, the formation feature identification in step S3 includes: S3-1. Boundary extraction: Based on morphological gradient operation, a combination of a 5×5 rectangular kernel and a 7×7 elliptical kernel is used to extract the initial stratum boundary line; S3-2, color clustering: performing K-means clustering on the pixel points within the bounding area of the dividing line in the LAB color space to generate a stratum main color classification result; S3-3, noise filtering: performing spatial continuity analysis on the classification results, and removing abnormal pixels whose discreteness exceeds a preset spatial distribution density threshold; S3-4, logical verification: matching the filtered classification results with the symbol information recognized by OCR in step S2 for spatial position and semantic consistency.

[0011] Preferably, in the dynamic optimization of the model in step S4: The early stopping strategy is automatically triggered by monitoring the fluctuation of the F1 value of the validation set; The feature fusion strategy dynamically adjusts the weight ratio of text features and color features.

[0012] Preferably, the three stages in the multi-stage enhanced training include: In the first stage, Gaussian noise is added to simulate low-quality scanned images; In the second stage, motion blur is applied to simulate the folding deformation of geological maps; In the third stage, downsampling processing is performed to improve the generalization ability of the model.

[0013] Preferably, the triggering condition of the early stopping strategy satisfies the following formula: ; Wherein, is the proportion of the number of iterations with performance degradation, is the number of samples in the validation set, represents the F1 value of the validation set at the th iteration, is an indicator function, and when > 0.7 in five consecutive iterations, the parameter rollback mechanism is triggered.

[0014] Preferably, the parameter rollback mechanism updates the model parameters using the adaptive momentum algorithm: ; In the formula, is the model parameter at the th iteration, is the adaptive learning rate factor, which decays dynamically according to , and are updated according to the following rules respectively: ; ; Wherein, , are momentum decay coefficients, is a numerical stability constant, is the current parameter gradient.

[0015] The present invention provides an intelligent extraction and recognition method for geological environment information based on the GCR model. It has the following beneficial effects: 1. The present invention realizes real-time optimization and adjustment of the model training process through the linkage between the difference analysis module and the early stopping determination mechanism. Compared with the traditional training method with a fixed number of iterations, this solution can autonomously identify the overfitting trend and adjust the learning process, solving the problem of insufficient stability in the learning of complex geological features by the model.

[0016] 2. The present invention uses a multi-head attention mechanism to integrate geological symbols, color attributes, and spatial features, establishing cross-modal feature associations. Compared with the OCR model that processes text or images alone, this design enhances the collaborative parsing ability of multiple elements of geological maps and improves the integrity of feature recognition in complex scenarios.

[0017] 3. The present invention is based on a momentum adjustment algorithm for the importance of geological elements to achieve differential learning of different feature channels. The traditional uniform parameter update method is difficult to adapt to the non-uniform distribution characteristics of geological data. This solution improves the model convergence efficiency and feature learning accuracy by dynamically adjusting the parameter update amplitude. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a flowchart of the method of the present invention; Figure 2 is a flowchart of key feature image recognition of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0020] Please refer to the attached Figure 1 , the embodiment of the present invention provides an intelligent extraction and recognition method for geological environment information based on the GCR model, including the following steps: S1. Data standardization processing: Perform format conversion and feature extraction on multi-source geological data to construct a feature index set including symbols, colors, and geometric features; The data standardization processing is jointly realized by a scan conversion unit, a coordinate verification unit, and a symbol mapping unit. The scan conversion unit is connected to the image input interface and is configured to execute the following processing flow: First, convert the scanned geological map into a standard image format, and use a bilateral filter for noise elimination. Its pixel update formula is: ; In the formula, is the spatial weight kernel function, is the value range weight kernel function, and the filter window size k is dynamically adjusted according to the image resolution. is the pixel value of the input image at the coordinate , and is the difference in grayscale value between the neighboring pixel and the central pixel. Subsequently, morphological opening operation is performed, and the structuring element uses a rectangular kernel.

[0021] The coordinate verification unit is connected to the GIS data interface and is configured to perform coordinate system conversion, and solve the conversion parameters by the least squares method: ; where and are the source coordinate system coordinates of the th control point, and are the target coordinate system coordinates of the th control point, are the elements of the affine transformation parameter matrix, and is the number of control points.

