A method for identifying and extracting image information from meteorological equipment

Abstract

This invention provides a method for identifying and extracting image information from meteorological equipment, belonging to the field of meteorological equipment information processing technology. The method includes: constructing a sample set of key areas of meteorological equipment; training a dedicated target detection model to locate key areas; preprocessing sub-blocks of key areas and performing coarse detection using an adapted text detection algorithm; optimizing the OCR recognition model based on a meteorological corpus and combining it with an error correction mechanism to obtain refined recognition results; and outputting target information based on a scene-predefined template through rule matching and NLP verification. This invention solves the problems of excessive irrelevant interference, low efficiency, poor accuracy, and weak scene adaptability in existing general OCR technologies, achieving efficient and accurate extraction of meteorological equipment information. It is applicable to various types of meteorological equipment and complex environments, possessing strong flexibility and scalability.

Description

Technical Field

[0001] This invention relates to the field of meteorological equipment information processing technology, and in particular to a method for identifying and extracting image information from meteorological equipment. Background Technology

[0002] In meteorological observation and maintenance, image information from meteorological equipment (such as sensor readings, equipment identification, and parameter configurations) is the core basis for meteorological data acquisition and equipment status monitoring. Currently, the industry mainly uses general OCR (Optical Character Recognition) technology to directly extract information from the entire image of meteorological equipment, but this method has the following significant drawbacks:

[0003] Significant interference from irrelevant information: Meteorological equipment images contain a large amount of content unrelated to the target information, such as the equipment's metal casing, the surrounding observation field fence, trees / buildings in the background, as well as environmental interference such as rain streaks in rainy weather, particle noise in dusty weather, and light spots at night. General OCR technology cannot distinguish between target information areas and irrelevant areas, easily misidentifying casing textures as text and rain streaks as numbers, leading to deviations in target information extraction.

[0004] Extremely low recognition efficiency: General-purpose OCR requires character detection and recognition for each pixel area of ​​the entire meteorological equipment image, while the target information area of ​​the meteorological equipment only accounts for 5% to 15% of the entire image (e.g., the reading area of ​​a temperature and humidity sensor is only 200×100 pixels). The recognition process for a large number of invalid areas wastes more than 90% of the computing resources, resulting in a recognition time of 200 to 500 ms for a single image, which cannot meet the requirement of "real-time data acquisition" in meteorological observation (requiring a recognition delay of <100 ms).

[0005] Insufficient accuracy: Key information from meteorological equipment exhibits unique forms and scene-dependent characteristics. For example, temperature and humidity readings from automatic weather stations are often displayed on LED digital screens, which are prone to glare and character breakage under strong light. Barometric pressure parameters on radar equipment are etched text next to the pointer, small in size and overlapping with the scale lines. Equipment identification tags may fade or wear due to prolonged outdoor use. General-purpose OCR is not optimized for these characteristics, achieving an accuracy rate of less than 60% for recognizing special symbols (°C, hPa) and a misrecognition rate as high as 40% for blurred / faded text.

[0006] Unable to adapt to customized scenarios: Meteorological observation involves various scenario requirements, such as real-time temperature and humidity monitoring requiring the extraction of temperature (XX.XX℃) and humidity (XX.XX%RH), equipment maintenance verification requiring the extraction of equipment number (QX-2024-001), and radar parameter calibration requiring the extraction of reflectivity factor (XXdBZ). General OCR can only output unstructured text strings and cannot filter and match target information according to scenario requirements, requiring secondary manual filtering and increasing maintenance costs.

[0007] Therefore, this invention proposes a method for identifying and extracting image information from meteorological equipment. Summary of the Invention

[0008] This invention provides a method for identifying and extracting image information from meteorological equipment, in order to solve the aforementioned technical problems.

[0009] This invention proposes a method for identifying and extracting image information from meteorological equipment, comprising:

[0010] Step 1: Construct a key area sample set for meteorological equipment, and use a deep learning target detection model to train a special target detection model for meteorological equipment using the sample set as training data. The sample set contains images of different types of meteorological equipment in various environmental scenarios, and labels the key areas and visual feature parameters corresponding to the information to be extracted in each image.

