A visual recognition-based full-process core automatic logging method and system

Abstract

The application discloses a full-process core automatic logging method and system based on visual identification, and belongs to the technical field of core logging. The method comprises the following steps: obtaining an original image of a target core box, and establishing associated index information between the original image and a preset engineering-drilling core box three-level hierarchical architecture; sequentially performing field of view distortion correction, illumination normalization processing and label region positioning on the original image, and outputting a standardized core image; using a multi-level deep learning model to perform identification on the standardized core image to output visual identification result data; superimposing the visual identification result data on the standardized core image to generate a labeled prediction image, so that a user can correct the labeled prediction image to obtain a corrected logging image and corresponding corrected identification result data; and storing the corrected logging image and the corresponding corrected identification result data in association according to the associated index information, so that the corrected logging image and the corresponding corrected identification result data can be visualized and traced back based on the engineering-drilling core box three-level hierarchical architecture.

Description

Technical Field

[0001] This application relates to the field of core logging technology, and in particular to a full-process automatic core logging method and system based on visual recognition. Background Technology

[0002] Core logging is a fundamental procedure in mineral exploration, oil and gas exploration, regional geological surveys, and geotechnical engineering investigations. Geological technicians observe, measure, and record the complete sets of core samples retrieved from the boreholes to delineate strata, identify structural fracture zones, and calculate rock quality designations (RQDs). This provides core quantitative evidence for stratigraphic delineation, reservoir evaluation, resource and reserve estimation, and rock mass quality assessment.

[0003] Traditional core logging primarily relies on on-site manual observation and recording by geological technicians: staff observe each core box, recording the start and end depths, the length of intact core segments, and the distribution of fractures and breccia zones. The RQD value is then manually calculated based on industry standards (the percentage of intact core segments longer than 10cm in the total footage of that segment). This manual logging method has the following problems: it requires a large number of technicians to be stationed on-site for extended periods, resulting in high labor costs; subjective differences exist in the judgment of breccia zone boundaries among different personnel, leading to poor consistency in RQD calculation results; and a large number of historical core photographs are stored in a scattered manner, making unified archiving and traceability difficult.

[0004] With the development of computer vision and deep learning technologies, research reports on image recognition technology-assisted core analysis have emerged in recent years. Related studies attempt to process core photographs using algorithms such as object detection, semantic segmentation, and OCR text recognition, aiming to achieve a certain degree of automated cataloging. For example, some studies use convolutional neural networks for feature extraction and transfer learning of core images to establish core recognition models to assist in RQD acquisition; others use instance segmentation models such as Mask R-CNN for core image recognition; still others use object detection algorithms such as YOLO for core detection and combine them with segmentation models for RQD calculation. Furthermore, some solutions also attempt to convert core photographs into digital images and then use geometric transformation techniques for image correction.

[0005] However, existing core logging technologies, whether traditional manual logging methods or image recognition-based assisted solutions, suffer from low logging efficiency, insufficient recognition accuracy, lack of flexible manual correction mechanisms, and non-standard data management. Summary of the Invention

[0006] In view of this, embodiments of this application provide a method and system for automatic core logging throughout the entire process based on visual recognition.

[0007] According to one aspect of this application, a method and system for automated core logging throughout the entire process based on visual recognition are provided, the method comprising: The original image of the target core box is obtained, and the association index information between the original image and the preset three-level hierarchical structure of engineering-drilling-core box is established. The association index information is composed of the engineering level identifier, drilling level identifier and core box level identifier to which the original image belongs, and is used to uniquely determine the belonging path of the original image in the three-level hierarchical structure of engineering-drilling-core box. The original image is sequentially subjected to field distortion correction, illumination normalization processing and label region pre-positioning to output a standardized core image; Using a multi-level deep learning model, the standardized core images are subjected to core structure analysis, label information extraction, and length quantization calculation based on pixel scale conversion, and visual recognition result data is output. The visual recognition result data is superimposed on the standardized core image to generate an annotated prediction image, which allows the user to correct the segmentation contours and recognition text in the annotated prediction image, resulting in a corrected catalog image and corresponding corrected recognition result data. Based on the associated index information, the corrected cataloged image and the corresponding corrected recognition result data are associated and stored, so that the corrected cataloged image and the corresponding corrected recognition result data can be visualized and traced back based on the three-level hierarchical architecture of engineering-drilling-core box.

