A multi-type meteorological self-recording paper trace recognition method based on deep learning

Abstract

The present invention relates to the field of meteorological technology and discloses a deep learning-based method for identifying traces on multiple types of meteorological recording paper. By utilizing deep learning technology and combining it with the cross-platform computer vision library OpenCV, the method extracts trace data from multiple types of meteorological recording paper and quantifies the results. A U-net semantic segmentation network is used to segment and extract traces from vertical axis wind direction / speed recording paper. Template matching and edge detection algorithms are used to implement grid extraction, and the extraction results are then quantified according to relevant meteorological specifications. This method comprehensively considers the differences in size, color, scale, and linear shape among different types of recording paper from multiple dimensions, including grid extraction, trace extraction, and result quantification, resulting in highly restored trace information. The method is suitable for the digital extraction and business processing of traces for wind, precipitation, temperature, and air pressure, significantly reducing workload and labor costs.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorology, and in particular to a method for recognizing multi-type meteorological self-recording paper traces based on deep learning. Background Art

[0002] Before the advent of automated observation in the 2000s, historical meteorological observation data in my country was primarily recorded on paper and manually transcribed. Prior to the founding of the People's Republic of China, various types of self-recording paper, including those for air pressure, diurnal temperature, diurnal humidity, French wind speed, vertical axis wind direction, Little Dyne wind, and R1 / R2 / R3 precipitation, were recorded. These data are extremely valuable due to their age, variety, and inconsistent formats and recording methods. However, their application and processing are also extremely difficult.

[0003] The trace identification and extraction of historical meteorological self-recording paper can provide timely, accurate and convenient information services for information departments, which is particularly important for carrying out weather monitoring business, enriching meteorological scientific and technological data, and conducting climate forecasts.

[0004] However, the key to trace extraction lies in the accuracy of grid extraction, trace extraction, and the accuracy of the algorithm used to convert the image data extracted from the self-recording paper into actual data. Traditional image processing methods primarily rely on single image information, such as color or texture. This results in ineffective background noise removal when traces on different self-recording papers become faded, blurred, or smudged, resulting in low extraction accuracy.

[0005] The rescue of precious paper meteorological archive data information is a major task, but in the process of extracting self-recording paper trace data, due to the differences in types and the characteristics of different self-recording papers, the following problems exist:

[0006] 1. Traditional methods for trace reconstruction are easily affected by interference factors such as noise, ink contamination, breakage, and pencil scratches, resulting in low precision and accuracy of trace extraction.

[0007] 2. Traditional extraction methods have their own scope of application. When there are many types of self-recording paper and they vary greatly, an algorithm that performs well in extracting traces on one type of self-recording paper may not be transferable to other types of self-recording paper.

[0008] 3. Traditional image processing techniques can only utilize information such as color and texture of the image, and cannot fully exploit the differences in self-recording paper size, color, size, scale, line shape, etc., which limits the extraction accuracy;

[0009] 4. When traditional methods encounter large changes in the steepness of the curve during curve tracking, they cannot accurately track it, resulting in excessive manual adjustment work in the later stage. Summary of the Invention

[0010] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a deep learning-based multi-type meteorological self-recording paper trace recognition method. The method uses edge gradient features to perform grid template matching for various types of self-recording paper such as air pressure before the founding of the People's Republic of China, diurnal temperature, diurnal humidity, French wind speed, vertical axis wind direction, small Dyne wind, R1 / R2 / R3 precipitation, etc., to achieve accurate extraction of self-recording paper grids; the deep learning-based multi-type meteorological self-recording paper trace recognition method constructs a high-quality data set of different self-recording papers, combines multiple scale information, and uses a U-net convolutional neural network to achieve semantic segmentation, eliminate background noise, and complete high-quality and high-precision trace extraction.