[0011] Step 2: Input the meteorological equipment image to be processed into the dedicated target detection model, output the coordinates and bounding boxes of the corresponding key areas and crop the image sub-blocks, select the text detection algorithm according to the morphological characteristics of the meteorological equipment text, and use the text detection algorithm to scan the text region of the image sub-blocks, and output the coordinates of the rectangular bounding box of the text region;

[0012] Step 3: Construct a dedicated text corpus for meteorological equipment and use an OCR recognition model. Optimize the OCR recognition model using the dedicated text corpus as pre-training data, and input the image of the text region into the optimized OCR recognition model to output the text content of the text region, forming the initial recognition result. The dedicated text corpus contains special symbols, numerical formats, and equipment-specific terminology in the meteorological field.

[0013] Step 4: Extract templates based on predefined information of the application scenario of meteorological equipment, and use a combination of rule matching and natural language processing to match and verify the initial recognition results with the predefined templates, remove invalid text that does not conform to the format, correct recognition errors, and output specific scenario information that has passed the match, thus completing the recognition and extraction of meteorological equipment image information.

[0014] Preferably, the deep learning object detection model is selected from one of YOLOv8, SSDMobileNetV3, and Faster R-CNN;

[0015] During training, Mosaic data augmentation, random rotation, and noise addition were used to simulate real-world environmental interference.

[0016] Preferably, the text detection algorithm is selected based on the text type in the key area: if it is digital display text, the DB algorithm is used;

[0017] For printed / etched text, the EAST algorithm is used.

[0018] Preferably, the OCR recognition model is a combination of CRNN and Attention mechanism, wherein the convolutional layer uses ResNet-50 to extract visual features of text, the recurrent layer uses bidirectional LSTM to capture text sequence dependencies, and the Attention layer focuses on key regions of text.

[0019] Preferred options also include:

[0020] A meteorological text error corpus was constructed, and abnormal fields in the initial identification results were replaced and corrected.

[0021] Preferably, before training the target detection model specifically for meteorological equipment, the following steps are also included:

[0022] Construct key region feature vectors based on the visual feature parameters of each sample. ,in, This is the type identifier for the key regions of the corresponding sample, where C is the contrast, D is the edge density, and S is the size ratio. () represents the coordinate range of the key area;

[0023] The environmental interference characteristics are determined, and F is expanded for each sample to obtain the expanded vector. ,in, Rain streak density; G represents the density of sand and dust particles; A represents the intensity of strong light reflection; and A represents the proportion of dark areas at night.

[0024] Calculate the multidimensional feature similarity between any two samples i and j. ;

[0025] ;

[0026] in, Let be the weight coefficient in the m-th dimension excluding the coordinate range, and ; Let m be the similarity in the m-th dimension; The cosine similarity for type identifiers, and , The type identifier encoding vector for sample i; Encode the type identifier vector for sample j; The normalized similarity is given by m = 2, 3, 4, 5, 6, 7, 8; This is the coordinate range weight, with a value of 0.1. , , The key regions for samples i and j; It is an area function; For the intersection and union comparison results;

[0027] Will The samples were grouped together, among which, This is a preset similarity threshold.

[0028] Preferably, after grouping the samples into the same group, the method further includes:

[0029] Four samples are randomly selected from each group, and the information importance weight Wk of the corresponding key area is calculated for each sample k.

[0030] ;

[0031] in, , These are the weighting coefficients; The coordinate range of the key region of sample k; The number of meteorological semantic units in the key area, k=1,2,3,4;

[0032] Semantic centers of four key sample regions For reference, the splicing canvas is divided into a 4×4 grid. Based on the semantic center positions of the four samples, a grid region is assigned to each sample, where... , , The coordinates of the center of the nth semantic unit of sample k;

[0033] For any pixel within the grid Attention-weighted fusion is used to obtain the stitched image, where the fusion formula is:

[0034] ;

[0035] in, For sample k in pixels Pixel value at; Let k be the attention weight for sample k at that pixel, and , Let be the dynamic standard deviation of the Gaussian function, and ; for The weighted fusion result;

[0036] Calculate the multi-constraint fusion loss of stitched images ;

[0037] ;

[0038] in, , , For loss weights; This is an extended feature vector for the key regions of the stitched image; This is an extended feature vector concatenated from the first four samples; The concatenated semantic segmentation graph semantic segmentation graph with sample k The intersection-union ratio function; The pixel overlap region between the stitched image and sample k; These are the pixel values ​​of the stitched image; For the sample k pixel value;

[0039] Iteratively adjust the splicing parameters until... A qualified stitched image is obtained;

[0040] Using qualified stitched images and original sample set images as training data, a deep learning target detection model was trained with the key area positioning accuracy as the optimization objective, resulting in a dedicated target detection model for meteorological equipment.

