CN-122024263-A - Key information extraction method and system for engineering drawing detail table

CN122024263ACN 122024263 ACN122024263 ACN 122024263ACN-122024263-A

Abstract

The invention discloses a key information extraction method and a key information extraction system for an engineering drawing list, wherein the key information extraction method comprises the steps of configuring rotation angles and corresponding perspective transformation strategies for different engineering drawing list image samples respectively, constructing an automatic rotation correction model for correcting direction deviation between the direction of a target engineering drawing list image and the direction of a standard engineering drawing list image sample, dividing different image samples into a target drawing and a non-target drawing, constructing a non-target drawing filtering model for screening the target drawing, configuring header area marks and table body area marks for different image samples respectively, constructing a target detection model for identifying a key information area, correcting the direction deviation between the direction of a current target image and the direction of the standard image sample by utilizing each model, screening the target drawing, and further identifying the key information area to extract key information. The invention realizes the efficient conversion of the engineering drawing detail table from the unstructured image to the structured data.

Inventors

CAI JIALI
ZHANG LEI
TIAN MUYANG
WANG WEN
WEN WEN
PENG SIFAN
LUO ANHUA

Assignees

湖南联诚轨道装备有限公司

Dates

Publication Date: 20260512
Application Date: 20251225

Claims (10)

1. The key information extraction method for the engineering drawing detail table is characterized by comprising the following steps of: respectively configuring rotation angles and corresponding perspective transformation strategies for different engineering drawing detail table image samples, and constructing an automatic rotation correction model for correcting the direction deviation between the target engineering drawing detail table image orientation and the standard engineering drawing detail table image sample orientation; dividing different image samples into a target drawing and a non-target drawing, and constructing a non-target drawing filtering model for screening the target drawing; respectively configuring a header region mark and a body region mark for the different image samples, and constructing a target detection model for identifying a key information region; and correcting the direction deviation between the current target image orientation and the standard image sample orientation by utilizing each model, screening the target drawing, and further identifying the key information area to extract the key information.
2. The key information extraction method according to claim 1, wherein the key information includes text content and spatial location information thereof, and wherein in the step of extracting key information, it includes: And carrying out multilingual text detection and recognition in the recognized key information area by adopting a PPOCRv model, so as to obtain the key information.
3. The key information extraction method according to claim 2, characterized in that after the key information is extracted, the key information extraction method further comprises: And carrying out semantic correction and format normalization processing on the extracted key information by adopting a post-processing mechanism based on a regular expression.
4. A key information extraction method according to claim 2 or 3, wherein in the step of extracting key information, further comprising: Constructing a table for accommodating the current text content, and then mapping the current text content to a corresponding cell in the table according to coordinate data matched with the current space position information; after the mapping is completed, the header region text content and the body region text content are aligned.
5. The key information extraction method of claim 4, wherein in constructing a table for accommodating current text contents, comprising: And analyzing the distribution density of rows and columns in the current text content by adopting gradient analysis or an elbow rule to determine the number of rows and columns of the table.
6. The method for extracting key information according to any one of claims 1 to 5, wherein the step of configuring rotation angles and corresponding perspective transformation strategies for the different engineering drawing detail table image samples respectively includes: if the direction deviation between the image sample orientation and the standard image sample orientation is greater than or equal to 90 degrees, determining the rotation angle which can lead the image sample to rotate and meet the direction deviation smaller than 90 degrees and belongs to the integral multiple of 90 degrees, and then formulating the perspective transformation strategy for the new direction deviation obtained after rotation according to the determined rotation angle, or If the direction deviation between the image sample orientation and the standard image sample orientation is smaller than 90 degrees, the perspective transformation strategy is directly formulated aiming at the current direction deviation.
7. The key information extraction method according to any one of claims 1 to 6, characterized in that, The automatic rotation correction model, the non-target drawing filtering model and the model are all obtained based on training of corresponding preset models.
8. The key information extraction method according to any one of claims 1 to 7, characterized in that, The non-target drawings include, but are not limited to, a pattern catalog and borrowing.
9. A computer readable storage medium containing a series of instructions for performing the key information extraction method steps for engineering drawing details tables according to any one of claims 1 to 8.
10. A key information extraction system for engineering drawing detail tables, characterized in that the key information extraction system comprises the following modules: The first model construction module is used for respectively configuring rotation angles and corresponding perspective transformation strategies for different engineering drawing detail table image samples and constructing an automatic rotation correction model for correcting the direction deviation between the target engineering drawing detail table image orientation and the standard engineering drawing detail table image sample orientation; The second model construction module is used for dividing different image samples into a target drawing and a non-target drawing and constructing a non-target drawing filtering model for screening the target drawing; the third model construction module is used for respectively configuring a header region mark and a body region mark for the different image samples and constructing a target detection model for identifying a key information region; And the key information extraction module is used for correcting the direction deviation between the current target image orientation and the standard image sample orientation by utilizing each model, screening the target drawing, and further identifying the key information area so as to extract the key information.

