Engineering pipeline drawing analysis and reading optimization method

Abstract

The invention provides an engineering pipeline drawing analysis and reading optimization method, which comprises the following steps: S1, providing drawings of different styles as format labeling samples, extracting table features in the drawings and taking the table features as minimum recognition units, training the minimum recognition units by using a neural network algorithm, and generating a unit template database set; s2, using a neural network algorithm to extract each table feature of a to-be-recognized drawing, and performing optimal matching with a minimum recognition unit in the unit template database set; s3, combining the minimum recognition units with the highest matching degree to form a drawing recognition template with the highest matching degree with the to-be-recognized drawing; and S4, according to the drawing recognition template, performing analysis and character recognition on the to-be-recognized drawing to obtain information. The method can efficiently and accurately realize analysis of different pipeline axonometric drawings.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to the field of drawing information analysis. Background Art

[0002] Some pipeline drawings include not only graphic information such as pipeline routes, but also text information such as material information. Different design institutes have drawings in different styles or formats.

[0003] Currently, there are methods on the market for automatic recognition of drawings based on optical character recognition (OCR), which can export recognition results with one click, but there are problems such as slow recognition speed and insufficient recognition accuracy for large-volume drawings.

[0004] How to achieve automatic parsing and accurate identification of drawing information is an urgent problem to be solved. Summary of the invention

[0005] An object of the present invention is to provide an engineering pipeline drawing parsing and reading optimization method, which can efficiently and accurately implement the parsing of different pipeline axonometric drawings.

[0006] The engineering pipeline drawing parsing and reading optimization method for achieving the above purpose includes the following steps:

[0007] S1. Provide drawings of different styles as layout annotation samples, extract the features of each table in each drawing and use it as the minimum recognition unit, use a neural network algorithm to train the minimum recognition unit, and generate a unit template database set;

[0008] S2. Using a neural network algorithm to extract the features of each table of drawings to be identified, and optimally matching the smallest identification unit in the unit template database set;

[0009] S3. Combining the smallest recognition units with the highest matching degree to form a drawing recognition template with the highest matching degree with the drawing to be recognized;

[0010] S4. According to the drawing recognition template, the drawing to be recognized is parsed and text recognized to obtain information.

[0011] In one or more embodiments, in S1, processing the layout annotation sample includes the following steps: performing table cropping, manually screening out qualified table features without intersections; and performing cell annotation and text annotation on different types of tables.

[0012] In one or more embodiments, for a table without a frame, the number of rows, columns, header rows, and data rows of the table without a frame are manually determined, and are marked and extracted.

[0013] In one or more embodiments, the table characteristics include one or more of table location, table type and table format.

[0014] In one or more embodiments, in S1 and S2, a convolutional neural network method and a recurrent neural network algorithm are used for training and feature extraction.

[0015] In one or more embodiments, in S2, the optimal match includes the following steps: extracting the table features of the drawing to be identified; determining the table type by a fully connected layer and locating the table area; using a loss function to measure the similarity between the two feature vectors of the table features and the minimum recognition unit in the unit template database set; and continuously adjusting the filter and other parameters through a back-propagation algorithm to minimize the loss function and achieve the optimal mapping from the input image to the feature vector.

[0016] In one or more embodiments, the drawings include isometric drawings, transformation diagrams, and flow charts.

[0017] In one or more embodiments, in S3, the drawing recognition template that has the highest matching degree with the drawing to be recognized is a layout within the layout annotation sample or a new layout. When it is a new layout, the drawing recognition template obtained each time is added to the unit template database set.

[0018] Due to the complexity of the layout and the large number of types, there is no logic in the table when identifying drawings. The above-mentioned engineering pipeline drawing parsing and reading optimization method extracts the table features of drawings of different styles as the minimum recognition unit, simplifies the table on the drawing step by step, and uses the table as the unit to improve the recognition accuracy; for different drawings to be identified, this method continuously expands the unit template database set of drawing table features, extracts and optimally matches them through algorithms, and flexibly combines the most suitable drawing recognition template according to the matching of the minimum recognition unit and the combination of table features, thereby improving the adaptability to drawings of different styles. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The above and other features, properties and advantages of the present invention will become more apparent through the following description in conjunction with the accompanying drawings and embodiments, in which:

[0020] Figure 1 It is a schematic diagram of one embodiment of an engineering piping drawing;

