Chip package drawing intelligent recognition and standardization conversion method

Abstract

The application provides a chip packaging drawing intelligent recognition and standardized conversion method, relates to the fields of intelligent manufacturing and semiconductor packaging technology, and comprises the following steps: converting a drawing into a bitmap image, completing two-level classification of customers and packaging types by relying on an identification model and a packaging identification model, and adaptively matching an OCR coordinate template, a connection rule template and a packaging standard template. A multi-candidate region OCR is combined with a multi-dimensional weighted scoring mechanism at the text end to improve the character recognition accuracy. At the graphic end, six types of YOLOv8 special models work cooperatively to complete packaging detection, wafer extraction, connection positioning, geometric distortion correction and graphic reconstruction in sequence, and realize full-automatic standardized conversion of the drawing. The scheme solves the problems of low analysis precision, poor adaptability and inability of full-automatic conversion of traditional technical drawings, can effectively shorten the drawing circulation period, and improve the production efficiency and product consistency.

Description

Technical Field

[0001] This application relates to the fields of intelligent manufacturing and semiconductor packaging technology, and in particular to a method for intelligent recognition and standardized conversion of chip packaging drawings. Background Technology

[0002] In the field of intelligent manufacturing for semiconductor chip packaging, the packaging drawings delivered by customers come from a wide range of sources and are diverse in style. Drawings from different cooperating manufacturers have issues such as inconsistent file formats, varying layouts, and significant differences in drawing styles. Even packaging drawings from different production batches of the same customer often show variations in text and image arrangement, annotation positions, and graphic orientation. Currently, the industry mainly uses three methods to handle such non-standardized drawings: manual review and conversion, traditional single optical character recognition (OCR), and basic image recognition technology.

[0003] In the manual processing mode, staff need to analyze the text parameters of the drawings page by page, manually extract the packaging graphics, replicate the wafer and bonding line layout and match it with the company's internal standard template. This not only results in low work efficiency and long drawing review and circulation cycles, but also makes it easy for human negligence to cause input errors and graphic replication deviations, which directly affect the yield and batch consistency of subsequent packaging production.

[0004] Traditional OCR technology mostly adopts the full-image scanning recognition mode, without setting regional filtering rules for chip packaging drawings. Background noise, text and image overlap, and offset annotation positions can all cause character misrecognition and omission. At the same time, conventional OCR lacks classification and adaptation logic, and cannot dynamically switch recognition rules according to different customers and different packaging types, resulting in extremely poor adaptability to drawings with varied layouts.

[0005] Conventional image recognition solutions mostly use a single deep learning model, which can only recognize fixed-style packaging graphics. They cannot distinguish between multiple types of packaging structures, accurately extract wafer areas, or automatically correct geometric distortions such as drawing rotation and mirroring. They are also unable to automatically migrate non-standard graphics to standard packaging templates and reconstruct bonding connections. They can only achieve simple graphic classification and cannot meet the standardized conversion requirements of the entire process.

[0006] Some existing technologies attempt to combine deep learning for drawing analysis, but they generally suffer from problems such as a single model architecture and independent text and image processing. They have not established a collaborative mechanism for text recognition and image analysis, nor have they designed hierarchical classification and result verification strategies for the diversity of drawings. When faced with complex working conditions such as irregularly shaped wafers, multi-chip composite packaging, and low-resolution drawings, the recognition and conversion effects drop significantly, and the overall solution's robustness and scenario adaptability cannot be adapted to large-scale semiconductor production scenarios. Summary of the Invention

[0007] The purpose of this invention is to provide a method for intelligent recognition and standardized conversion of chip packaging drawings, in order to solve the common technical problems in the current chip packaging manufacturing process, such as the large variety of customer drawings, inconsistent formats, and significant differences in layout style between different batches of drawings from the same customer.