[0008] Compared with existing technologies, this application has significant advantages: This application employs a pre-defined three-tiered architecture of engineering, borehole, and core box, along with associated index information. This anchors each original image and each recognition result to a unique hierarchical node, solving the problems of scattered data storage, difficulty in unified archiving, and traceability in existing technologies. It achieves unified organization, associated storage, and visualized backtracking of data throughout its entire lifecycle. By sequentially performing field-of-view distortion correction, illumination normalization, and label region pre-positioning on the original images, spatial deformation and illumination interference introduced during the shooting process are eliminated, improving the quality of input images for subsequent recognition. Furthermore, a multi-level deep learning model is used to perform structural analysis, label extraction, and length determination on standardized core images. Quantitative calculations enabled precise differentiation between the core body and fracture zones, as well as accurate recognition of text in the identifier area, improving the automation and accuracy of core logging. By overlaying the visual recognition results onto standardized core images to generate labeled predictive images, users can intuitively correct the segmented contours and recognized text, achieving synergy between flexible manual correction and automatic AI recognition. By associating and storing the corrected logged images and corresponding corrected recognition results based on the associated index information, the corrected data can be visualized and traced back based on a three-level hierarchical architecture, laying the foundation for further integration with external systems and achieving a closed-loop business process.

[0009] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below. Attached Figure Description

[0010] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating a full-process automatic core logging method based on visual recognition provided in an embodiment of this application; Figure 2 This is a schematic diagram of the framework of a full-process automatic core logging system based on visual recognition provided in an embodiment of this application. Detailed Implementation

[0011] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0012] This embodiment provides a full-process automatic core logging method based on visual recognition, such as... Figure 1 As shown, this includes S110 to S150: S110: Obtain the original image of the target core box and establish the association index information between the original image and the preset three-level hierarchical structure of engineering-drilling-core box. The association index information is composed of the engineering level identifier, drilling level identifier and core box level identifier to which the original image belongs, and is used to uniquely determine the belonging path of the original image in the three-level hierarchical structure of engineering-drilling-core box.

[0013] In some embodiments, the original image of the target core box is obtained by at least one of the following methods: (1) Data collection via mobile shooting app: The shooting app has built-in shooting guidelines to help users place the standard ruler and adjust the shooting angle to reduce perspective distortion; (2) Import existing core box image files in batches using desktop software; (3) Receive images pushed by external devices in real time via local area network or Internet.

[0014] After acquiring the original images, they are stored in a pre-set storage system. The storage system adopts a hybrid storage architecture: the associated index information and shooting parameter metadata of the original images are stored as structured data in a relational database, while the original images themselves are stored as unstructured data in a digital warehouse system. The relational database and the digital warehouse system are linked through the associated index information to ensure that the structured metadata and unstructured image files of each original image can be retrieved and traced through the same index information.

[0015] Relational databases (such as PostgreSQL) are database systems that organize data using a relational model. Their core feature is storing data in tables as rows and columns, with tables linked by primary and foreign keys. They are primarily used for storing structured data. Digital warehousing systems (such as DSpace) are systems specifically designed for the long-term preservation and management of various digital resources. They can collect, store, index, and save files in any digital format. They primarily store unstructured data (such as images and documents).

[0016] Here, the three-tiered architecture of engineering—drilling—core box serves as the pre-defined data organization framework for this application. It is used to hierarchically organize and manage image data, identification result data, and logging outcome data involved in the entire core logging process, providing a unified hierarchical index foundation for the associated storage, retrieval, and visualization of data in subsequent steps S120 to S150. The three-tiered architecture of engineering—drilling—core box includes an engineering level, a drilling level, and a core box level; and within this three-tiered architecture: (1) Engineering level, which includes at least one engineering node, each engineering node is bound to an exploration project and has a corresponding engineering level identifier; (2) Drilling level, associated with the engineering level, includes at least one drilling node. Each drilling node is used to bind the location information and drilling depth information of a borehole under the exploration project bound to the engineering level, and each drilling node corresponds to a drilling level identifier. (3) Core box level, associated with borehole level, includes at least one core box node. Each core box node is used to bind the image of a core box under the borehole associated with the borehole level and the corresponding logging results information. Each core box node corresponds to a core box level identifier.

[0017] It should be noted that each exploration project corresponds to one engineering node, and each engineering node is associated with at least one borehole node, which in turn is associated with at least one core box node. In other words, an exploration project contains multiple boreholes, each borehole corresponds to multiple core boxes, forming a top-down tree-like hierarchical structure. The engineering level identifier uniquely identifies an engineering node within the system; the borehole level identifier uniquely identifies a borehole node within its respective engineering node (the same borehole level identifier is allowed to appear under different engineering nodes); and the core box level identifier uniquely identifies a core box node within its respective borehole node (the same core box level identifier is allowed to appear under different borehole nodes).