[0021] Figure 2 It is a flowchart of the method for optimizing the analysis and reading of engineering pipeline drawings;

[0022] Figure 3 It is a table feature extraction and matching process based on convolutional neural network;

[0023] Figure 4It is a flow chart of a specific embodiment of drawing processing. DETAILED DESCRIPTION

[0024] The present invention is further described below in conjunction with specific embodiments and drawings. More details are elaborated in the following description to facilitate a full understanding of the present invention. However, the present invention can obviously be implemented in a variety of other ways different from the description herein. Those skilled in the art can make similar generalizations and deductions based on actual application situations without violating the connotation of the present invention. Therefore, the protection scope of the present invention should not be limited by the content of this specific embodiment.

[0025] It should be noted that these and other subsequent drawings are only examples and are not drawn to scale, and should not be used to limit the actual scope of protection required by the present invention.

[0026] like Figure 1 As shown, the engineering pipeline axonometric drawing includes pipeline information, data information on the pipeline and table information. The table includes material list information, dimension information, engineering information and other data.

[0027] At present, OCR-based automatic recognition of pipeline single-line diagrams is commonly used to extract material lists, export recognition results with one click, and summarize material lists. However, for the fast and accurate recognition and extraction function, the recognition speed of large-volume drawings is slow and the recognition accuracy is insufficient. For example, when uploading multiple drawings at one time, the recognition is slow and it takes a lot of time to wait for the recognition to be completed; and the processing capacity for complex engineering drawings is limited, making it difficult to achieve high-precision information analysis. There is garbled code in the recognition, and a lot of manual intervention is still required after recognition, and the degree of automation is insufficient. For the analysis of drawings of different styles or formats from different design institutes, the accuracy rate drops significantly and the adaptability is poor.

[0028] Based on this, the present application provides a method for parsing and optimizing the information of engineering pipeline axonometric drawings. Based on the classified collection of table features of hanging drawings, vector matching of table features, template combination, and structured processing of table information, it can realize efficient parsing of pipeline axonometric drawing information, greatly improve the speed of recognizing pipeline axonometric drawing attribute information, and further improve the accuracy of large-scale simultaneous recognition.

[0029] Reference Figure 1 It is understood that the method comprises the following steps:

[0030] S1. Provide drawings of different styles as layout annotation samples, extract the features of each table in each drawing and use it as the minimum recognition unit, use the neural network algorithm to train the minimum recognition unit, and generate a unit template database set;

[0031] S2. Use a neural network algorithm to extract the features of each table of drawings to be identified, and optimally match them with the minimum recognition unit in the unit template database set;

[0032] S3. Combining the smallest recognition units with the highest matching degree to form a drawing recognition template with the highest matching degree with the drawing to be recognized;

[0033] S4. Based on the drawing recognition template, analyze and recognize text on the drawing to be recognized to obtain information.

[0034] Specifically, step S1 includes the following steps.

[0035] First, the preprocessing module is used to normalize the different styles of drawings used as layout annotation samples, including denoising, contrast enhancement, etc. The different styles of drawings include but are not limited to axonometric drawings, conversion drawings, flow charts, etc. of different design institutes.

[0036] Subsequently, the table is manually cropped, and features such as graphics, symbols, and lines in the layout annotation samples are extracted for table cropping and annotation. Furthermore, during the table cropping and annotation process, qualified table features without intersections are manually screened out, that is, problems such as intersections between table frames and text, intersections between table frames, and blank text lines are identified to ensure the accuracy of the annotation process, thereby ensuring the accuracy of the extraction of the minimum recognition unit.

[0037] Table features include, but are not limited to, one or more of table position, table type, and table format, etc. Any style drawing includes multiple types of table features.

[0038] Form types include but are not limited to Figure 1 The material table A, parameter table B, pipeline table C, pipe section table D, basic table, etc. are shown. The table position refers to the position of a single table on the overall drawing. For example, the material table A is located at Figure 1 In the upper right shown, in other embodiments, the material table A may also be located in the middle or upper left of the drawing, etc. Each large table is composed of multiple cells, and the table format refers to the cell distribution form such as the number of rows and columns in any type of table cut.

[0039] After completing the division of the table, perform cell annotation and text annotation to process the extracted text information and other data in the table. For example, when annotating text rows, select rows one by one under the cell range and record them one by one. When annotating cells, record the table header, number of rows, number of columns, row width, column width and other information.