[0008] The above-mentioned objective of this application is achieved through the following technical solution: The method includes a parallel text and image processing flow, which comprises: a standardized text recognition flow for drawings and a standardized reconstruction flow for packaged graphics. The method is implemented based on multiple pre-trained YOLOv8 deep learning models, a hierarchical template library, and an OCR scoring and filtering model. The YOLOv8 deep learning models include: a marker recognition model, a package recognition model, a die-bonding detection model, a line recognition model, a positional relationship model, and a refined classification model. The hierarchical template library stores OCR coordinate templates, package standard templates, and connection rule templates corresponding to different customers and different package types. The method specifically includes the following steps: S1: Convert the packaging drawing files submitted by the customer into high-resolution bitmap images page by page, which will be used as the original input images for subsequent image and text recognition; S2: Based on the logo recognition model, identify the enterprise logo in the original input image to complete the first level of customer category classification; based on the encapsulation recognition model, identify the encapsulation graphic recognition result in the original input image to complete the second level of encapsulation type subdivision; based on the two-level classification results, retrieve the matching OCR coordinate template, encapsulation standard template and connection rule template from the hierarchical template library; S3: Using the OCR coordinate template and the connection rule template, the text standardization recognition process and the encapsulated graphic standardization reconstruction process are executed in parallel to obtain standardized parameter data and standardized encapsulated images, respectively. S4: Integrate standardized parameter data and standardized packaging images into a unified packaging standard template to generate the final standardized chip packaging drawing.

[0009] Optionally, all YOLOv8 deep learning models are trained using transfer learning, and the training dataset consists of factory history labeled and encapsulated drawings. Data augmentation strategies are employed during model training, including random image flipping, adaptive brightness adjustment, and Gaussian noise superposition.

[0010] Optionally, the two-level classification results are specifically as follows: First-level classification: By identifying the company logo and specific text fields in the drawings using the logo recognition model, the customer's category is determined; Second-level classification: Based on the output of the packaging identification model, and combined with the product series, the packaging sub-types are divided. By linking the two-level classification results to call the corresponding OCR coordinate template and encapsulated standard template, adaptive template matching for multiple clients and encapsulation types can be achieved.

[0011] Optionally, the standardized text recognition process for drawings includes the following steps: Target Area Cropping: Based on the retrieved OCR coordinate template, locate and crop multiple candidate Regions of Interest (ROIs) corresponding to the key parameter fields of the drawing. Partitioned OCR recognition: Optical character recognition is performed independently for each cropped candidate region of interest (ROI), resulting in multiple sets of candidate recognition results; Multi-dimensional comprehensive scoring: Preset scoring indicators are used to score the candidate recognition results of each group in turn to obtain a comprehensive score; the scoring indicators include: format matching degree, character confidence degree, regional clarity and historical data consistency; Results screening and post-processing: The candidate recognition results with the highest comprehensive score are selected as the final text recognition results. The recognized text of the candidate recognition results is processed by unit conversion and numerical rounding to obtain standardized parameter data. Parameter entry: Fill the standardized parameter data into the data table corresponding to the packaging standard template.

[0012] Optionally, the comprehensive score is calculated using a weighted scoring formula.

[0013] Optionally, the encapsulated graphic standardization reconstruction process includes the following steps: Encapsulation type recognition: Call the encapsulation recognition model to detect the original input image and output the encapsulation category and the corresponding independent encapsulation image; Wafer region extraction: Based on the die bonding detection model, the die bonding region is located within the independent package image, cropped to obtain the wafer image, and the wafer center coordinates are recorded; Bonding coordinate detection: The line recognition model is invoked to extract the initial position coordinates of the bonding lines on the wafer image; Geometric distortion correction: The rotation angle and mirror state of the independent package image relative to the package standard template are identified by the positional relationship model, an affine transformation matrix is ​​generated, and the wafer image is corrected by the affine transformation matrix to make the wafer image orientation consistent with the package standard template. Subtype refinement differentiation: Input the encapsulation category and independent encapsulation image into the refinement classification model, identify the subtype label of the current encapsulation, and load the refinement template parameters of the corresponding subtype; Graphic splicing and connection reconstruction: The corrected wafer image is superimposed onto the corresponding area of ​​the package standard template according to the coordinate position defined by the refined template parameters; By combining the initial position coordinates of the connection with the connection rule template, the bonding connection is automatically drawn, and the standardized reconstruction of the encapsulation graphic is completed.