[0011] The present invention provides the following technical solution: a method for identifying multi-type meteorological self-recording paper traces based on deep learning, comprising the following steps:

[0012] (1) Read different types of meteorological scanning image data according to path parameters;

[0013] (2) Parse the read data and determine whether its data format and data type are correct based on the file name;

[0014] (3) Preprocess the read meteorological scanning image data, mainly completing image correction and noise reduction;

[0015] (4) Using a template matching algorithm based on edge gradient features to extract grid traces from the paper;

[0016] (5) Based on different self-recording paper types such as temperature, humidity, precipitation, wind direction / speed, etc., a convolutional neural network algorithm based on deep learning is used to mine the spectral and spatial geometric information of the image, such as color, texture, and shape, to build a training database of different types of self-recording paper and generate a prediction model;

[0017] (6) Using the prediction models generated by different types of self-recording paper, image features are extracted to generate binary images;

[0018] (7) Using the cross-platform computer vision library OpenCV, the trace extraction results are optimized to remove noise and fill gaps, and to delete the wrong traces;

[0019] (8) According to relevant meteorological specifications such as temperature, humidity, precipitation, wind direction, and wind speed, the trace extraction results are quantified and the result file is generated.

[0020] Preferably, the specific processing in step 1 includes respectively reading the scanned image data of different stations and different types of self-recording paper for which trace extraction is required.

[0021] Preferably, the specific processing in step 2 includes determining whether the format and data type of the data read in step 1 meet the requirements.

[0022] Preferably, the specific processing in step 3 includes correcting the image to reduce data extraction and quantization errors caused by the tilt of the self-recording paper; using a directional filtering method to reduce the interference of image noise on the grid extraction and trace extraction of the self-recording paper.

[0023] Preferably, the specific processing in step 4 includes selecting characteristic characters in the image as a basis for template matching, processing the image feature template into discrete points through an edge detection algorithm; achieving template matching by measuring the similarity between the target image and the character template, and completing the self-recording paper grid trace extraction.

[0024] Preferably, a convolutional neural network algorithm based on deep learning is used to construct a training database of different types of self-recording papers, complete model training, and generate a prediction model.

[0025] Preferably, the step 5 of constructing a training database of different types of self-charging paper comprises manually cutting the image self-charging paper, generating thumbnails, labeling the images, and binarizing the training images using a method combining color channel attributes with dynamic threshold setting, i.e., generating image training labels.

[0026] The generation of the prediction model in step 5 refers to using the U-net convolutional neural network model, based on the training database generated in step 7, to fully utilize the spectral information and spatial geometric information of the image such as color, texture, and shape to generate a training model.

[0027] Preferably, the prediction model generated in step 6 and the different types of self-recording paper images preprocessed in step 3 are used to perform image feature extraction to generate a binary image.

[0028] Preferably, the binary feature image generated in step 6 is subjected to a directional filtering algorithm of the computer vision library OpenCV to remove burrs and edge points.

[0029] Preferably, according to relevant meteorological specifications such as temperature, humidity, precipitation, wind direction, and wind speed, the self-recording paper grid generated in step 5 is used as a calculation basis to quantify the extraction results of each pixel point on the extracted self-recording paper image and generate a result file.

[0030] The present invention has the following beneficial effects:

[0031] This deep learning-based method for multi-type meteorological chart trace recognition uses a template matching algorithm based on edge gradient features to extract chart grids. Compared with traditional image edge detection algorithms, it can adapt to a wider range of chart types (of varying sizes, styles, and coordinate systems), making it more universal. The deep learning-based convolutional neural network algorithm employed in this paper mines the image's spectral and spatial geometric information, including color, texture, and shape. It also eliminates the effects of chart trace noise caused by faded, blurred, smeared ink, and pencil marks, resulting in a highly accurate restoration of chart traces. Compared with traditional chart trace recognition algorithms, this method achieves more accurate and universal recognition results.

[0032] 2. This deep learning-based multi-type meteorological self-recording paper trace recognition method uses deep learning technology and combines it with the cross-platform computer vision library OpenCV to realize the extraction and result quantification of multi-type meteorological self-recording paper trace data; the extraction results of this method comprehensively consider various factors such as the size, grid type, scale, trace recording method, color, etc. of different types of self-recording paper. For temperature, humidity, precipitation and other self-recording paper before the founding of the People's Republic of China, a curve tracking method based on directional filtering is used to complete trace extraction; for vertical axis wind direction / wind speed self-recording paper, a U-net semantic segmentation network is used to complete trace segmentation and extraction; template matching and edge detection algorithms are used to realize grid extraction, and then the extraction results are quantified according to relevant meteorological specifications; this method comprehensively considers the differences in size, color, size, scale, line shape and other aspects of different types of self-recording paper from multiple dimensions such as grid extraction, trace extraction, and result quantification, and highly restores trace information. It is suitable for the digital extraction and business processing of traces such as wind, precipitation, temperature, and air pressure, which can greatly reduce workload and labor costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a flow chart of the identification method of the present invention;