[0040] For the table of the wireless frame, the number of rows, columns, header rows and data rows of the table of the wireless frame are manually determined, and are marked and extracted as initial input data of the minimum recognition unit of the table of the wireless frame.

[0041] After the minimum recognition units of all types of drawings that may be involved, such as axonometric drawings, conversion diagrams, flow charts, etc., are extracted, they are used as input values, and the input minimum recognition units are trained using a neural network algorithm to generate a unit template database set.

[0042] For example, by using convolutional neural network methods and recurrent neural network algorithms (CNN+RNN), a rich set of unit template databases for algorithm recognition and matching is formed through model training of massive feature samples. Using the CNN convolutional neural algorithm, it is possible to extract local features and spatial hierarchical structures of drawing image features in the recognition area. By applying multi-layer convolution, pooling and nonlinear activation functions, local features in the drawings and their spatial hierarchical structures can be effectively extracted, thereby generating high-level feature representations that are conducive to subsequent processing. Even if a new drawing layout appears, the most suitable drawing recognition template can be flexibly combined according to the matching and combination of table features, which greatly solves the problem of poor recognition effect of new drawing layouts.

[0043] In steps S2 and S3, the method for extracting table features of drawings to be identified and performing template matching includes the following steps.

[0044] Step S2 is also processed using the CNN+RNN algorithm.

[0045] like Figure 3 As shown, the image information is used as the input layer 101, and the first convolution layer 102 uses a set of filters (or convolution kernels) to perform convolution operations on the input drawings to be recognized. Each filter of the first pooling layer 104 is responsible for extracting a specific feature, such as edges, corners, or basic wireframes. Each filter slides over the entire image information to generate a feature map, which contains feature information in the image that matches the filter.

[0046] Subsequently, a nonlinear activation function (such as ReLU) is applied to the output of the convolution layer in the pooling layer to add nonlinear characteristics so that the network can capture complex patterns; then the data dimension is reduced by reducing the resolution of the feature map while retaining important feature information, which helps prevent overfitting and makes the features have a certain spatial invariance; finally, after passing through multiple convolutional layers 104 and multiple pooling layers 105, the image is subjected to more refined feature extraction, and the high-level abstract features are sent to the fully connected layer 106 for the final classification or regression task, and finally output in the output layer 107. The fully connected layer 106 stores the final table as the minimum feature unit.

[0047] In drawing table recognition, the fully connected layer can determine the type of table based on the extracted features, such as material table, basic information table or parameter table, and achieve precise positioning of the entire table area.

[0048] In order to calculate the similarity with the template, a loss function is used to measure the distance or similarity between two feature vectors.

[0049] During the training phase, the CNN algorithm continuously adjusts filters and other parameters through the back-propagation algorithm to minimize the loss function, thereby achieving the best mapping from the input image to the feature vector.

[0050] The RNN algorithm relies on its ability to process feature sequence information and is suitable for processing long text sequences of indefinite length. It captures the long-distance dependencies of text sequences in the drawings to be identified, thereby determining the exact location of text label predictions.

[0051] In this way, when the drawing is submitted for recognition, the key features can be quickly extracted and optimally compared with the feature vectors of the table template library to obtain the minimum recognition unit with the highest matching degree.

[0052] Finally, the table features are combined to form a drawing recognition template that has the highest matching degree with the drawing format to be recognized. The drawing recognition template that has the highest matching degree with the drawing to be recognized can be an existing format in the format annotation sample or a new format.

[0053] Through the template's feature matching mechanism, a new type of drawings with complex layouts can be quickly matched and identified. After quickly reading the table features, the process of disassembling and template-based processing of the complex layout is completed.

[0054] Finally, step S4 is performed, and the drawing to be identified can be accurately parsed and text recognized based on the drawing recognition template to obtain information. For example, after identifying a new drawing including material table A, parameter table B, pipeline table C and other data with different matching relationships, based on the known information and features of the table, efficient parsing and rapid recognition of the drawing information can be achieved. Under the guidance of known data, compared with single OCR recognition, each line of text and data can be specifically recognized.