[0018] Based on this, the associated index information is specifically used to: anchor the original image to the corresponding core box node in the three-level hierarchical architecture of engineering-drilling-core box, and associate it with the corresponding drilling node and engineering node level by level through the association relationship between the levels.

[0019] Through the above-mentioned three-level hierarchical architecture of engineering-drilling-core box, each original core box image and its subsequent generated preprocessed images, predicted images, corrected and cataloged images, visual recognition result data, corrected recognition result data, cataloged result information, etc., are all anchored to the corresponding core box node through the association index information, and then associated with the corresponding drilling node and engineering node, realizing the unified organization, association storage and visualization backtracking of data throughout the entire life cycle.

[0020] S120: Perform field distortion correction, illumination normalization, and label region pre-positioning on the original image in sequence to output a standardized core image.

[0021] Specifically, S120 includes: acquiring a corrected image obtained after the original image has undergone field distortion correction and illumination normalization processing; performing size normalization processing on the corrected image to unify the spatial resolution and color space of the image, thereby obtaining a unified core image; based on a target detection model, performing region of interest pre-location on the unified core image to crop out a target region containing the main core region and the identifier region from the unified core image, and outputting the target region as a standardized core image; wherein, the identifier region includes at least one of the box number label region and the depth label region.

[0022] Here, field-of-view distortion correction includes: performing geometric transformations on the original image based on the core box corner detection results to straighten the perspective-distorted image into a frontal view, eliminating spatial distortion caused by the shooting angle. Core box corner detection is implemented based on Hough line detection or object detection models.

[0023] Illumination normalization processing includes performing at least one of color equalization, contrast stretching, and homomorphic filtering on the original image to eliminate interference from shooting reflections, local shadows, and uneven illumination, and restore the texture characteristics of the rock core.

[0024] It is important to note that the order of field distortion correction and illumination normalization is not interchangeable; field distortion correction must be performed first, followed by illumination normalization. This is because geometric transformations alter the spatial positions of pixels. If illumination normalization is performed before field distortion correction, the corrected interpolation calculations will change the pixel value distribution, causing a shift in the effect of illumination normalization.

[0025] Here, the object detection model is used to select the core body, standard scale, and identifier regions in the unified core image to narrow the computational scope of the subsequent AI recognition stage and improve the inference speed.

[0026] Furthermore, in the above-mentioned S120 image preprocessing stage, the set of adjustable parameters involved in each processing step is visualized, allowing users to manually fine-tune them according to the actual image quality in order to obtain the optimal imaging effect.

[0027] S130: Utilizing a multi-level deep learning model, it performs core structure analysis, label information extraction, and length quantization calculation based on pixel scale conversion on standardized core images, and outputs visual recognition result data.

[0028] Here, the multi-level deep learning model includes a semantic segmentation model, an instance segmentation model, and a character recognition model. These three models can be implemented using the following specific models, but are not limited to: the semantic segmentation model can use U... Net, DeepLabV3+, or YOLO segmentation models; instance segmentation models can use Mask R. CNN or YOLO instance segmentation models; character recognition models can use PaddleOCR or EasyOCR.

[0029] S130 specifically includes: using a semantic segmentation model to perform semantic segmentation of the core body in a standardized core image, extracting the core body mask, and obtaining the core body mask; within the core body mask area, using an instance segmentation model to perform instance segmentation, dividing the image into a first segmentation region contour set for referring to continuous structural regions and a second segmentation region contour set for referring to discontinuous structural regions; calculating the span dimensions of the structural regions contained in the first and second segmentation region contour sets respectively to generate rock quality indicators; using a character recognition model to perform character recognition on the identifier regions in the standardized core image, extracting character recognition data; and outputting the first segmentation region contour set, the second segmentation region contour set, the character recognition data, the span dimensions of the first and second segmentation region contour sets, and the rock quality indicators as visual recognition result data.

[0030] It should be noted that in core logging scenarios, a continuous structural region refers to a geologically intact section without obvious cracks or fractures. This continuous structural region appears as a closed two-dimensional area on the image, defined by the boundary coordinate data set output by the instance segmentation model. The length value of the structural region measures the span dimension along the length direction of the core (i.e., the borehole depth direction), not its width or area. A discontinuous structural region refers to a geologically intact section of the core containing cracks, joints, or fracture zones. This discontinuous structural region appears as one or more closed two-dimensional areas on the image, defined by the boundary coordinate data set output by the instance segmentation model. When multiple spatially separated discontinuous structural regions exist, the cumulative length value is obtained by calculating the span dimensions of each discontinuous structural region separately and then summing them.