[0041] Preferably, the initial identification result is matched and verified against a predefined template, including:

[0042] The fields in the initial recognition results are matched using regular expressions. If a field is missing or the format is missing, it is marked as a format error and the error field is recorded.

[0043] For numeric fields that pass format validation, calculate the numeric confidence score. If the numeric confidence score is less than a preset value, mark it as a numeric anomaly.

[0044] For fields that pass the numerical validity check, calculate the semantic matching degree Ms between the fields;

[0045] ;

[0046] Where n0 is the total number of fields; The term in the p-th field of the initial recognition result Standard terminology in the template Cosine similarity; For the p-th field With the qth field The degree of logical consistency;

[0047] like , marked as semantic anomalies, where a1 is the matching degree threshold;

[0048] The system calls upon a dedicated text error corpus for meteorological equipment and historical normal data from the same equipment to correct fields marked with formatting, numerical, or semantic errors until all fields pass validation or are confirmed to be beyond correction. Finally, it outputs structured meteorological equipment information that has passed the matching validation.

[0049] Compared with the prior art, the beneficial effects of this application are as follows:

[0050] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0051] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0052] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0053] Figure 1 This is a flowchart of a method for identifying and extracting image information from meteorological equipment according to an embodiment of the present invention. Detailed Implementation

[0054] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0055] This invention provides a method for identifying and extracting image information from meteorological equipment, such as... Figure 1 As shown, it includes:

[0056] Step 1: Construct a key area sample set for meteorological equipment, and use a deep learning target detection model to train a special target detection model for meteorological equipment using the sample set as training data. The sample set contains images of different types of meteorological equipment in various environmental scenarios, and labels the key areas and visual feature parameters corresponding to the information to be extracted in each image.

[0057] Step 2: Input the meteorological equipment image to be processed into the dedicated target detection model, output the coordinates and bounding boxes of the corresponding key areas and crop the image sub-blocks, select the text detection algorithm according to the morphological characteristics of the meteorological equipment text, and use the text detection algorithm to scan the text region of the image sub-blocks, and output the coordinates of the rectangular bounding box of the text region;

[0058] Step 3: Construct a dedicated text corpus for meteorological equipment and use an OCR recognition model. Optimize the OCR recognition model using the dedicated text corpus as pre-training data, and input the image of the text region into the optimized OCR recognition model to output the text content of the text region, forming the initial recognition result. The dedicated text corpus contains special symbols, numerical formats, and equipment-specific terminology in the meteorological field.

[0059] Step 4: Extract templates based on predefined information of the application scenario of meteorological equipment, and use a combination of rule matching and natural language processing to match and verify the initial recognition results with the predefined templates, remove invalid text that does not conform to the format, correct recognition errors, and output specific scenario information that has passed the match, thus completing the recognition and extraction of meteorological equipment image information.

[0060] Preferably, the deep learning object detection model is selected from one of YOLOv8, SSDMobileNetV3, and Faster R-CNN;

[0061] During training, Mosaic data augmentation, random rotation, and noise addition were used to simulate real-world environmental interference.

[0062] Preferably, the text detection algorithm is selected based on the text type in the key area: if it is digital display text, the DB algorithm is used;

[0063] For printed / etched text, the EAST algorithm is used.

[0064] Preferably, the OCR recognition model is a combination of CRNN and Attention mechanism, wherein the convolutional layer uses ResNet-50 to extract visual features of text, the recurrent layer uses bidirectional LSTM to capture text sequence dependencies, and the Attention layer focuses on key regions of text.

[0065] In this embodiment, the sample set is constructed as follows: to address the diversity of meteorological equipment, images of more than 10 types of equipment are collected, including automatic weather stations (temperature and humidity sensors, barometric pressure sensors), weather radars (control panels, parameter displays), and observation signs (equipment numbers, range markings). The images cover four typical environments: rainy days (with added rain streaks), sandstorms (with added particle noise), strong light (with adjusted exposure), and nighttime (with reduced brightness and added light spots). A total of more than 50,000 images are collected. Each image is labeled with key areas and visual feature parameters to form a dedicated sample set.

[0066] In this embodiment, model training and optimization are performed as follows: YOLOv8 is selected as the basic target detection model. During training, the focus is on optimizing the feature extraction of key areas of meteorological equipment. For example, for the high brightness and rectangular border features of LED digital areas, the convolutional layer is enhanced to capture the edge contours; for the low contrast and small size features of etched text areas, the anchor frame size is adjusted, and small anchor frames of 16×16 and 32×32 are set. Mosaic data augmentation is used during the training process. Four samples are randomly spliced ​​and randomly rotated (-15°~15°) to simulate the installation angle deviation of the equipment. After 50 rounds of training, the accuracy of key area localization of the model reaches 98.5%.