[0034] Figure 2 This is a schematic diagram of wind direction training data of the present invention;

[0035] Figure 3 This is a second schematic diagram of wind direction training data of the present invention;

[0036] Figure 4 This is a schematic diagram of wind speed training data of the present invention;

[0037] Figure 5 This is a second schematic diagram of wind speed training data of the present invention;

[0038] Figure 6 This is a schematic diagram of a binary image of a wind speed trace of the present invention;

[0039] Figure 7This is a schematic diagram of the optimization of the binary results of the wind speed trace of the present invention;

[0040] Figure 8 This is an example of a trace binary image of the present invention;

[0041] Figure 9 The filter template of the present invention;

[0042] Figure 10 This is a schematic diagram of the trace binary filtering result of the present invention. DETAILED DESCRIPTION

[0043] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0044] See also Figure 1 A multi-type meteorological self-recording paper trace recognition method based on deep learning includes the following steps:

[0045] (1) Read different types of meteorological scanning image data according to path parameters;

[0046] The specific processing in step 1 includes respectively reading different stations and different types of self-recording paper scan image data that need to be extracted;

[0047] The input data contains:

[0048]

[0049]

[0050] (2) Parsing the read data and determining whether its data format and data type are correct based on the file name; the specific processing in step 2 includes determining whether the format and data type of the data read in step 1 meet the requirements.

[0051] (3) Preprocessing the read meteorological scanning image data mainly completes image deflection correction and noise reduction; the specific processing in step 3 includes correcting the image to reduce the data extraction and quantization errors caused by the tilt of the self-recording paper; using a directional filtering method to reduce the interference of image noise on the grid extraction and trace extraction of the self-recording paper.

[0052] First, a smoothing filter is used to remove noise from the image. The threshold is automatically determined based on local features such as the image's color information, and the image is binarized using an adaptive binarization method.

[0053] Secondly, the Hough transform algorithm is used to extract straight lines from the table. The horizontal lines of the table are removed using the slope, and the slopes of the vertical lines of the table are calculated. The image is corrected using the average of the statistical slopes.

[0054] The specific steps are as follows:

[0055] 1) Connect the upper endpoints of the grid to obtain the upper line of the grid frame;

[0056] 2) Calculate the tilt angle of the grid line. If the tilt angle is greater than 0.2 degrees, use the affine transformation method to correct the image.

[0057] 3) Affine transformation

[0058] The image should be rotated around the center point, but the origin of the image is in the upper left corner. When calculating, the origin of the upper left corner must be moved to the center of the image, and the Y axis needs to be flipped. Let's assume a point (X0, Y0), the image width is W, the height is H, and the point after the origin transformation is (X1, Y1). The transformation is as follows:

[0059]

[0060] The image rotation angle is θ. Suppose the point after the origin is transformed and rotated by the rotation matrix by θ is (X2, Y2). The formula is as follows:

[0061]

[0062] The width of the rotated image is W″ and the height is H″. The formula for transforming from the Cartesian coordinate origin back to the upper left corner is as follows:

[0063]

[0064] In summary, a point (X0, Y0) in the original image can be transformed to the rotated point (X3, Y3) using the following formula:

[0065]

[0066] Similarly, the rotated point (X3, Y3) can be transformed back to the original image (X0, Y0) through the following formula:

[0067]

[0068] (4) Using a template matching algorithm based on edge gradient features to extract grid traces from the paper;

[0069] The specific processing in step 4 includes selecting the characteristic characters in the image as the basis for template matching, processing the image characteristic template into discrete points through the edge detection algorithm; achieving template matching by measuring the similarity between the target image and the character template, and completing the self-recording paper grid trace extraction.

[0070] 1) Make a grid template for your self-recording paper

[0071] The production of the grid template is divided into two steps. First, the characteristic characters in the image are selected as the target of template matching. The selected characters can be the time numbers on the edge of the grid, etc.; then, the relative position coordinates of these characteristic characters and the key nodes of the grid are recorded.

[0072] ① Perform a series of rotations, scaling, and pyramid downsampling on the template image to generate a series of templates with different rotation angles and pyramid levels;

[0073] ② Extract the edge of the template and extract the edge points based on the principle of Canny algorithm;

[0074] ③Calculate the gradient value of the edge point in the x, y direction and the total gradient value;

[0075] ④ Save the x, y gradients corresponding to the edge points, normalize the gradient intensity to eliminate the influence of uneven illumination, and convert the edge point coordinates to relative coordinates relative to the center of gravity.