[0031] Then, the span dimensions of the structural regions contained in the first segmented region contour set and the second segmented region contour set are calculated respectively to generate rock quality indicators. This includes: establishing a conversion ratio between pixels and actual length based on the scale in the standardized core image or a preset standard physical reference; based on the conversion ratio, calculating the span dimension along the core length direction of each continuous structural region in the first segmented region contour set, and the cumulative span dimension along the core length direction of the discontinuous structural regions in the second segmented region contour set; filtering out the span dimensions in the first segmented region contour set that are greater than a preset threshold, and summing them to obtain a first accumulated value; adding the span dimensions of each continuous structural region in the first segmented region contour set and the cumulative span dimensions of the discontinuous structural regions in the second segmented region contour set to obtain a total length value; and dividing the first accumulated value by the total length value to obtain the rock quality indicators.

[0032] Here, the preset threshold is 10cm, corresponding to the minimum length of a complete core segment used in the calculation of the Rock Quality Index (RQD) in geological industry standards. The formula for calculating the RQD value is: .

[0033] Furthermore, the character recognition model is an OCR model specifically trained on core label samples. The identifier region contains at least one of the box number label region and the depth label region. The box number label region records the core box number information, and the depth label region records the borehole start and end depth numerical information corresponding to that core box. The recognized character data includes the recognized text content and its position information in the standardized core image.

[0034] It should be noted that the inference order of the three models in S130 is not interchangeable. The reason is as follows: the semantic segmentation model is used to determine the core body mask, limiting the computational area of ​​the instance segmentation model and avoiding processing irrelevant background; since it relies on the core body mask to distinguish between continuous and discontinuous structural regions, the instance segmentation model must be executed after the semantic segmentation model is completed; the length quantization calculation targets the first and second segmentation region contours output by the instance segmentation model, therefore it must be executed after the instance segmentation model is completed; the character recognition model processes the identifier region, which is located differently in the image from the core body region processed by the segmentation model and does not overlap with it. Therefore, the character recognition model can be executed in parallel with the segmentation model, or before or after the segmentation model, and the order does not affect the recognition result. However, in the multi-model cascade architecture of this application, for ease of data management and process control, the character recognition model is scheduled to be executed after the segmentation model is completed.

[0035] S140: The visual recognition result data is superimposed on the standardized core image to generate an annotated prediction image, which allows users to correct the segmentation contours and recognition text in the annotated prediction image, resulting in a corrected catalog image and the corresponding corrected recognition result data.

[0036] In some embodiments, the first segmentation region contour set and the second segmentation region contour set do not overlap spatially, and their union completely covers the range of the core body mask, meaning that they together constitute a complete division of the core body mask. Therefore, the first segmentation region contour set and the second segmentation region contour set are superimposed on the standardized core image with different colors, and character recognition data and rock quality indicators are displayed in the interface to generate an annotated prediction image.

[0037] For example, the method of layering different colors involves rendering the first segmented region outline set with a first color (e.g., green) and the second segmented region outline set with a second color (e.g., red), where the first and second colors are different. Users can intuitively distinguish the distribution of complete core segments (first color) and fractured zones (second color) by color. Character recognition data is displayed as text boxes superimposed next to or above the corresponding identifier area, and rock quality indicators are displayed in the interface as numerical labels.

[0038] In some embodiments, the corrections performed by the user include dragging and adjusting the segmentation contours in the labeled prediction image, adding or deleting the second segmentation region contour set, and / or modifying the character recognition data in the labeled prediction image. For example, dragging the boundary points of the first or second segmentation region contour set to adjust the precise position of the segmentation boundary; adding a fracture zone region in the image (i.e., adding a region labeled as the second segmentation region contour set within the main core mask area), or deleting a labeled fracture zone region (i.e., removing the labeling of a certain region as the second segmentation region contour set); modifying the text content with incorrect recognition in the character recognition data (such as correcting incorrect box numbers or depth values); recalculating and confirming the rock quality indicators, etc.

[0039] Here, the corrected catalog image refers to the visualization image generated by overlaying the user-corrected segmentation contours, corrected character recognition data, and rock quality indicators onto a standardized core image. The corrected recognition result data refers to the structured data set corresponding to the corrected catalog image, containing corrected segmentation contour coordinate data, corrected character recognition data, corrected length calculation values, and corrected rock quality indicators.

[0040] S150: Based on the associated index information, the corrected cataloged image and the corresponding corrected recognition result data are associated and stored, so that the corrected cataloged image and the corresponding corrected recognition result data can be visualized and traced back based on the three-level hierarchical architecture of engineering-drilling-core box.