[0022] The number of control points meets the requirements of the preset threshold. The parameter matrix is solved by the matrix decomposition method, and the plane residuals conform to the engineering standard tolerances. The converted spatial data is written into the spatial database, and metadata integrity verification is performed.

[0023] The symbol mapping unit includes a feature extraction subunit and a relationship construction subunit. The feature extraction subunit uses a target detection model to locate the geological symbol area and outputs the symbol boundary coordinates. The color feature calculates the histogram distribution in the standard color space: ; where is the color component value at the pixel (x, y), is the reference value of the th color interval, and is the impulse function. [[ID=5**]]

[0024] The geometric feature calculates the compactness parameter: ; where is the pixel area of the geological symbol area, and is the boundary perimeter of the geological symbol area.

[0025] The relationship construction subunit establishes an index relationship between the multi-dimensional feature vector and the lithology coding, and the feature dimension conforms to the preset engineering specifications.

[0026] Example: When processing scanned geological maps, first perform resolution standardization to the engineering requirement value, and use bilateral filtering to eliminate scanning noise. For coordinate transformation, select control points that meet the specification requirements and calculate the affine transformation parameters. For symbol detection, use a preset confidence threshold, and establish a mapping relationship based on the similarity algorithm for feature matching.

[0027] S2. Model construction and training: Based on the feature index set, integrate a geological feature enhancement module in the OCR framework, and generate an optimized model through multi-stage enhancement training; The model construction and training are realized through the cooperation of the feature enhancement module and the multi-stage training framework. The feature enhancement module includes a convolution optimization part and a rule verification part. The convolution optimization part is configured to perform deformable convolution operations: ; In the formula, is the learnable offset, is the preset convolution kernel size, is the feature value of the output feature map at position ; is the weight parameter of the th convolution kernel; is the input feature map tensor; is the preset spatial offset at the position in the standard convolution kernel. The rule verification part is connected to the timing analysis unit, and when the stratigraphic sequence violates the geological time scale, the correction function is triggered: Among them, is the stratigraphic timing constraint threshold, set according to the engineering standard, is the stratigraphic transition probability matrix, is the current detected stratigraphic age code, is the previous stratigraphic age code, is the stratigraphic timing transition probability, sourced from the geological time scale database.

[0028] The multi-stage enhancement training framework includes three levels of data enhancement pipelines. In the first stage, Gaussian noise is added to the input image, and the noise intensity conforms to the standard of the scanned image degradation model. In the second stage, a motion blur kernel is applied, and the parameter settings conform to the folding deformation characteristics of the geological map. In the third stage, downsampling processing is performed, and the sampling rate is set according to the engineering specifications.

[0029] Example: When training the GCR model, clean samples, noise samples, and downsampled samples are loaded in stages. The rule verification unit loads the standard geological time scale database, and automatically triggers the correction mechanism when the stratigraphic sequence is inverted. The convolution optimization layer adopts a progressive parameter unfreezing strategy, and freezes the weights of the bottom network in the initial stage.

[0030] Please see the attached Figure 2 ,S3, Stratum feature recognition: Apply the optimization model, combine morphological operations and color clustering algorithm to output stratigraphic boundaries and lithologic attributes; Stratum feature recognition is achieved through a multi-stage processing pipeline, including a boundary extraction module, a color clustering module, a noise filtering module, and a logic check module. The boundary extraction module is configured to perform morphological gradient operations: ; Where, is a rectangular structural element, It is an elliptical structural element, and its size parameters are dynamically adjusted according to the scale of the geological map. is the input geological image matrix, is a morphological dilation operation used to expand bright areas. This is a morphological erosion operation used to shrink bright areas. The extracted boundary lines are optimized by the edge detector and then output to the color analysis unit.

[0031] The color clustering module is connected to the color space conversion component to perform K-means clustering on the bounding area of the dividing line: ; in, Pixel The feature vector in LAB color space, No. The characteristic vectors of cluster centers, is the number of dynamic clusters, which is automatically determined by the regional lithology type. Belong to A set of pixels that are clustered.

[0032] The number of clusters is automatically determined based on the regional lithology, and an improved sampling strategy is used for initialization. The main color classification results are transmitted to the verification unit after spatial distribution density analysis.

[0033] The noise filtering module is configured to calculate the pixel neighborhood connectivity index: ; in, Pixel The number of similar pixels in the eight-neighborhood, is the total number of neighborhood pixels (fixed to 8), It is a spatial distribution density index that reflects the reliability of pixel attribution.