[0014] Optionally, the affine transformation matrix is ​​used to achieve image rotation and mirror correction.

[0015] A chip packaging drawing intelligent recognition and standardization conversion system is used to implement the chip packaging drawing intelligent recognition and standardization conversion method as described in any one of claims 1-7, characterized in that the system includes: a drawing preprocessing module, a classification and template matching module, a text recognition and graphic reconstruction module, and a result output module; The drawing preprocessing module is used to convert the packaged drawing files submitted by the customer into high-resolution bitmap images page by page, which serve as the original input images for subsequent image and text recognition. The classification and template matching module is used to identify enterprise logos in the original input image based on the logo recognition model to complete the first level of customer category classification; and to identify the encapsulation graphic recognition results in the original input image based on the encapsulation recognition model to complete the second level of encapsulation type subdivision; based on the two-level classification results, it retrieves the matching OCR coordinate template, encapsulation standard template and connection rule template from the hierarchical template library. The text recognition and image reconstruction module is used to execute the text standardization recognition process and the encapsulated image standardization reconstruction process in parallel using OCR coordinate templates and connection rule templates, to obtain standardized parameter data and standardized encapsulated images respectively. The output module is used to integrate standardized parameter data and standardized packaging images into a packaging standard template to generate the final standardized chip packaging drawing.

[0016] The beneficial effects of the technical solution provided in this application are: The overall design incorporates two synergistic core innovation directions. The first is an OCR intelligent recognition method based on template classification and scoring mechanisms; the second is a packaging structure and wafer interconnect standardization method based on YOLO multi-model fusion. Together, these two directions form a complete technology chain from drawing input to standardized output, effectively solving the diversity and heterogeneity problems that traditional methods struggle to handle. This invention proposes an intelligent text recognition method based on template classification and scoring mechanisms. First, customer drawings are classified in two levels: the first level categorizes companies based on their logos or specific text fields; the second level further subdivides based on packaging type or product series. Through this two-level classification, the system can dynamically match corresponding OCR recognition templates and coordinate rules. Based on this, key information areas in the drawings are automatically marked and cropped, and OCR recognition is performed only on the target areas, avoiding noise interference from full-image recognition. After recognition, this invention introduces a multi-candidate label scoring mechanism: for the same key field, the system performs OCR recognition based on multiple preset coordinate groups, and scores each group's results according to rules such as format matching degree, character confidence, and historical consistency. The result with the highest score is then selected as the output. This mechanism effectively solves the problem of OCR misreading caused by changes in drawing layout.

[0017] The packaging image structure analysis and reconstruction method based on multi-model collaboration is used to obtain packaging images by training a YOLO model to distinguish twenty packaging types for the image information in the drawings. At the same time, the wafer image in the packaging image is extracted and moved to the corresponding company standard template. Finally, the connection processing is performed. In the connection processing, the YOLO connection model, position model and thinning model of the original image are used. At the same time, the standardized packaging image is restored according to the secondary classification results in (1).

[0018] To address the challenge of recognizing and reconstructing graphic information in drawings, this invention designs a fusion recognition scheme based on the YOLO deep learning model. This scheme encompasses multiple stages, including package identification, wafer extraction, line recognition, and orientation correction. The system trains six dedicated YOLO models. These models work collaboratively, first identifying the initial coordinates of the connection between the package outline and the extracted wafer region. Then, based on the package orientation relationship model, they determine the geometric transformation between the customer's drawing and the company's standard template. Finally, they automatically complete the connection and alignment of the wafer drawing to the standard template. This innovation overcomes the limitations of traditional single models in handling complex package structures, achieving fully automated conversion from non-standardized customer drawings to standardized templates within the factory. It is particularly suitable for complex scenarios such as multi-chip and irregularly shaped wafers.