[0041] Here, we first describe the overall architecture and storage logic of data storage in this application. As shown in S110, this application adopts a hybrid storage architecture for unified storage management of data throughout the entire process. The hybrid storage architecture consists of two components: a relational database and a digital warehouse system. The two establish data associations through related index information, ensuring that each data object can be retrieved and traced back through the same index information. Specifically, the relational database is used to store structured data (such as corrected identification result data, operation logs, shooting parameter metadata, etc.), and the digital warehouse system is used to store unstructured data (such as corrected cataloged images, original images, preprocessed images, labeled predicted images, etc.). The two establish associations through related index information, enabling structured data and unstructured image files of the same core box node to be mutually retrieved and synchronously presented.

[0042] Based on the aforementioned hybrid storage architecture, the corrected cataloged images and their corresponding corrected recognition results are stored together according to the associated index information. This includes: writing the corrected recognition results data as structured data into the relational database under the hierarchical nodes corresponding to the engineering level identifier, borehole level identifier, and core box level identifier, according to the associated index information; storing the corrected cataloged images in the digital warehouse system under the hierarchical paths corresponding to the engineering level identifier, borehole level identifier, and core box level identifier, according to the associated index information; and establishing the association between the corrected cataloged images and the corrected recognition results data through the associated index information, so as to realize the visual backtracking of the corrected cataloged images and their corresponding corrected recognition results data based on the three-level hierarchical architecture of engineering-borehole-core box.

[0043] Here, the specific implementation method for writing to the hierarchical nodes corresponding to the engineering level identifier, borehole level identifier, and core box level identifier in the relational database is as follows: The relational database pre-constructs a data table structure corresponding to a three-level hierarchical architecture of engineering, borehole, and core box, including an engineering table, a borehole table, and a core box record table. The engineering table contains an engineering level identifier field and exploration project information fields; the borehole table contains a borehole level identifier field, a foreign key field related to the engineering level, and borehole location and depth information fields; the core box record table contains a core box level identifier field, a foreign key field related to the borehole level, and logging result information fields. During writing, based on the engineering level identifier, borehole level identifier, and core box level identifier in the associated index information, the corresponding engineering node, borehole node, and core box node are located respectively. The corrected identification result data is written to the record in the core box record table corresponding to the core box level identifier, and then linked level by level to the corresponding borehole node and engineering node through foreign key associations.

[0044] Furthermore, the specific implementation method for storing the data in the digital warehousing system under the hierarchical paths corresponding to the engineering level identifier, borehole level identifier, and core box level identifier is as follows: The digital warehousing system pre-establishes a folder directory structure corresponding to the three-level hierarchy of engineering—drilling—core box, including engineering-level directories, drilling-level directories, and core box-level directories. During storage, based on the associated index information, a storage path corresponding to the three-level hierarchy of engineering—drilling—core box (e.g., " / engineering-level identifier / drilling-level identifier / core box-level identifier / ") is generated. The corrected and cataloged images are stored under this path, ensuring that the physical storage location of the image files corresponds one-to-one with their respective hierarchical nodes.

[0045] Using the storage method described above, each corrected and cataloged image and its corresponding corrected recognition result data are anchored to a unique node in the three-tiered architecture of engineering—borehole—core box through associated index information. Users can search using any dimension of engineering level identifier, borehole level identifier, or core box level identifier, locating the target core box node level by level, and viewing all the data stored under that node, including the original image, images from each intermediate processing stage, corrected and cataloged images, and corrected recognition result data. Users can also view the details of each correction operation recorded in the operation log (operator, modification time, modification content), thereby achieving visualized backtracking of the entire lifecycle of data.

[0046] Based on the aforementioned storage, this method further includes: responding to user input instructions, performing annotation information overlay processing on the corrected cataloged image to generate a visual verification chart; calling preset report templates and / or graph templates to format the corrected identification result data to generate a standardized cataloged report and / or a rock quality index-depth variation graph; and outputting the visual verification chart, standardized cataloged report, and / or rock quality index-depth variation graph as a visual retrospective form of the corrected cataloged image and the corresponding corrected identification result data through a standardized interface for visualization presentation or further analysis by external systems.

[0047] The preset report template is a standardized table template in Excel or CSV format, which includes fields such as project name, borehole number, core box number, start and end depth, length of each continuous segment, cumulative length of fracture zone, total length, rock quality index (RQD), operator, and correction time; the curve template is a preset two-dimensional coordinate graph template with borehole depth as the vertical axis and RQD as the horizontal axis, used to generate a curve reflecting the trend of RQD value with borehole depth.