[0076] 2) Template matching to locate characteristic characters

[0077] ①Determine the similarity measure

[0078] When calculating the similarity metric of edge template matching, the edge point gradient vector is used as the matching information. After the template is processed by the edge detection algorithm, the template is converted into a series of points p i =(x i ,y i ) T , i = 1, 2, ?, n, corresponding to each point there is a gradient direction vector d i =(t i ,u i ) T .

[0079] The target image is converted into a point (x, y) and its corresponding direction vector e by the same process. x，y =(v x，y , w x，y ) T In the matching process, the similarity measure is calculated between the transformed template and the sub-image at a certain position of the target image. The gradient vector of the edge point in the template is transformed to d′ i =R(θ)d i, then at a certain position q=[x, y] in the target image T At the sub-image, the similarity metric function is described by the following formula:

[0080]

[0081] The formula calculates the average dot product of the gradient vectors of the corresponding edge points of the template and the sub-image. The dot product reaches its maximum value when the gradient directions of a pair of corresponding edge points are the same. The dot product is zero when the gradient directions are orthogonal. The similarity metric calculated using the above formula theoretically spans the entire real number domain, making it inconvenient for determining whether a sub-image meets the requirements. In other words, it is difficult to determine a threshold to filter the desired results. Therefore, the above formula is normalized:

[0082] (7)

[0083] A desirable feature of the above similarity measure is that when object recognition is based on a threshold S of the similarity measure that a potential match must reach min , there is no need to fully evaluate them. Let S j represents the partial sum of dot products to the jth element of the model. For the matching metric using the normalized dot product sum,

[0084]

[0085] Obviously, all the remaining terms of the sum are ≤ 1. Therefore, if S j <S min -1+j / n, so as long as this condition is met, the summation can stop calculating after the jth element.

[0086] ② Template matching

[0087] Calculate the gradient information of the edge points of the searched image; use two 3*3 matrix operators to convolve with the original image respectively to obtain the gradient values of the horizontal Gx and vertical Gy respectively;

[0088]

[0089] Horizontal and vertical gradient values

[0090] The gradient value of the image is calculated by the following formula:

[0091]

[0092] B) Calculate the gradient information of the edge points of the search image;

[0093] C) Pyramid downsampling to obtain edge point gradient information of the multi-level pyramid image;

[0094] D) The similarity between the template edge gradient and the target image edge gradient vector is calculated using the Normalized Cross Correlation (NCC) algorithm.

[0095] 3) Extract the mesh

[0096] After determining the characteristic character according to step 2), the grid endpoint is found according to the relative position of the character and the grid line endpoint, and the grid is extracted according to the obtained endpoint.

[0097] (5) Based on different self-recording paper types such as temperature, humidity, precipitation, wind direction / speed, etc., a convolutional neural network algorithm based on deep learning is used to mine the spectral and spatial geometric information of the image, such as color, texture, and shape, to build a training database of different types of self-recording paper and generate a prediction model; a convolutional neural network algorithm based on deep learning is used to build a training database of different types of self-recording paper, complete model training, and generate a prediction model;

[0098] The step 5 of constructing a training database of different types of self-marking paper refers to manually cutting the image self-marking paper, generating thumbnails, labeling the images, and binarizing the training images using a method combining color channel attributes with dynamic threshold setting, i.e., generating image training labels.

[0099] Generating a prediction model in step 5 refers to using a U-net convolutional neural network model, based on the training database generated in step 7, to fully utilize the spectral information such as color, texture, shape and spatial geometric information of the image to generate a training model;

[0100] Construct different types of self-recording paper training databases:

[0101] Since the image size of a single self-recording paper is large, considering the hardware facilities and computing efficiency, the self-recording curve image needs to be cropped into smaller images for training. Taking all factors into consideration, the image size is set to 320*320, and cross cropping is used for cropping.