[0019] In summary, the innovation of this invention lies not only in introducing deep learning and OCR technology into encapsulated drawing processing, but also in systematically solving the problem of automatic conversion from non-standardized drawings to standardized templates through template classification, multi-candidate scoring, and multi-model fusion. Compared with existing technologies, this invention has significant advantages in recognition accuracy, cross-version robustness, and scene adaptability. Attached Figure Description

[0020] The present application will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a flowchart of an embodiment of this application; Figure 2 This is a diagram showing the connection results of package type I in the embodiments of this application; Figure 3 This is a diagram showing the wiring results of package type II in the embodiments of this application; Figure 4 This is a diagram showing the wiring results of package type III in the embodiments of this application; Figure 5 This is a diagram showing the wiring results of package type IV in the embodiments of this application; Figure 6 This is an example diagram of incorrect wiring in the embodiments of this application; Figure 7 This is the first result diagram of the layered training of the encapsulation type in the embodiments of this application; Figure 8 This is a second result diagram of the layered training of the encapsulation type in the embodiments of this application; Figure 9 This is a rotation result diagram from an embodiment of this application. Detailed Implementation

[0021] To provide a clearer understanding of the technical features, objectives, and effects of this application, the specific embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0022] The embodiments of this application provide a method for intelligent recognition and standardized conversion of chip packaging drawings.

[0023] Please refer to Figure 1 , Figure 1 This is a flowchart of a method for intelligent recognition and standardized conversion of chip packaging drawings according to an embodiment of this application, including: The method includes a parallel text and image processing flow, which comprises: a standardized text recognition flow for drawings and a standardized reconstruction flow for packaged graphics. The method is implemented based on multiple pre-trained YOLOv8 deep learning models, a hierarchical template library, and an OCR scoring and filtering model. The YOLOv8 deep learning models include: a marker recognition model, a package recognition model, a die-bonding detection model, a line recognition model, a positional relationship model, and a refined classification model. The hierarchical template library stores OCR coordinate templates, package standard templates, and connection rule templates corresponding to different customers and different package types. The method specifically includes the following steps: S1: Convert the packaging drawing files submitted by the customer into high-resolution bitmap images page by page, which will be used as the original input images for subsequent image and text recognition; S2: Based on the logo recognition model, identify the enterprise logo in the original input image to complete the first level of customer category classification; based on the encapsulation recognition model, identify the encapsulation graphic recognition result in the original input image to complete the second level of encapsulation type subdivision; based on the two-level classification results, retrieve the matching OCR coordinate template, encapsulation standard template and connection rule template from the hierarchical template library; S3: Using the OCR coordinate template and the connection rule template, the text standardization recognition process and the encapsulated graphic standardization reconstruction process are executed in parallel to obtain standardized parameter data and standardized encapsulated images, respectively. S4: Integrate standardized parameter data and standardized packaging images into a unified packaging standard template to generate the final standardized chip packaging drawing.

[0024] By adopting the above-mentioned technical solution, this application can be widely used in the automatic parsing and processing of diverse customer drawings in smart manufacturing factories. It can standardize the conversion of complex packaging drawings in various formats and styles, significantly shorten the review and circulation cycle of production drawings, and improve the factory's automated production efficiency and product yield consistency.

[0025] As one embodiment, the technical solution of this application can be adapted to single-chip packaging, dual-chip packaging and irregular wafer scenarios. For low-quality drawings with blurry drawings and faint lines, the recognition tolerance is improved by relying on the multi-candidate ROI scoring mechanism and multi-model collaborative detection. For distorted drawings rotated by 90°, 180°, 270° and mirror flipped, the geometric distortion correction success rate is 100%.