[0048] Here, the annotation information overlay processing refers to overlaying the length values ​​of each continuous segment, the rock quality index values, and the corresponding annotation boxes or text labels onto the corrected and cataloged image. For example, the span dimensions are marked above each complete core segment, and the RQD values ​​are marked in the blank areas or fixed areas of the image, forming a high-definition verification image that can be directly printed or archived. The visual verification image can serve as a direct source for the attached diagrams in the exploration report.

[0049] In some embodiments, the standardized interface is an API interface conforming to RESTful specifications or an interface conforming to industry data exchange standards, used for data exchange with external geological databases, resource estimation software, or 3D geological modeling platforms. For example, rock quality indicators and layer depth fields can be automatically backfilled into external geological databases for use in resource estimation or 3D geological modeling; alternatively, visual verification maps can be output to external report generation systems for direct use in the preparation of attached figures in exploration reports, realizing a closed-loop business process from core logging to results archiving.

[0050] Corresponding to the aforementioned vision-based automatic core logging method, this application also provides a vision-based automatic core logging system. For example... Figure 2 As shown, the system includes: (1) Data acquisition layer, used to acquire the original image of the target core box and establish the association index information between the original image and the preset three-level hierarchical structure of engineering-drilling-core box. The association index information is composed of the engineering level identifier, drilling level identifier and core box level identifier to which the original image belongs, and is used to uniquely determine the belonging path of the original image in the three-level hierarchical structure of engineering-drilling-core box.

[0051] The data acquisition layer acquires raw images through at least one of the following methods: mobile camera app, batch image import from desktop, or real-time image push from the network. The raw images are then linked and bound to the three-level hierarchical architecture through the engineering-drilling-core box three-level data organization module.

[0052] (2) Image preprocessing layer, used to perform field distortion correction, illumination normalization and label region pre-positioning on the original image in sequence, and output standardized core image.

[0053] The image preprocessing layer includes an image distortion correction module for performing field distortion correction on the original image, a light and shadow equalization correction module for performing illumination normalization on the original image, an image normalization module for performing size normalization on the corrected image to unify the spatial resolution and color space of the image, and a label region pre-positioning module. The label region pre-positioning module is used to pre-position the region of interest in the unified core image to crop the target region containing the main core region and the identifier region from the unified core image, and output the target region as the normalized core image.

[0054] (3) AI recognition and quantization calculation layer, which uses a multi-level deep learning model to perform core structure analysis, label information extraction, and length quantization calculation based on pixel scale conversion on standardized core images, and outputs visual recognition result data.

[0055] The AI ​​recognition and quantization calculation layer includes the YOLO object detection model, U... The Net / DeepLabV3+ semantic segmentation model, the specially trained PaddleOCR text recognition module, and the core length and RQD quantization calculation unit.

[0056] (4) Human-machine collaborative calibration layer, which is used to overlay the visual recognition result data onto the standardized core image to generate annotated prediction image, so that users can correct the segmentation contours and recognition text in the annotated prediction image and obtain the corrected catalog image and the corresponding corrected recognition result data.

[0057] The human-machine collaborative verification layer includes an image layering visualization unit, an interactive manual correction tool module, a multi-person collaborative annotation control unit, and an operation log recording module. The image layering visualization unit is used to overlay the first segmentation region contour set and the second segmentation region contour set onto the standardized core image using different colors. The interactive manual correction tool module is used to respond to user correction operations on the segmentation contours and recognized text. The multi-person collaborative annotation control unit is used to support multiple people to simultaneously annotate the same borehole data. The operation log recording module is used to record the operator, modification time, and modification content corresponding to the correction operation.

[0058] (5) Data storage and output layer, which is used to store the corrected cataloged image and the corresponding corrected recognition result data in association according to the associated index information, so that the corrected cataloged image and the corresponding corrected recognition result data can be visualized and traced back based on the three-level hierarchical architecture of engineering-drilling-core box.

[0059] The data storage and output layer comprises a hybrid storage area and a results output area. The hybrid storage area includes a relational database (such as PostgreSQL, a structured database) and a digital warehousing system (such as DSpace, an unstructured image repository), used to store structured and unstructured data respectively. The results output area includes a visual verification image generation unit, a standardized report export module, an external system interface API, and a multi-user hierarchical permission management module. Specifically, the visual verification image generation unit overlays annotation information onto corrected cataloged images to generate a visual verification map; the standardized report export module uses preset report templates and / or graph templates to format the corrected identification results data, generating standardized cataloged reports and / or rock quality index-depth variation curves; the external system interface API outputs the visual verification map, standardized cataloged reports, and / or rock quality index-depth variation curves to external systems; and the multi-user hierarchical permission management module controls access and operation permissions for data.