[0067] Key region output: The image to be processed is input into the trained model, and the bounding box coordinates of each key region are output (e.g., the coordinates of the temperature and humidity region are: x1=500, y1=300, x2=700, y2=400). This region is cropped as the image sub-block for subsequent processing to eliminate irrelevant background interference.

[0068] In this embodiment, image preprocessing involves optimizing the extracted key area sub-blocks by using adaptive histogram equalization to enhance the contrast between text and background, thus solving the problems of strong light reflection and fading, and using 5×5 Gaussian filtering to remove rain streaks and dust noise. If the sub-block is a nighttime image, gamma correction is used to improve brightness.

[0069] Algorithm selection and adaptation: Select the detection algorithm according to the text type: For digital display text (such as LED temperature and humidity readings), use the DB algorithm (which can handle irregular text areas and has a fast detection speed), and set the binarization threshold to 0.4 (to balance detection integrity and accuracy); For printed / etched text (such as equipment numbers), use the EAST algorithm (which has high detection accuracy for small-sized text), and set the text line score threshold to 0.7 (to filter non-text areas); The algorithm outputs a rectangular bounding box of the text area and excludes non-text parts in key areas (such as sensor terminals and instrument panel scale lines).

[0070] In this embodiment, a dedicated corpus is constructed: a corpus containing special symbols in the meteorological field (℃, %RH, hPa, dBZ, km / h), numerical formats (such as -15.2°C, 99.9 decimals), and equipment terms (wind direction: NE, visibility: 10km) is constructed, which contains more than 10,000 valid texts and annotates the image features corresponding to the texts (such as the visual form of ℃ being: circle + letter C).

[0071] Model optimization and training: The CRNN+Attention model is used as the core of OCR recognition: the convolutional layer uses ResNet-50 to extract the visual features of the text, the recurrent layer uses bidirectional LSTM (hidden layer dimension 256) to capture text sequence dependencies, that is, to solve the confusion of 0 and O, 6 and 9, etc., and the attention layer focuses on key parts of the text, such as the % symbol in %RH and the decimal point; the model parameters are fine-tuned using a dedicated corpus as pre-training data, so that the accuracy of meteorological special symbol recognition is improved to 99.2%.

[0072] Error correction of recognition results: Construct an error corpus that records common misidentification cases, such as ℃→C, hPa→hP, -→_. Call the error corpus to perform replacement and correction, such as correcting temperature: 25.3C to temperature: 25.3℃.

[0073] In this embodiment, template definition and editing include: predefined multi-scene templates.

[0074] Temperature and humidity monitoring template: Temperature: {number}.{number}℃, Humidity: {number}.{number}%RH;

[0075] Barometric Pressure Monitoring Template: Barometric Pressure: {number}.{number}hPa;

[0076] Equipment Number Template: Equipment Number: QX-{Year}-{Number}; Provides a visual template configuration interface, allowing users to add templates, such as wind speed: {Number}.{Number}m / s, and modify format rules, such as changing temperature to T).

[0077] Matching and Validation: A dual validation process using rule matching and NLP is employed.

[0078] Rule matching: Check whether the initial recognition result conforms to the template format, such as temperature range -50℃~50℃ and humidity 0%~100%. If it does not conform, trigger a second recognition.

[0079] NLP verification: The recognition results and template keywords, such as temperature, humidity, and equipment number, are converted into Word2Vec word vectors, and the cosine similarity is calculated. If the similarity is greater than or equal to 0.85, the match is considered successful.

[0080] The beneficial effects of the above technical solution are: by locating key areas to eliminate irrelevant interference, and combining a meteorological corpus with an error correction mechanism, the accuracy of special symbol recognition is improved, solving the recognition deviation problems caused by reflection, fading, and noise. Post-processing is only performed on key areas, meeting the latency requirements of real-time meteorological data acquisition. It supports multi-scenario template customization, quickly adapting to different needs such as temperature and humidity monitoring, equipment maintenance, and radar calibration, requiring only template adjustments and a small amount of training data.