[0026] The YOLOv8 deep learning models are all trained using transfer learning, and the training dataset consists of historical labeled and encapsulated factory drawings. Data augmentation strategies are employed during model training, including random image flipping, adaptive brightness adjustment, and Gaussian noise superposition.

[0027] This application provides an embodiment as follows, with six specific dedicated models: 1) Package identification model: The input is a complete customer drawing image, and the output is the package category and an independent package image, used to distinguish twenty types of chip package structures; 2) Die bonding inspection model: The input is an independent package image, and the output is a wafer trimming image and wafer center coordinates; 3) Line recognition model: input is a wafer image, output is the initial position coordinates of the bonding lines; 4) Identification model: The input is a complete customer drawing image, and the output is the company name, used to achieve the first level of customer classification; 5) Positional relationship model: The input is an independently encapsulated image, and the output is a rotation marker, a mirror marker, and an affine transformation matrix; 6) Refine the classification model. The input is an independent encapsulated image, and the output is an encapsulated subtype label, which is used to match the subdivision template parameters.

[0028] This application provides an embodiment as follows: This system designs and trains six dedicated deep learning models, all based on the YOLOv8 architecture, employing a transfer learning strategy, and trained using a dataset of historically accumulated factory-annotated drawings. The functions and input / output of each model are described in the table below: Table 1 Introduction to each training model

[0029] During model training, at least 2000 labeled images were collected for each model class. Data augmentation techniques such as random flipping, brightness variation, and Gaussian noise were employed. The final mean accuracy (mAP@0.5) on the validation set was no less than 0.95. The overall logical flow of the entire method is as follows: Figure 1 As shown.

[0030] The specific results of the two-level classification are as follows: First-level classification: By identifying the company logo and specific text fields in the drawings using the logo recognition model, the customer's category is determined; Second-level classification: Based on the output of the packaging identification model, and combined with the product series, the packaging sub-types are divided. By linking the two-level classification results to call the corresponding OCR coordinate template and encapsulated standard template, adaptive template matching for multiple clients and encapsulation types can be achieved.

[0031] The standardized text recognition process for the drawings includes the following steps: Target Area Cropping: Based on the retrieved OCR coordinate template, locate and crop multiple candidate Regions of Interest (ROIs) corresponding to the key parameter fields of the drawing. Partitioned OCR recognition: Optical character recognition is performed independently for each cropped candidate region of interest (ROI), resulting in multiple sets of candidate recognition results; Multi-dimensional comprehensive scoring: Preset scoring indicators are used to score the candidate recognition results of each group in turn to obtain a comprehensive score; the scoring indicators include: format matching degree, character confidence degree, regional clarity and historical data consistency; Results screening and post-processing: The candidate recognition results with the highest comprehensive score are selected as the final text recognition results. The recognized text of the candidate recognition results is processed by unit conversion and numerical rounding to obtain standardized parameter data. Parameter entry: Fill the standardized parameter data into the data table corresponding to the packaging standard template.

[0032] The overall score is calculated using a weighted scoring formula.

[0033] The standardized reconstruction process for the encapsulated graphics includes the following steps: Encapsulation type recognition: Call the encapsulation recognition model to detect the original input image and output the encapsulation category and the corresponding independent encapsulation image; Wafer region extraction: Based on the die bonding detection model, the die bonding region is located within the independent package image, cropped to obtain the wafer image, and the wafer center coordinates are recorded; Bonding coordinate detection: The line recognition model is invoked to extract the initial position coordinates of the bonding lines on the wafer image; Geometric distortion correction: The rotation angle and mirror state of the independent package image relative to the package standard template are identified by the positional relationship model, an affine transformation matrix is ​​generated, and the wafer image is corrected by the affine transformation matrix to make the wafer image orientation consistent with the package standard template. Subtype refinement differentiation: Input the encapsulation category and independent encapsulation image into the refinement classification model, identify the subtype label of the current encapsulation, and load the refinement template parameters of the corresponding subtype; Graphic splicing and connection reconstruction: The corrected wafer image is superimposed onto the corresponding area of ​​the package standard template according to the coordinate position defined by the refined template parameters; By combining the initial position coordinates of the connection with the connection rule template, the bonding connection is automatically drawn, and the standardized reconstruction of the encapsulation graphic is completed.