[0060] Compared with existing technologies, this application has significant advantages: This application employs a pre-defined three-tiered architecture of engineering, drilling, and core box, along with associated indexing information. This ensures that each original image and each recognition result is anchored to a unique hierarchical node, addressing the issues of scattered data storage, difficulty in unified archiving, and traceability in existing technologies. It achieves unified organization, associated storage, and visual backtracking of data throughout its entire lifecycle, improving the data consistency and manageability of cataloged results. By sequentially correcting field-of-view distortion and illumination, it eliminates spatial deformation and lighting interference introduced during the shooting process, improving the quality of input images for subsequent recognition. Finally, through multi-level deep learning models and sequential inference, it achieves core box... The system improves the accuracy of core structure analysis and text recognition by extracting the main mask, finely distinguishing between complete segments and fractured zones, and accurately identifying text in the identifier area. By displaying the segmented contours in layers with different colors and providing interactive correction tools, users can intuitively drag, adjust, add, delete, and modify the recognition results, achieving a synergy between flexible manual correction and automatic AI recognition. At the same time, by generating and outputting visual verification charts, standardized logging reports, and / or graphs, the system provides standardized data interfaces for external systems, laying the foundation for further realizing a closed-loop business process from core logging to results archiving.

[0061] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0062] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A fully automated core logging method based on visual recognition, characterized in that, The method includes: The original image of the target core box is obtained, and the association index information between the original image and the preset three-level hierarchical structure of engineering-drilling-core box is established. The association index information is composed of the engineering level identifier, drilling level identifier and core box level identifier to which the original image belongs, and is used to uniquely determine the belonging path of the original image in the three-level hierarchical structure of engineering-drilling-core box. The original image is sequentially subjected to field distortion correction, illumination normalization processing and label region pre-positioning to output a standardized core image; Using a multi-level deep learning model, the standardized core images are subjected to core structure analysis, label information extraction, and length quantization calculation based on pixel scale conversion, and visual recognition result data is output. The visual recognition result data is superimposed on the standardized core image to generate an annotated prediction image, which allows the user to correct the segmentation contours and recognition text in the annotated prediction image, resulting in a corrected catalog image and corresponding corrected recognition result data. Based on the associated index information, the corrected cataloged image and the corresponding corrected recognition result data are associated and stored, so that the corrected cataloged image and the corresponding corrected recognition result data can be visualized and traced back based on the three-level hierarchical architecture of engineering-drilling-core box.

2. The full-process automatic core logging method based on visual recognition according to claim 1, characterized in that, The three-tiered architecture of engineering-drilling-core box includes an engineering level, a drilling level, and a core box level; and within this three-tiered architecture: The engineering hierarchy includes at least one engineering node, each engineering node is bound to an exploration project and has a corresponding engineering hierarchy identifier; The borehole level, associated with the engineering level, includes at least one borehole node. Each borehole node is used to bind the location and depth information of a borehole under the exploration project bound to the engineering level, and each borehole node corresponds to a borehole level identifier. The core box level, associated with the borehole level, includes at least one core box node. Each core box node is used to bind the image of a core box belonging to the borehole associated with the borehole level and the corresponding logging results information, and each core box node corresponds to a core box level identifier.

3. The full-process automatic core logging method based on visual recognition according to claim 1, characterized in that, The process of sequentially performing field distortion correction, illumination normalization, and label region pre-location on the original image to output a standardized core image includes: The corrected image is obtained by processing the original image after field distortion correction and illumination normalization; The corrected image is subjected to size normalization processing to unify the spatial resolution and color space of the image, resulting in a unified core image; Based on the target detection model, the region of interest is pre-located in the unified core image to crop out the target region containing the main core region and the identifier region from the unified core image, and the target region is output as the standardized core image; wherein, the identifier region includes at least one of the box number label region and the depth label region.

4. The full-process automatic core logging method based on visual recognition according to claim 3, characterized in that, The multi-level deep learning model includes a semantic segmentation model, an instance segmentation model, and a character recognition model; The process utilizes a multi-level deep learning model to perform core structure analysis, label information extraction, and length quantization calculation based on pixel scale conversion on the standardized core images, outputting visual recognition result data, including: Using the semantic segmentation model, the standardized core image is subjected to semantic segmentation of the core subject, and the core subject mask is extracted to obtain the core subject mask; Within the masked area of ​​the core body, instance segmentation is performed using the instance segmentation model to divide the area into a first segmentation region contour set for referring to continuous structural regions and a second segmentation region contour set for referring to discontinuous structural regions. Calculate the span dimensions of the structural regions contained in the first segmented region contour set and the second segmented region contour set respectively to generate rock quality indicators; Using the character recognition model, character recognition is performed on the identifier regions in the standardized core image to extract character recognition data; The first segmented region contour set, the second segmented region contour set, the character recognition data, the span size corresponding to each of the first segmented region contour set and the second segmented region contour set, and the rock quality index are output as the visual recognition result data.