[0102] 1) Training image labels:

[0103] Due to the complexity of different types of self-recording paper, manual labeling requires a considerable amount of time. To reduce labor costs and save time, the present invention designs a labeling method that combines dynamic thresholding and manual correction. First, the image is dynamically binarized. The principle is as follows: vertical block processing is performed according to the grid lines. For each block, the blue channel values of a single row of pixels are statistically sorted. The value at the center position is subtracted by 15 as the threshold T for that row. At the same time, other band operations are combined with channel thresholds to binarize the single row of pixels, realize dynamic thresholding segmentation of the image, and obtain a preliminary image binary map.

[0104] Binarization formula: (B<T&&(BR)<100)||(R<150)||(BG)>50), where B, G, and R are the pixel values of the blue, green, and red channels of the image pixels.

[0105] Manually correct the initial binary image, add the missed extraction parts and delete the over-extraction parts according to the original image, and obtain the final label after correction. The training image and label correspond one to one, as shown in the figure:

[0106] (6) Using the prediction model generated by different types of self-recording paper, perform image feature extraction to generate a binary image; using the prediction model generated in step 6 and the different types of self-recording paper images preprocessed in step 3, perform image feature extraction to generate a binary image

[0107] The binary image extraction method based on Unet network mainly includes the following parts: data preprocessing, U-Net network training, and trace binary image extraction.

[0108] ① Data preprocessing

[0109] The original wind recorder contains two parts: wind direction and wind speed. In order to avoid interference from other factors and obtain wind direction and wind speed information more accurately, the original wind recorder image needs to be cropped into two parts, wind direction and wind speed parts, and binary images of wind direction and wind speed are extracted respectively.

[0110] Other types of self-recording paper do not need to be cut.

[0111] ②U-net network training

[0112] We selected the TensorFlow platform as the training environment, built a U-net network, and conducted learning and training on different types of self-recording paper data. The parameter settings are as follows:

[0113] Lr=0.0001, batch_size=8, epoch=30, save the optimal model.

[0114] ③ Wind binary image extraction

[0115] Input the image and call the optimal model of different self-recording paper types to complete the binary image extraction of different self-recording paper types;

[0116] (7) Using the cross-platform computer vision library OpenCV, the trace extraction results are optimized to remove noise and fill gaps, and to delete the incorrectly extracted traces; for the binary feature image generated in step 6, the directional filtering algorithm of the computer vision library OpenCV is used to remove burrs and edge points;

[0117] 1) Divide the binary image into blocks of 100 pixels wide. Count the number of ink pixels in each row. The row with the largest number of ink pixels among the ten rows is used as the center of the block image. From this starting point, determine if any two rows have a non-zero number of ink pixels and a distance greater than 50 pixels between them. Pixels outside this range are removed to eliminate noise that is far from the main ink area.

[0118] 2) The binary image is extracted using area, and regions with an area less than 200 and a width less than 20 are extracted. The direction of the connected domain is determined using image moments, and connected domains with angles α < 70 or α > 110 are eliminated.

[0119] 3) The missing parts of the trace are supplemented using inherent rules to form the overall ink area, and the upper edge line of the area is extracted to obtain the wind speed trace.

[0120] 4) Horizontal filtering optimization

[0121] The lateral filtering method is used for filtering to remove the false positives on both sides of the trace.

[0122] Filtering method:

[0123] ① Traverse the binary results. If the pixel value of the blue light band of the pixel point is greater than the threshold 150 and it is the target pixel of the binary result, set the point as the point to be removed;

[0124] ② Extract the pixel values of the three pixels on the left and right of the point to be removed, and calculate the difference between each pixel value and the point to be removed. If the difference is less than -50, the point is considered to be the edge of the trace and is removed. Otherwise, the point is retained as the trace.

[0125] 5) Remove mistaken edge points

[0126] After horizontal filtering, there are a few protruding pixels on the edge of the result. These protrusions are edge errors, and these edges are removed by filtering. The filtering template is shown in the figure:

[0127] (8) Based on relevant meteorological specifications such as temperature, humidity, precipitation, wind direction, and wind speed, the trace extraction results are quantified and a result file is generated. Based on relevant meteorological specifications such as temperature, humidity, precipitation, wind direction, and wind speed, the self-recording paper grid generated in step 5 is used as the calculation basis to quantify the extraction results of each pixel on the extracted self-recording paper image and generate a result file.