[0081] This invention provides a method for identifying and extracting image information from meteorological equipment. Before training a dedicated target detection model for meteorological equipment, the method further includes:

[0082] Construct key region feature vectors based on the visual feature parameters of each sample. ,in, This is the type identifier for the key regions of the corresponding sample, where C is the contrast, D is the edge density, and S is the size ratio. () represents the coordinate range of the key area;

[0083] The environmental interference characteristics are determined, and F is expanded for each sample to obtain the expanded vector. ,in, Rain streak density; G represents the density of sand and dust particles; A represents the intensity of strong light reflection; and A represents the proportion of dark areas at night.

[0084] Calculate the multidimensional feature similarity between any two samples i and j. ;

[0085] ;

[0086] in, Let be the weight coefficient in the m-th dimension excluding the coordinate range, and ; Let m be the similarity in the m-th dimension; The cosine similarity for type identifiers, and , The type identifier encoding vector for sample i; Encode the type identifier vector for sample j; The normalized similarity is given by m = 2, 3, 4, 5, 6, 7, 8; This is the coordinate range weight, with a value of 0.1. , , The key regions for samples i and j; It is an area function; For the intersection and union comparison results;

[0087] Will The samples were grouped together, among which, This is a preset similarity threshold.

[0088] Preferably, after grouping the samples into the same group, the method further includes:

[0089] Four samples are randomly selected from each group, and the information importance weight Wk of the corresponding key area is calculated for each sample k.

[0090] ;

[0091] in, , These are the weighting coefficients; The coordinate range of the key region of sample k; The number of meteorological semantic units in the key area, k=1,2,3,4;

[0092] Semantic centers of four key sample regions For reference, the splicing canvas is divided into a 4×4 grid. Based on the semantic center positions of the four samples, a grid region is assigned to each sample, where... , , The coordinates of the center of the nth semantic unit of sample k;

[0093] For any pixel within the grid Attention-weighted fusion is used to obtain the stitched image, where the fusion formula is:

[0094] ;

[0095] in, For sample k in pixels Pixel value at; Let k be the attention weight for sample k at that pixel, and , Let be the dynamic standard deviation of the Gaussian function, and ; for The weighted fusion result;

[0096] Calculate the multi-constraint fusion loss of stitched images ;

[0097] ;

[0098] in, , , For loss weights; This is an extended feature vector for the key regions of the stitched image; This is an extended feature vector concatenated from the first four samples; The concatenated semantic segmentation graph semantic segmentation graph with sample k The intersection-union ratio function; The pixel overlap region between the stitched image and sample k; These are the pixel values ​​of the stitched image; For the sample k pixel value;

[0099] Iteratively adjust the splicing parameters until... A qualified stitched image is obtained;

[0100] Using qualified stitched images and original sample set images as training data, a deep learning target detection model was trained with the key area positioning accuracy as the optimization objective, resulting in a dedicated target detection model for meteorological equipment.

[0101] In this embodiment, the type identifier is used to distinguish the text presentation type of key areas. For example, the LED digital display area of ​​the temperature and humidity sensor of an automatic weather station is marked with T to indicate LED digital type; the etched text area of ​​the weather radar instrument panel is marked with T to indicate etched text type.

[0102] Contrast ratio is the difference in brightness between text and background in a key area, calculated using the difference in grayscale values. For example, the measured C value for text (high-brightness white) in an LED digital area and a dark panel background is 0.85 (range 0~1); the contrast ratio for faded signage text and background is low, so it is taken as 0.3.

[0103] Edge density is the percentage of pixels at the edges of text in key areas, calculated using Canny edge detection. For example, the edge density of the number 8 is 0.6, and the edge density of the number 1 is 0.2.

[0104] Size proportions refer to the length-to-width ratio of key areas.

[0105] Rain streak density is the percentage of rain streak pixels in key areas during rainy weather. For example, by segmenting an image to identify rain streaks and calculating the percentage, we can determine the percentage. For instance, during moderate rain, 30% of the pixels in the temperature and humidity area are covered by rain streaks, so R=0.3; on non-rainy days, R=0.

[0106] Dust particle density is the percentage of dust particles in a key area of ​​a dusty scene. Target detection identifies dust particles and calculates their percentage; for example, in light dust, 15% of pixels are covered, so Ds=0.15; in non-dust conditions, Ds=0.

[0107] Strong light reflection intensity is the percentage of highlight pixels in key areas under strong light. A brightness threshold is set to statistically determine this percentage. For example, under midday strong light, 25% of the LED screen pixels are highlight pixels, so G=0.25; under non-strong light, G=0.