[0034] When the indicator value falls below the preset threshold, the rejection mechanism is triggered. The spatial distribution density threshold is dynamically calculated based on regional characteristics and is positively correlated with the complexity of the formation.

[0035] The logic verification module connects to the OCR recognition database and performs spatial position matching: ; in, It is the overlapping area between the OCR recognition symbol area and the color cluster area. is the union area of the symbol area and the cluster area, It is a spatial semantic consistency indicator, and the threshold is set according to engineering standards.

[0036] If the matching degree meets the engineering standard, it is considered semantically consistent; otherwise, the review process is triggered. The verification results are linked to the formation attribute database and updated.

[0037] Example: When processing a typical geological map, stratigraphic boundaries are first extracted to generate topological boundaries. Adaptive clustering is then performed on the target area to identify key lithologic color features. A spatial filtering mechanism automatically eliminates discrete noise points, and a logic verification unit spatially associates color features with OCR-recognized symbols to correct three lithologic labeling conflicts.

[0038] S4. Dynamic model optimization: Based on the difference analysis between the recognition results and the verification indicators, the model parameters are updated through feature fusion and early stopping strategy.

[0039] The model dynamic optimization module includes a difference analysis unit, a feature fusion unit, and a parameter update unit. The difference analysis unit is configured to calculate the multi-dimensional differences between the recognition results and the verification indicators: ; Where, is the feature weight coefficient, is the distribution difference weight, KL divergence measures the predicted distribution and the true distribution The difference, For the Class feature verification index benchmark value, For the The actual recognition results of class features, is the distribution difference adjustment factor, which controls the KL divergence item weight, The model predicts the feature distribution probability, Geological prior knowledge feature distribution probability. Feature fusion unit connected to multi-head attention mechanism: ; in The projection matrices from geological features, symbol features, and color features, respectively, is the scaling factor.

[0040] The early stopping strategy module monitors the F1 value fluctuation of the validation set: ; wherein, is the proportion of the number of iterations with performance degradation, is the number of samples in the validation set, the F1 value of the validation set in the th iteration,

[0041] When the performance continuously degrades during iterations, the parameter rollback mechanism is triggered. The parameter update adopts the adaptive momentum algorithm: ; wherein, the model parameters in the th iteration, is the adaptive learning rate factor, is the first-order momentum estimator, is the second-order momentum estimator, is the numerical stability coefficient, and the momentum term and the second moment are updated exponentially, specifically according to dynamically decays, and are updated according to the following rules respectively: ; wherein, , are the momentum decay coefficients, is the numerical stability constant, is the current parameter gradient.

[0042] Example: When processing a geological map, the difference analysis module detects a sudden increase in the symbol recognition error rate, and the feature fusion unit automatically adjusts the geometric feature weights. After the early stopping strategy triggers parameter rollback, the model maintains a stable accuracy rate in the validation set.

[0043] This embodiment fully discloses the optimization mechanism through the above formula, and each module realizes state synchronization through the control bus. The multi-head attention mechanism improves the feature fusion accuracy, the adaptive momentum algorithm accelerates convergence, and the parameter rollback mechanism reduces the overfitting risk. The difference threshold is strictly aligned with the engineering standard to ensure that the dynamic optimization of the model complies with the geological feature recognition specification.

[0044] Test Case 1: In this experiment, 168 geological maps at a scale of 1:50,000 in the northeastern margin of the Yangtze Platform were selected as test samples, including 1,235 symbols (■) of the Permian Qixia Formation limestone and 892 symbols (▲) of the Triassic Yankeng Formation sandstone. The experimental environment was configured with an Intel Xeon Gold 6248R processor, an NVIDIA A100 graphics card, a Python 3.8 environment, and the PaddleOCR 2.5 framework.

[0045] Experimental method: Four groups of comparison models were established: Control group A: Tesseract 4.0 standard OCR model; Control group B: PaddleOCR v2.5 pre-trained model; Control group C: Customized geological model with ResNet-50 architecture; Experimental group: GCR model (integrated multi-head attention mechanism + dynamic early stopping strategy).