[0128] The technical solution of the present invention is aimed at various types of self-recording papers, such as air pressure, diurnal temperature, diurnal humidity, French wind speed, vertical axis wind direction, small Dyne wind, R1 / R2 / R3 precipitation, etc. before the founding of the People's Republic of China. A multi-type meteorological self-recording paper trace recognition method based on deep learning is used to perform grid template matching using edge gradient features to achieve accurate extraction of self-recording paper grids; by constructing a high-quality data set of different self-recording papers, combining multiple scale information, and using a U-net convolutional neural network, semantic segmentation is achieved, background noise is eliminated, and high-quality and high-precision trace extraction is completed.

[0129] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0130] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying multi-type meteorological self-recording paper traces based on deep learning, characterized by: The steps include: (1) Read different types of meteorological scanning image data according to path parameters; (2) Parse the read data and determine whether its data format and data type are correct based on the file name; (3) Preprocess the read meteorological scanning image data, mainly completing image correction and noise reduction; (4) Using a template matching algorithm based on edge gradient features to extract grid traces from the self-recording paper; (5) Based on different self-recording paper types such as temperature, humidity, precipitation, wind direction / speed, etc., a convolutional neural network algorithm based on deep learning is used to mine the spectral and spatial geometric information of the image, such as color, texture, and shape, to build a training database of different types of self-recording paper and generate a prediction model; (6) Using the prediction models generated by different types of self-recording paper, image features are extracted to generate binary images; (7) Using the cross-platform computer vision library OpenCV, the trace extraction results are optimized to remove noise and fill gaps, and to delete incorrectly extracted traces; (8) Complete the quantification of trace extraction results according to relevant meteorological specifications such as temperature, humidity, precipitation, wind direction, and wind speed, and then generate the results file; Step 4 includes: selecting characteristic characters in the image as a basis for template matching, processing the image feature template into discrete points through an edge detection algorithm; achieving template matching by measuring the similarity between the target image and the character template, and completing the self-recording paper grid trace extraction; wherein, after determining the characteristic characters, finding the grid endpoints based on the relative position of the characters and the grid line endpoints, and extracting the grid based on the obtained endpoints.

2. The method for recognizing multiple types of meteorological self-recording paper traces based on deep learning according to claim 1, characterized in that: The specific processing in step 1 includes respectively reading the self-recording paper scan image data of different stations and different types that require trace extraction.

3. The method for recognizing multiple types of meteorological self-recording paper traces based on deep learning according to claim 1, characterized in that: The specific processing in step 2 includes determining whether the format and data type of the data read in step 1 meet the requirements.

4. The method for recognizing multiple types of meteorological self-recording paper traces based on deep learning according to claim 1, characterized in that: The specific processing in step 3 includes correcting the image to reduce the data extraction and quantization errors caused by the tilt of the self-recording paper; using a directional filtering method to reduce the interference of image noise on the grid extraction and trace extraction of the self-recording paper.

5. The method for recognizing multiple types of meteorological self-recording paper traces based on deep learning according to claim 1, characterized in that: Using the convolutional neural network algorithm based on deep learning, a training database of different types of self-recording papers was constructed to complete model training and generate a prediction model.

6. The method for recognizing multiple types of meteorological self-recording paper traces based on deep learning according to claim 1, characterized in that: The step 5 of constructing a training database of different types of self-recording paper refers to manually completing the cutting of the image self-recording paper, generating a small image, labeling the image, and using a method combining color channel attributes with dynamic threshold setting to complete the binarization of the training image, that is, generating image training labels; the step 5 of generating a prediction model refers to using a U-net convolutional neural network model, based on the training database generated in step 7, to fully utilize the spectral information and spatial geometric information of the image such as color, texture, and shape to generate a training model.

7. The method for recognizing multiple types of meteorological self-recording paper traces based on deep learning according to claim 1, characterized in that: Using the prediction model generated in step 6 and the different types of self-recording paper images preprocessed in step 3, image features are extracted to generate binary images.

8. The method for recognizing multiple types of meteorological self-recording paper traces based on deep learning according to claim 1, characterized in that: For the binary feature image generated in step 6, the directional filtering algorithm of the computer vision library OpenCV is used to remove burrs and edge points.

9. The method for recognizing multiple types of meteorological self-recording paper traces based on deep learning according to claim 1, characterized in that: According to relevant meteorological specifications such as temperature, humidity, precipitation, wind direction, and wind speed, the self-recording paper grid generated in step 5 is used as the calculation basis to quantify the extraction results of each pixel point on the extracted self-recording paper image and generate a result file.

Images