[0108] The nighttime dark area ratio is the percentage of dark pixels (grayscale value < 50) in key nighttime areas. It involves statistically analyzing the percentage of dark pixels; for example, if 40% of pixels are too dark at night, A = 0.4; otherwise, A = 0.

[0109] The weights for each dimension feature are pre-defined. For example, the weight for type T is w1=0.2, the weight for contrast C is w2=0.15, the weight for edge density D is w3=0.15, and the weights for the other dimensions (S, R, Ds, G, A) are wm=0.1. These weights can be used directly.

[0110] In this embodiment, the type identifier is encoded as a one-hot vector and then the cosine similarity is calculated. For example, if sample i is an LED digital type with an encoding of [1,0] and sample j has the same type encoding of [1,0], then sim1=1; if j is an etched text type, then the encoding is [0,1], and sim1=0.

[0111] For normalized similarity, for example, if the contrast of sample i is 0.8 and the contrast of sample j is 0.7, then sim2 = 1 - .

[0112] In this embodiment, =0.6, =0.4, which adjusts the influence of the number of semantic units and the size of the region; this is a pre-set value.

[0113] In this embodiment, , , The values ​​are 0.4, 0.3, and 0.3, respectively, to adjust the proportion of each loss.

[0114] In this embodiment, the value of y0 is 0.8, which was obtained based on experimental settings.

[0115] The beneficial effects of the above technical solution are: through multi-dimensional feature encoding grouping, semantic center-guided attention splicing and multi-constraint loss optimization, the generated enhanced samples can accurately simulate complex meteorological environments, enabling the trained dedicated target detection model to reduce irrelevant interference and retain core semantic information in key area localization under rainy, dusty and other scenarios, significantly improving localization accuracy and robustness, and laying a solid foundation for subsequent OCR recognition and scene matching.

[0116] This invention provides a method for identifying and extracting image information from meteorological equipment, comprising matching and verifying the initial identification result with a predefined template, including:

[0117] The fields in the initial recognition results are matched using regular expressions. If a field is missing or the format is missing, it is marked as a format error and the error field is recorded.

[0118] For numeric fields that pass format validation, calculate the numeric confidence score. If the numeric confidence score is less than a preset value, mark it as a numeric anomaly.

[0119] For fields that pass the numerical validity check, calculate the semantic matching degree Ms between the fields;

[0120] ;

[0121] Where n0 is the total number of fields; The term in the p-th field of the initial recognition result Standard terminology in the template Cosine similarity; For the p-th field With the qth field The degree of logical consistency;

[0122] like , marked as semantic anomalies, where a1 is the matching degree threshold;

[0123] The system calls upon a dedicated text error corpus for meteorological equipment and historical normal data from the same equipment to correct fields marked with formatting, numerical, or semantic errors until all fields pass validation or are confirmed to be beyond correction. Finally, it outputs structured meteorological equipment information that has passed the matching validation.

[0124] Preferably, a meteorological text error corpus is constructed to replace and correct abnormal fields in the initial identification results.

[0125] In this embodiment, the format regular expression is a string pattern used to verify whether the format of a text field conforms to preset rules. It can precisely constrain the text structure of the field, such as the field name, symbols, and numerical format. For example, the temperature field text in the initial recognition result, such as temperature: 25.3℃, is matched. If the recognition result is temperature: 25C, the colon format or suffix is ​​incorrect, or temperature: 25.3℃ (incorrect field name), it is marked as a format anomaly and the field is recorded.

[0126] In this embodiment, numerical confidence is an indicator that measures the meteorological logic of numerical fields, such as temperature, humidity, and air pressure values, integrating dimensions such as numerical range conformity and spatiotemporal correlation. Taking the temperature field as an example, the normal range is set to ℃℃, while also referencing the historical temperatures identified by the same device within the adjacent 5 minutes (e.g., historical values ​​of ℃, ℃) and the concurrent temperatures of three adjacent observation points in the same area (e.g., ℃, ℃, ℃). If the current identified temperature is ℃, the following is calculated:

[0127] Range compliance: 1 because it is within ℃;

[0128] Time correlation: The Pearson correlation coefficient with historical values ​​is 0.95 (0.95 after normalization).

[0129] Spatial consistency: The deviation rate from the mean ℃ of adjacent observation points is (25.3−24.77) / 24.77=0.021 (normalized consistency is 0.98);

[0130] Final confidence level: According to the formula Confidence Level = 0.5 × Range Compliance + 0.3 × Temporal Correlation + 0.2 × Spatial Consistency, we get 0.5 × 1 + 0.3 × 0.95 + 0.2 × 0.98 = 0.981. If the preset value threshold is 0.6, the value passes the verification; if the identified temperature is ℃, it is outside the normal range. The calculated confidence level is 0.73, and the preset threshold is 0.8, so it is marked as an abnormal value.