[0046] Evaluation metrics: Symbol recognition rate: , where is the number of correctly recognized geological symbols (such as limestone symbol ■ and sandstone symbol ▲), is the total number of geological symbols marked in the test sample, is the percentage of symbol recognition accuracy; Anti-noise F1 value: (adding Gaussian noise σ = 0.2), where is the precision, the proportion of correctly recognized symbols in high-noise data, is the recall, the proportion of actual symbols recognized in high-noise data, is the comprehensive evaluation metric; Training efficiency: Average time consumption per epoch Overfitting index: , where is the training set accuracy (e.g., the GCR model training set reaches 98%), is the validation set accuracy (e.g., the GCR validation set reaches 95.2%), is the quantization index of overfitting degree.

[0047] Experimental results: Model Symbol recognition rate High-noise F1 value Training time consumption Overfitting fluctuation Tesseract 4.0 34% 0.52 9 hours 30 minutes 10% PaddleOCR 63% 0.68 7 hours 50 minutes 8% ResNet 63% 0.73 9 hours 12 minutes 5% GCR 75% 0.81 5 hours 24 minutes ＜3% As shown in the above table, the GCR model performs significantly better than the control groups in key metrics: The symbol recognition rate is 41 percentage points higher than that of Tesseract (75% vs 34%) and 12 percentage points higher than that of ResNet; The F1 value reaches 0.81 in high-noise scenarios, a 19.1% improvement compared to PaddleOCR; The training time is 5h24m, a 42% reduction compared to ResNet, and the performance fluctuation of the validation set is <3%; The overfitting index is controlled at 2.8%, 5 - 8 percentage points better than the control group.

[0048] Typical embodiment: Taking the geological map of H50E003015 as an example, the processing flow of the GCR model is as follows: Extract the formation boundary through morphological gradient operation: ; Among them, is a 5×5 rectangular structuring element, is an elliptical element with a major axis of 15px, is the input geological image matrix; Use K-means clustering (k = 6) to segment the lithology area, and match the cluster centers with the symbol library; The dynamic early stopping mechanism is triggered at the 1,200th iteration and rolls back to the parameter snapshot at the 1,150th iteration. The final output contains the recognition results of 28 formations. The symbol-color matching accuracy rate reaches 92.3%, a 17.6 percentage point improvement compared to the ResNet model. This test case confirms that the GCR model has significant technical advantages in feature fusion accuracy, training stability, and anti-noise ability.

[0049] Test case two: Verification of the multi-stage training efficiency of the GCR model Experimental settings: Model architecture: Build a dual-module structure based on the PaddleOCR framework. The text detection module uses the DB algorithm (ResNet50 backbone network), and the text recognition module uses the CRNN + Attention mechanism Training data: Include 800 geological map pieces (including elements such as histograms and formation boundaries), and divide the training / validation / test sets according to 7:2:1 Hardware configuration: NVIDIA RTX3090 GPU, PyTorch 1.10 environment Experimental steps: Three-stage iterative training: Basic training stage (600 iterations): Fix the backbone network parameters, learning rate 0.001; Enhanced training stage (1200 iterations): Unfreeze the top layer network, and the learning rate is reduced to 0.0OO3; Fine-tuning stage (2000 iterations): All network parameters are adjustable, learning rate 0.0001.

[0050] Dynamic early stopping monitoring: Calculate the best evaluation index of the validation set every 100 iterations; Early stopping is triggered when the index fluctuates by < 0.5% for 5 consecutive times; Optimal model saving condition: the index improvement > 1%.

[0051] Experimental results: Evaluation metrics 600 iterations 1200 iterations 2000 iterations Best evaluation metrics 0.63 0.72 0.78 Precision 0.59 0.63 0.65 Recall 0.94 0.96 0.97 Inference speed (FPS) 0.12 0.15 0.16 Best training cycle 600 1200 1250 Training duration 2 hours 40 minutes 5 hours 24 minutes 11 hours 38 minutes Conclusion analysis: The model performance shows diminishing marginal benefits with the number of iterations: The increase rate of the best evaluation index decreases from 14.3% (0.63 → 0.72) in the range of 600 - 1200 iterations to 8.3% (0.72 → 0.78) in the range of 1200 - 2000 iterations; The improvement rate of accuracy narrows from 6.8% (0.59 → 0.63) to 3.2% (0.63 → 0.65); Analysis of the trigger of the dynamic early stopping mechanism: In the 2000 - iteration stage, the best training cycle is advanced to 1250 times (actually saving 37.5% of the iteration times); The comparison of training durations shows that it takes 11.6 hours to complete all 2000 iterations, while the early stopping mechanism only takes 7.2 hours at 1250 times (saving 38% of the time cost); Precision - recall balance verification: The recall rate is stable above 94%, indicating that the model has a strong ability to capture geological elements; The accuracy is improved by 6 percentage points (0.59 → 0.65), proving that the multi - stage training effectively reduces the false detection rate; The experimental data verifies the effectiveness of the dynamic early stopping mechanism in controlling overfitting and optimizing training resources, and at the same time proves that the multi - stage iteration strategy can significantly improve the recognition accuracy of the model for geological symbols.