[0034] The affine transformation matrix is ​​used to achieve image rotation and mirror correction.

[0035] A chip packaging drawing intelligent recognition and standardized conversion system, the system comprising: a drawing preprocessing module, a classification and template matching module, a text recognition and graphic reconstruction module, and a result output module; The drawing preprocessing module is used to convert the packaged drawing files submitted by the customer into high-resolution bitmap images page by page, which serve as the original input images for subsequent image and text recognition. The classification and template matching module is used to identify enterprise logos in the original input image based on the logo recognition model to complete the first level of customer category classification; and to identify the encapsulation graphic recognition results in the original input image based on the encapsulation recognition model to complete the second level of encapsulation type subdivision; based on the two-level classification results, it retrieves the matching OCR coordinate template, encapsulation standard template and connection rule template from the hierarchical template library. The text recognition and image reconstruction module is used to execute the text standardization recognition process and the encapsulated image standardization reconstruction process in parallel using OCR coordinate templates and connection rule templates, to obtain standardized parameter data and standardized encapsulated images respectively. The output module is used to integrate standardized parameter data and standardized packaging images into a packaging standard template to generate the final standardized chip packaging drawing.

[0036] In one embodiment, regarding the image recognition process, the system performs standardized conversion in the following order: First, the PDF file submitted by the customer is converted page by page into a high-resolution bitmap, and the logo.pt model (logo recognition model) is used to locate the company logo area. Simultaneously, the company name is obtained by recognizing key text using OCR, thereby loading the corresponding OCR coordinate configuration file from the template library. Subsequently, the package recognition model is used to identify the package category, and within the package image area, the die bonding detection model is used to crop the wafer region image.

[0037] Based on this, the line recognition model is called to obtain the starting coordinates of the connection, the positional relationship model is called to determine the directional difference between the customer's package diagram and the standard template and outputs the affine transformation matrix to correct the wafer image to be consistent with the direction of the standard template.

[0038] For a specific company, the model further refines the molecular type and loads the corresponding refinement template parameters. Finally, the corrected wafer image is overlaid onto the corresponding area of ​​the company's standard template according to the coordinates defined in the refinement template. Based on the connection rules set in the refinement template and the starting points detected by the line recognition model, connections are automatically drawn to generate a complete standardized package image. The entire process achieves fully automated conversion from original drawings to standardized images.

[0039] In terms of text recognition, the system first retrieves the index of key information pages from the template library based on the company name, and performs OCR recognition only on those pages. For each target field, the system reads multiple preset candidate ROI coordinate groups from the template, performs OCR recognition on each, and then performs a comprehensive score based on indicators such as format matching degree, character confidence degree, historical consistency, and ROI area clarity, selecting the candidate group with the highest total score as the recognition result. After recognition, the system performs post-processing on the original text according to preset calculation rules, such as unit conversion and numerical rounding, to obtain standardized data, which is automatically filled into the corresponding data table of the company's standard template. Through the parallel and collaborative processing of the above image and text processes, this invention achieves comprehensive standardized conversion of customer drawings.

[0040] In another embodiment, the present invention also employed multiple testing methods in the image recognition performance test. Some encapsulation test results are shown below. Figure 2 , Figure 3 , Figure 4 and Figure 5As shown, the test results for some package types are listed. Through a full sample traversal test, 338 valid package drawing files accumulated from the factory's history were fully identified. The overall test included 236 single-chip packages and 102 dual-chip packages. Based on manual verification results, 22 errors were found in single-chip packages (accuracy rate 90.7%) and 11 errors were found in dual-chip packages (accuracy rate 89.2%). Errors mainly occurred in extreme cases such as blurry drawings and light-colored wiring, such as... Figure 6 As shown, the errors are: too blurry, internal connections, and source file errors.