[0131] Among them, the normal range of values ​​is set based on meteorological operational experience, and historical data and data from adjacent observation points are retrieved from the database. Combined with statistical methods (Pearson correlation, deviation rate calculation) to achieve multi-dimensional integrated evaluation.

[0132] In this embodiment, n0 is the number of fields included in the initial recognition result. For example, if the recognition result is temperature: 25.3℃, humidity: 60%RH, air pressure: 1013.2hPa, then n0=3.

[0133] Using pre-trained word vector models, such as Word2Vec trained on meteorological texts, cosine similarity is calculated after vectorizing terms.

[0134] Determine whether the logical relationship between fields conforms to meteorological patterns, such as low humidity when the temperature is high and low wind speed when the air pressure is stable. For example, if the temperature is ℃ and the humidity is 60%RH, it conforms to the pattern that humidity is mostly between 40% and 70% when the temperature is moderate, so the Logic value is 1; if the temperature is ℃ and the humidity is 90%RH, it does not conform to the common pattern that humidity is low when the temperature is high, so the Logic value is 0.6.

[0135] We analyze the logical patterns between meteorological fields, such as the relationships between temperature and humidity, and temperature and air pressure, and transform these patterns into numerical compliance rules, such as 1 for compliance and assigning scores based on the degree of deviation for partial compliance.

[0136] a1 is the preset semantic matching score threshold. If it is lower than this value, it is judged as semantic anomaly. If a1=0.8, Ms=0.816 is calculated and the verification is passed; if Ms=0.75, it is marked as semantic anomaly.

[0137] In this embodiment, the dedicated text error corpus for meteorological equipment collects common OCR error cases in the meteorological field and establishes a mapping relationship between error text and correct text. For example, the corpus stores mappings such as C→℃, %R→%RH, and hP→hPa. These are manually labeled meteorological OCR error cases compiled into the corpus, and the automatic replacement of error text is achieved through string matching technology.

[0138] Historical normal data for the same device involves retrieving previously identified normal data from that meteorological device to help correct obviously abnormal values. For example, if a device's historical temperature is mostly in the range of ℃, but the current identified value is ℃ with low confidence, combining historical data with semantic matching logic can correct it to a reasonable value around ℃. Historical identification data for the device is stored in a database. When an anomaly is detected, historical data is retrieved and corrected using semantic rules (such as historical value fluctuation range and typical values ​​for the same scene).

[0139] The beneficial effects of the above technical solution are: by strictly constraining the field format through regular expressions, calculating the numerical confidence level through multi-dimensional meteorological logic, evaluating the semantic matching degree through word vectors and domain rules, and then using error corpus and historical data to correct anomalies throughout the entire process, it is possible to accurately screen and correct format, numerical and semantic errors in the initial recognition results, and significantly improve the accuracy and structure of meteorological equipment image information extraction.