Description

Key information extraction method and system for engineering drawing detail table Technical Field The invention belongs to the technical field of image information identification, and particularly relates to a key information extraction method and system for an engineering drawing detail table. Background In modern equipment manufacturing, engineering drawing lists are used as key process documents to carry a number of important parameters. The traditional information extraction method is highly dependent on manual graph recognition, has the problems of low efficiency, easy error and the like, and is difficult to meet the high requirements of intelligent manufacturing on data driving and flow automation. In recent years, although some researches try to apply OCR or target detection technology to automatically analyze engineering drawing detail tables, many challenges still exist in practical application, such as various forms, non-uniform layout structures, multi-language mixing and other problems commonly existing in the engineering drawing detail tables, so that the generalization capability of a model is limited, the recognition precision of the model on complex layout is low, the logical structure of the table is difficult to accurately restore, most methods rely on direct calling of a general OCR tool chain, and a customized processing mechanism aiming at the characteristics of a process document is lacking. Disclosure of Invention In order to solve the problems, the embodiment of the invention provides a key information extraction method for an engineering drawing list, which comprises the steps of configuring rotation angles and corresponding perspective transformation strategies for different engineering drawing list image samples respectively, constructing an automatic rotation correction model for correcting direction deviation between the direction of a target engineering drawing list image and the direction of a standard engineering drawing list image sample, dividing different image samples into target drawings and non-target drawings, constructing a non-target drawing filtering model for screening the target drawings, configuring header area marks and table body area marks for the different image samples respectively, constructing a target detection model for identifying key information areas, correcting the direction deviation between the current target image direction and the direction of the standard image sample by utilizing each model, screening the target drawings, and further identifying the key information areas to extract key information. Preferably, the key information comprises text content and spatial position information thereof, wherein the step of extracting the key information comprises adopting PPOCRv model to carry out multilingual text detection and recognition in the identified key information area so as to obtain the key information. Preferably, after extracting the key information, the key information extraction method further comprises the step of carrying out semantic correction and format normalization processing on the extracted key information by adopting a post-processing mechanism based on a regular expression. Preferably, the step of extracting the key information further comprises the steps of constructing a table for accommodating the current text content, mapping the current text content to a corresponding cell in the table according to coordinate data matched with the current space position information, and aligning the text content of the header area with the text content of the body area after mapping is completed. Preferably, the process of constructing the table for accommodating the current text content includes analyzing the distribution density of the rows and columns in the current text content using gradient analysis or elbow rule to determine the number of rows and columns of the table. Preferably, the step of respectively configuring the rotation angles and the corresponding perspective transformation strategies for the image samples of different engineering drawing detail tables comprises the steps of firstly determining that the direction deviation is smaller than 90 degrees after the image samples are rotated and belongs to the rotation angles which are integer multiples of 90 degrees if the direction deviation between the direction of the image samples and the direction of a standard image sample is larger than or equal to 90 degrees, and then formulating the perspective transformation strategies for new direction deviation obtained after the rotation according to the determined rotation angles, or directly formulating the perspective transformation strategies for the current direction deviation if the direction deviation between the direction of the image samples and the direction of the standard image sample is smaller than 90 degrees. Preferably, the automatic rotation correction model, the non-target drawing filtering model and the model are