[0041] In one exemplary embodiment, the next step is a layered testing based on encapsulation type, calculating the recognition accuracy separately for each encapsulation type. The model training results are shown in the figure below. Figure 7 and Figure 8 As shown. The model training results show that the recognition accuracy for each encapsulation is above 99.5%, indicating high accuracy. The third test is the rotation and mirror invariance test. The original drawing was manually rotated by 90°, 180°, and 270°, and then mirrored to test the correction capability of this method. Figure 9 As shown in the figure. The results show that after processing with the orientation relationship model, the wafer images can be correctly restored to the standard template orientation, with an orientation correction success rate of 100%, and the wiring is restored without error.

[0042] This application provides an embodiment as follows: In the text recognition performance test, this invention employs a cross-version robustness test, selecting five different versions of drawings from the same customer to test the consistency of key field recognition. The results show that the coefficient of variation of the recognition results for core fields such as "wafer diameter" and "number of pads" across different versions is less than 2%, indicating good version robustness. Through the above tests, it can be concluded that the text recognition method of this invention has reached high standards for industrial applications in terms of accuracy, robustness, and anti-interference ability.

[0043] The above are merely exemplary embodiments of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure.

[0044] This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

Claims

1. A method for intelligent recognition and standardized conversion of chip packaging drawings, characterized in that, The method includes a parallel text and image processing flow, which comprises: a standardized text recognition flow for drawings and a standardized reconstruction flow for packaged graphics. The method is implemented based on multiple pre-trained YOLOv8 deep learning models, a hierarchical template library, and an OCR scoring and filtering model. The YOLOv8 deep learning models include: a marker recognition model, a package recognition model, a die-bonding detection model, a line recognition model, a positional relationship model, and a refined classification model. The hierarchical template library stores OCR coordinate templates, package standard templates, and connection rule templates corresponding to different customers and different package types. The method specifically includes the following steps: S1: Convert the packaging drawing files submitted by the customer into high-resolution bitmap images page by page, which will be used as the original input images for subsequent image and text recognition; S2: Based on the logo recognition model, identify the enterprise logo in the original input image to complete the first level of customer category classification; based on the encapsulation recognition model, identify the encapsulation graphic recognition result in the original input image to complete the second level of encapsulation type subdivision; based on the two-level classification results, retrieve the matching OCR coordinate template, encapsulation standard template and connection rule template from the hierarchical template library; S3: Using the OCR coordinate template and the connection rule template, the text standardization recognition process and the encapsulated graphic standardization reconstruction process are executed in parallel to obtain standardized parameter data and standardized encapsulated images, respectively. S4: Integrate standardized parameter data and standardized packaging images into a unified packaging standard template to generate the final standardized chip packaging drawing.

2. The method for intelligent recognition and standardized conversion of chip packaging drawings as described in claim 1, characterized in that, The YOLOv8 deep learning models are all trained using transfer learning, and the training dataset consists of historical labeled and encapsulated factory drawings. Data augmentation strategies are employed during model training, including random image flipping, adaptive brightness adjustment, and Gaussian noise superposition.

3. The method for intelligent recognition and standardized conversion of chip packaging drawings as described in claim 1, characterized in that, The specific results of the two-level classification are as follows: First-level classification: By identifying the company logo and specific text fields in the drawings using the logo recognition model, the customer's category is determined; Second-level classification: Based on the output of the packaging identification model, and combined with the product series, the packaging sub-types are divided. By linking the two-level classification results to call the corresponding OCR coordinate template and encapsulated standard template, adaptive template matching for multiple clients and encapsulation types can be achieved.