[0140] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method of identifying extracted meteorological equipment image information, characterized by, The method comprises the following steps: Step 1: Construct a key region sample set of meteorological equipment, and use a deep learning target detection model to train a meteorological equipment special target detection model by taking the sample set as training data, wherein the sample set contains images of different types of meteorological equipment in various environmental scenes, and the key regions and visual feature parameters corresponding to the information to be extracted in each image are labeled; Step 2: input the image of the meteorological equipment to be processed into the special target detection model, output the coordinates and bounding box of the corresponding key region, and intercept the image sub-block, select a character detection algorithm according to the morphological characteristics of the meteorological equipment characters, and use the character detection algorithm to scan the character region of the image sub-block, and output the rectangular bounding box coordinates of the character region; Step 3: Construct a special text corpus of meteorological equipment, and optimize the OCR recognition model by taking the special text corpus as pre-training data, and input the image of the character region into the optimized OCR recognition model, and output the text content of the character region to form an initial recognition result, wherein the special text corpus contains special symbols, numerical formats and equipment-specific terminologies in the meteorological field; Step 4: Predefine the information extraction template according to the application scenario of the meteorological equipment, and use a method combining rule matching and natural language processing to match and verify the initial recognition result with the predefined template, eliminate invalid text that does not conform to the format, correct recognition errors, output specific scene information that passes the matching, and complete the recognition and extraction of the meteorological equipment image information; Before training the meteorological equipment special target detection model, the method further comprises the following steps: Constructing a key region feature vector based on visual feature parameters of each sample wherein, is a type identifier of the key region of the corresponding sample, C is a contrast, D is an edge density, S is a size ratio, is a coordinate range of the key region; The environmental interference feature is determined to extend F of each sample to obtain an extended vector wherein, is the density of rain marks; is the density of sand particles; G is the intensity of light reflection in strong light; and A is the proportion of dark areas at night. Computing a multi-dimensional feature similarity of any two samples i, j ; ； wherein, is the weight coefficient in the mth dimension other than the coordinate range, and ; is the similarity in the mth dimension; is the cosine similarity of the type identifier, and , is the type identifier encoding vector of sample i; is the type identifier encoding vector of sample j; is the normalized similarity, and m = 2, 3, 4, 5, 6, 7, 8; is the coordinate range weight, taking a value of 0.1; , , is the key area of sample i and sample j; is the area function; is the intersection over union result; Will The samples were grouped together, among which, To set a preset similarity threshold; After grouping the samples into the same group, the method further comprises the following steps: Randomly select 4 samples from each group, and calculate the information importance weight Wk of the corresponding key region for each sample k; ； wherein, , is a weight coefficient; is a coordinate range of the key region of sample k; is the number of meteorological semantic units in the key region, k = 1, 2, 3, 4; with the semantic center of 4 sample key regions For reference, the stitching canvas is divided into a 4x4 grid, and each sample is assigned a grid region according to the semantic center position of the 4 samples, wherein, , , is the center coordinate of the nth semantic unit of sample k; For any pixel within the grid The stitching image is obtained by attention weighting fusion, and a fusion formula is: ； wherein, is the pixel value of sample k at pixel ; is the attention weight of sample k at this pixel, and , is the dynamic standard deviation of the Gaussian function, and ; is the weighted fusion result of ; Computing multi-constraint fusion loss for stitched images ; ; wherein, , , is a loss weight; is an extended feature vector of a key region of the stitched image; is an extended feature vector of the first four samples before stitching; is a semantic segmentation map after stitching is an intersection over union function of the semantic segmentation map of sample k; is a pixel overlap region of the stitched image and sample k; is a pixel value of the stitched image; is a pixel value of the stitched image; is a pixel value of sample k; The stitching parameters are iteratively adjusted until a qualified stitched image is obtained. Based on the qualified spliced image and the original sample set image as training data, a deep learning target detection model is used to train a meteorological equipment special target detection model with key region positioning accuracy as the optimization target.

2. The method of claim 1, wherein, The deep learning target detection model is selected from one of YOLOv8, SSDMobileNetV3 and FasterR-CNN; In the training process, Mosaic data enhancement, random rotation and noise addition are used to simulate real environment interference.

3. The method of claim 1, wherein the method further comprises: The character detection algorithm is selected according to the type of characters in the key region: if it is a digital display character, DB algorithm is used; If it is a printed / etched character, EAST algorithm is used.

4. The method of claim 1, wherein, The OCR recognition model is a combined model of CRNN and Attention mechanism, wherein the convolutional layer uses ResNet-50 to extract character visual features, the recurrent layer uses bidirectional LSTM to capture text sequence dependence, and the Attention layer focuses on the key region of the character.

5. The method of claim 1, wherein, The method further comprises the following steps: Construct a meteorological text error corpus to replace and correct abnormal fields in the initial recognition result.

6. The method of claim 1, wherein, The matching and verification of the initial recognition result with the predefined template comprises the following steps: Use a format regular expression to match the fields in the initial recognition result, and if a field is missing or the format is missing, mark it as a format anomaly and record the abnormal field; For the numeric field that passes the format check, calculate the numeric confidence, and if the numeric confidence is less than a preset value, mark it as a numeric anomaly; For the field that passes the numeric reasonableness check, calculate the semantic matching degree Ms between fields; ； wherein n0 is the total number of fields; is the cosine similarity of the pth field in the initial recognition result with the standard term in the template; is the logical association compliance of the pth field with the qth field ;​ If , is marked as a semantic anomaly, wherein a1 is a matching degree threshold value; Call the meteorological equipment-specific text error corpus and the same equipment historical normal data to correct the fields marked with format anomalies, numeric anomalies, or semantic anomalies until all fields pass the check or it is confirmed that they cannot be corrected, and finally output the structured meteorological equipment information that passes the matching check.

Images