5. The full-process automatic core logging method based on visual recognition according to claim 4, characterized in that, The step of calculating the span dimensions of the structural regions contained in the first segmented region contour set and the second segmented region contour set respectively to generate rock quality indicators includes: Based on the scale or preset standard physical reference in the standardized core image, establish the conversion ratio between pixels and actual length; Based on the conversion ratio, the span dimension of each continuous structural body region in the first segmented region contour set along the core length direction is calculated, and the cumulative span dimension of the discontinuous structural body region in the second segmented region contour set along the core length direction is calculated. Filter the span dimensions of the first segmented region contour set that are greater than a preset threshold, and sum them to obtain the first accumulated value; The total length value is obtained by adding the span dimensions of each continuous structural region in the first segmented region contour set and the cumulative span dimensions of the non-continuous structural regions in the second segmented region contour set. Divide the first accumulated value by the total length value to obtain the rock quality index.

6. The full-process automatic core logging method based on visual recognition according to claim 4, characterized in that, The step of overlaying the visual recognition result data onto the standardized core image to generate an labeled predicted image includes: The first segmented region contour set and the second segmented region contour set are superimposed onto the standardized core image in different colors, and the character recognition data and the rock quality index are displayed on the interface to generate a labeled prediction image.

7. The full-process automatic core logging method based on visual recognition according to claim 1, characterized in that, The correction includes dragging and adjusting the segmentation contours in the labeled prediction image, adding or deleting the second segmentation region contour set, and / or modifying the character recognition data in the labeled prediction image.

8. The full-process automatic core logging method based on visual recognition according to claim 2, characterized in that, The step of associating and storing the corrected cataloged image and the corresponding corrected recognition result data according to the associated index information includes: Based on the associated index information, the corrected identification result data is written as structured data into the relational database under the hierarchical nodes corresponding to the engineering level identifier, the borehole level identifier, and the core box level identifier; Based on the associated index information, the corrected cataloged image is stored in the digital warehousing system under the hierarchical path corresponding to the engineering level identifier, the borehole level identifier, and the core box level identifier; By using the associated index information, a relationship is established between the corrected cataloged image and the corrected identification result data, so as to realize the visual backtracking of the corrected cataloged image and the corresponding corrected identification result data based on the three-level hierarchical architecture of engineering-drilling-core box.

9. The full-process automatic core logging method based on visual recognition according to claim 1 or 8, characterized in that, The method further includes: In response to the user's input instruction, the corrected cataloged image is overlaid with annotation information to generate a visual verification image; Call the preset report template and / or graph template to format the corrected identification result data and generate a standardized catalog report and / or a rock quality index-depth change graph; The visualization verification chart, the standardized logging report, and / or the rock quality index-depth variation curve are used as the visualization retrospective form of the corrected logging image and the corresponding corrected identification result data. They are output to external systems through a standardized interface for visualization or further analysis.

10. A fully automated core logging system based on visual recognition, characterized in that, The system includes: The data acquisition layer is used to acquire the original image of the target core box and establish the association index information between the original image and the preset three-level hierarchical structure of engineering-drilling-core box. The association index information is composed of the engineering level identifier, drilling level identifier and core box level identifier to which the original image belongs, and is used to uniquely determine the belonging path of the original image in the three-level hierarchical structure of engineering-drilling-core box. The image preprocessing layer is used to sequentially perform field distortion correction, illumination normalization processing and label region pre-positioning on the original image, and output a standardized core image. The AI ​​recognition and quantization calculation layer is used to perform core structure analysis, label information extraction, and length quantization calculation based on pixel scale conversion on the standardized core image using a multi-level deep learning model, and output visual recognition result data. The human-machine collaborative calibration layer is used to overlay the visual recognition result data onto the standardized core image to generate an annotated prediction image, which allows users to correct the segmentation contours and recognition text in the annotated prediction image, and obtain the corrected catalog image and the corresponding corrected recognition result data. The data storage and output layer is used to associate and store the corrected cataloged image and the corresponding corrected recognition result data according to the associated index information, so that the corrected cataloged image and the corresponding corrected recognition result data can be visualized and traced back based on the three-level hierarchical architecture of engineering-drilling-core box.

Images