[0052] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent extraction and recognition method for geological environment information based on the GCR model, characterized in that The following steps are involved: S1. Data standardization: Perform format conversion and feature extraction on multi-source geological data to construct a feature index set including symbols, colors and geometric features; S2. Model construction and training: Based on the feature index set, a geological feature enhancement module is integrated into the OCR framework, and an optimized model is generated through multi-stage enhancement training; S3. Stratum feature identification: applying the optimization model, combining morphological operations and color clustering algorithms to output stratigraphic boundaries and lithologic attributes; S4. Dynamic model optimization: Based on the difference analysis between the recognition results and the verification indicators, the model parameters are updated through feature fusion and early stopping strategy.

2. The intelligent extraction and recognition method of geological environment information based on the GCR model according to claim 1, characterized in that The data standardization process in step S1 includes: Convert scanned geological maps into standard image formats and perform denoising; Verify the coordinate system and metadata of GIS files; Establish a mapping relationship library between geological symbols and lithologic properties.

3. The intelligent extraction and recognition method of geological environment information based on the GCR model according to claim 1, wherein The S2 step model construction and training: The geological feature enhancement module includes a convolution optimization layer and a geological rule verification unit; The multi-stage enhancement training consists of three stages of gradually increasing the complexity of image noise.

4. The intelligent extraction and recognition method of geological environment information based on the GCR model according to claim 3, characterized in that, The geological rule verification unit is configured as follows: When the sequence of layers violates the geological time scale, a correction mechanism is triggered; Detect symbolic logic conflicts and perform rule-based error correction.

5. The intelligent extraction and recognition method of geological environment information based on the GCR model according to claim 1, characterized in that The formation feature identification step S3 includes: S3-1. Boundary extraction: Based on morphological gradient operation, a combination of a 5×5 rectangular kernel and a 7×7 elliptical kernel is used to extract the initial stratum boundary line; S3-2, color clustering: performing K-means clustering on the pixel points within the bounding area of the dividing line in the LAB color space to generate a stratum main color classification result; S3-3, noise filtering: performing spatial continuity analysis on the classification results, and removing abnormal pixels whose discreteness exceeds a preset spatial distribution density threshold; S3-4, logical verification: matching the filtered classification results with the symbol information recognized by OCR in step S2 for spatial position and semantic consistency.

6. The intelligent extraction and recognition method of geological environment information based on the GCR model according to claim 1, characterized in that In the dynamic optimization of the S4 step model: The early stopping strategy is automatically triggered by monitoring the fluctuation of the F1 value of the validation set; The feature fusion strategy dynamically adjusts the weight ratio of text features and color features.

7. The intelligent extraction and recognition method of geological environment information based on the GCR model according to claim 3, characterized in that, The three phases of the multi-phase enhanced training include: In the first stage, Gaussian noise is added to simulate low-quality scanned images; In the second stage, motion blur is applied to simulate the fold deformation of geological maps; The third stage performs downsampling processing to improve the generalization ability of the model.

8. The intelligent extraction and recognition method of geological environment information based on the GCR model according to claim 6, characterized in that, The triggering condition of the early stopping strategy satisfies the following formula: ； Among them, is the proportion of the number of iterations with performance degradation, is the number of samples in the validation set, represents the F1 value of the validation set at the -th iteration, is an indicator function, and when > 0.7 for five consecutive iterations, the parameter rollback mechanism is triggered.

9. The intelligent extraction and recognition method of geological environment information based on the GCR model according to claim 6, characterized in that, The parameter rollback mechanism uses an adaptive momentum algorithm to update model parameters: ； In the formula, is the model parameter of the -th iteration, is the adaptive learning rate factor, which decays dynamically according to , and are updated according to the following rules respectively: ； ； Among them, , is the momentum decay coefficient, is the numerical stability constant, is the current parameter gradient.

Images