4. The method for intelligent recognition and standardized conversion of chip packaging drawings as described in claim 1, characterized in that, The standardized text recognition process for the drawings includes the following steps: Target Area Cropping: Based on the retrieved OCR coordinate template, locate and crop multiple candidate Regions of Interest (ROIs) corresponding to the key parameter fields of the drawing. Partitioned OCR recognition: Optical character recognition is performed independently for each cropped candidate region of interest (ROI), resulting in multiple sets of candidate recognition results; Multi-dimensional comprehensive scoring: Preset scoring indicators are used to score the candidate recognition results of each group in turn to obtain a comprehensive score; the scoring indicators include: format matching degree, character confidence degree, regional clarity and historical data consistency; Results screening and post-processing: The candidate recognition results with the highest comprehensive score are selected as the final text recognition results. The recognized text of the candidate recognition results is processed by unit conversion and numerical rounding to obtain standardized parameter data. Parameter entry: Fill the standardized parameter data into the data table corresponding to the packaging standard template.

5. The method for intelligent recognition and standardized conversion of chip packaging drawings as described in claim 4, characterized in that, The overall score is calculated using a weighted scoring formula.

6. The method for intelligent recognition and standardized conversion of chip packaging drawings as described in claim 1, characterized in that, The standardized reconstruction process for the encapsulated graphics includes the following steps: Encapsulation type recognition: Call the encapsulation recognition model to detect the original input image and output the encapsulation category and the corresponding independent encapsulation image; Wafer region extraction: Based on the die bonding detection model, the die bonding region is located within the independent package image, cropped to obtain the wafer image, and the wafer center coordinates are recorded; Bonding coordinate detection: The line recognition model is invoked to extract the initial position coordinates of the bonding lines on the wafer image; Geometric distortion correction: The rotation angle and mirror state of the independent package image relative to the package standard template are identified by the positional relationship model, an affine transformation matrix is ​​generated, and the wafer image is corrected by the affine transformation matrix to make the wafer image orientation consistent with the package standard template. Subtype refinement differentiation: Input the encapsulation category and independent encapsulation image into the refinement classification model, identify the subtype label of the current encapsulation, and load the refinement template parameters of the corresponding subtype; Graphic splicing and connection reconstruction: The corrected wafer image is superimposed onto the corresponding area of ​​the package standard template according to the coordinate position defined by the refined template parameters; By combining the initial position coordinates of the connection with the connection rule template, the bonding connection is automatically drawn, and the standardized reconstruction of the encapsulation graphic is completed.

7. The method for intelligent recognition and standardized conversion of chip packaging drawings as described in claim 6, characterized in that, The affine transformation matrix is ​​used to achieve image rotation and mirror correction.

8. A chip packaging drawing intelligent recognition and standardized conversion system, used to implement the chip packaging drawing intelligent recognition and standardized conversion method as described in any one of claims 1-7, characterized in that, The system includes: a drawing preprocessing module, a classification and template matching module, a text recognition and graphic reconstruction module, and a result output module; The drawing preprocessing module is used to convert the packaged drawing files submitted by the customer into high-resolution bitmap images page by page, which serve as the original input images for subsequent image and text recognition. The classification and template matching module is used to identify enterprise logos in the original input image based on the logo recognition model to complete the first level of customer category classification; and to identify the encapsulation graphic recognition results in the original input image based on the encapsulation recognition model to complete the second level of encapsulation type subdivision; based on the two-level classification results, it retrieves the matching OCR coordinate template, encapsulation standard template and connection rule template from the hierarchical template library. The text recognition and image reconstruction module is used to execute the text standardization recognition process and the encapsulated image standardization reconstruction process in parallel using OCR coordinate templates and connection rule templates, to obtain standardized parameter data and standardized encapsulated images respectively. The output module is used to integrate standardized parameter data and standardized packaging images into a packaging standard template to generate the final standardized chip packaging drawing.

Images