[0055] The above algorithm model has high scalability. In some embodiments, the new layout formed by the combination is added to the unit template database set, and the drawing table feature template data set is continuously expanded. Through the continuous cycle of the above steps, the general and rapid recognition ability of the drawings is improved. Even when the drawing recognition of a new layout appears in the future, the unit template database set still has good generalization ability.

[0056] The above method has the following advantages:

[0057] 1) By training various drawing table feature templates and continuously expanding the template database of the algorithm model, the general ability of rapid identification of pipeline drawings has been greatly improved, providing a technical basis for timely providing more accurate drawing material lists and other data for various projects.

[0058] 2) By combining the extraction of drawing layout table features with template feature vector mapping and matching, the generalization ability of the drawing model has been greatly improved. In terms of the range of pipeline drawings that can be supported for recognition, a more complete engineering pipeline professional drawing recognition system has been gradually formed.

[0059] 3) The algorithm model has good scalability. For different types of industrial drawings in different professional directions, it can collect according to the drawing sample type template - extract features of the table type in the drawing - classify and group the table features in the template - match and combine the table features to form an intelligent drawing recognition library covering all professions.

[0060] The present application uses specific words to describe the embodiments of the present application. For example, "one embodiment", "an embodiment", and / or "some embodiments" refer to a certain feature, structure or characteristic related to at least one embodiment of the present application. Therefore, it should be emphasized and noted that "one embodiment" or "an embodiment" or "an alternative embodiment" mentioned twice or multiple times in different positions in this specification does not necessarily refer to the same embodiment. In addition, some features, structures or characteristics in one or more embodiments of the present application can be appropriately combined.

[0061] In some embodiments, numbers describing the number of components and attributes are used. It should be understood that such numbers used in the description of the embodiments are modified by the modifiers "about", "approximately" or "substantially" in some examples. Unless otherwise specified, "about", "approximately" or "substantially" indicate that the numbers are allowed to vary by ±20%. Accordingly, in some embodiments, the numerical parameters used in the specification and claims are approximate values, which may change according to the required features of individual embodiments. In some embodiments, the numerical parameters should take into account the specified significant digits and adopt the general method of retaining digits. Although the numerical domains and parameters used to confirm the breadth of their range in some embodiments of the present application are approximate values, in specific embodiments, the setting of such numerical values ​​is as accurate as possible within the feasible range.

[0062] Although the present invention is disclosed as above with preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art may make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. The method for optimizing the analysis and reading of engineering pipeline drawings is characterized by: The steps include: S1. Provide drawings of different styles as layout annotation samples, extract the features of each table in each drawing and use it as the minimum recognition unit, use a neural network algorithm to train the minimum recognition unit, and generate a unit template database set; S2. Using a neural network algorithm to extract the features of each table of drawings to be identified, and optimally matching the smallest identification unit in the unit template database set; S3. Combining the smallest recognition units with the highest matching degree to form a drawing recognition template with the highest matching degree with the drawing to be recognized; S4. According to the drawing recognition template, the drawing to be recognized is parsed and text recognized to obtain information.

2. The method according to claim 1, characterized in that In S1, processing the format annotation sample includes the following steps: Perform table trimming and manually screen out qualified table features without overlap; Perform cell and text annotation on different types of tables.

3. The method according to claim 2, characterized in that For a table without a wireless frame, the number of rows, columns, header rows and data rows of the table without a wireless frame are manually determined, and then marked and extracted.

4. The method according to claim 1, characterized in that The table features include one or more of table location, table type and table format.

5. The method according to claim 1, characterized in that In S1 and S2, convolutional neural network method and recurrent neural network algorithm are used for training and feature extraction.

6. The method according to claim 1, characterized in that In S2, the optimal matching includes the following steps: Extracting table features of the drawing to be identified; The fully connected layer determines the table type and locates the table area; Using a loss function to measure the similarity between two feature vectors about the table feature and the minimum recognition unit in the unit template database set; The back-propagation algorithm continuously adjusts the filter and other parameters to minimize the loss function and achieve the best mapping from the input image to the feature vector.

7. The method according to claim 1, characterized in that The drawings include axonometric drawings, conversion diagrams, and flow charts.

8. The method according to claim 1, characterized in that In S3, the drawing recognition template with the highest matching degree with the drawing to be recognized is a layout in the layout annotation sample or a new layout. When it is a new layout, the drawing recognition template obtained each time is added to the unit template database set.

Images