Character information recognition method in electronic component symbol diagram

Abstract

The present invention relates to the field of electronic design automation technology and discloses a method for recognizing character information in electronic component symbol diagrams. The method includes an electronic component symbol diagram text detection module that detects image areas containing text information, an electronic component symbol diagram text recognition module that identifies each character and obtains the text information therein, and an electronic component symbol diagram PIN name text semantic repair module that constructs a PIN name vocabulary and uses a spelling correction language model to perform spelling correction on the PIN name information and output the final recognition result. The electronic component symbol diagram character information recognition method of the present invention can quickly and accurately extract various character information from image-type electronic component symbol diagrams and circuit schematics, and accurately obtain electronic component symbol diagram models.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic design automation, and in particular to a method for recognizing character information in an electronic component symbol diagram. Background Art

[0002] Component symbol libraries are essential for schematic design and component selection tools in PCB EDA. Electrical design engineers want to quickly retrieve component symbol models from component datasheets (containing a single component symbol) or existing circuit schematics (containing multiple component symbols). These symbols primarily contain the pin numbers and names of each component pin.

[0003] There are already some semi-automatic symbol library building tools on the market, but it is still very difficult to recognize image-based component symbols. The reason is that the component symbol diagrams and circuit schematics in electronic component manuals contain a large number of combined symbols such as overlines, underlines, and slashes. There are also some rare special symbols. Although they are printed, the recognition accuracy is not high when using general OCR tools, which will result in a large number of missed detections and false detections. Summary of the Invention

[0004] The purpose of the present invention is to address the deficiencies of the above-mentioned technology and provide a method for recognizing character information in electronic component symbol diagrams, which can quickly and accurately extract various character information in picture-type electronic component symbol diagrams and circuit schematics, and accurately obtain electronic component symbol diagram models.

[0005] To achieve the above-mentioned purpose, the present invention relates to a method for recognizing character information in an electronic component symbol diagram, comprising an electronic component symbol diagram text detection module, an electronic component symbol diagram text recognition module, and an electronic component symbol diagram PIN Name text semantic repair module;

[0006] The electronic component symbol image text detection module uses the segmentation-based scene text detection network DBNet++ to detect image areas containing text information for electronic component symbol images and circuit schematics of the image type;

[0007] The electronic component symbol image text recognition module consists of an overline extraction algorithm, a character segmentation algorithm, and a single-character classification CNN network recognition algorithm. First, the overline extraction algorithm is used to extract the overlines in the character area. Then, the character segmentation algorithm is used to segment the characters in the image area obtained by the electronic component symbol image text detection module. Finally, the classification CNN network recognition algorithm is used to recognize each character and obtain the text information therein.

[0008] The electronic component symbol PIN Name semantic logic repair module includes PIN Name vocabulary construction and spelling correction language model. After the PIN Name vocabulary is constructed, the spelling correction language model is used to perform spelling correction on the PIN Name information obtained by the electronic component symbol text recognition module, and the final recognition result is output.

[0009] Preferably, when the electronic component symbol diagram text detection module is used, several electronic component symbol diagrams of picture type are selected from various electronic component manuals of chip manufacturers, and several circuit schematic diagrams provided by the chip manufacturers are selected at the same time. All text information in the electronic component symbol diagrams and circuit schematic diagrams are marked and sent to the pre-trained DBNet++ network for training to obtain a proprietary DBNet++ network, thereby realizing accurate detection of text information in the electronic component symbol diagrams and circuit schematic diagrams.

[0010] Preferably, the overline extraction algorithm uses horizontal projection and vertical projection histograms to extract overlines. First, the horizontal projection histogram is used to determine whether there is an overline in the character string. If there is an overline, the overline is cut, and then the cut part is vertically projected to accurately locate the character corresponding to the overline.

[0011] Preferably, the character segmentation algorithm is a connected component analysis method.

[0012] Preferably, the single-character classification CNN network recognition algorithm is a three-layer CNN network, which collects and organizes 52 uppercase and lowercase English characters, 10 numeric characters and several special characters, and a total of several single-character images in common fonts in electronic component manuals and circuit schematics to form a CNN training set of single-character images, and trains a single-character classification CNN network recognition algorithm in electronic component symbol diagrams.

[0013] Preferably, when building a PIN Name thesaurus, electronic component symbol information of various device types from chip manufacturers is collected and organized, and all PIN names appearing therein are statistically analyzed to build a PIN Name thesaurus. The frequency of each PIN Name is counted, and for the PIN Name obtained by the electronic component symbol image text recognition module, the confidence level of the PIN Name is set based on whether it appears in the PIN Name thesaurus and the corresponding frequency.

[0014] Preferably, the spelling correction language model collects and organizes the electronic component symbol information of various device types of chip manufacturers, constructs a sentence form, and each electronic component corresponds to a text, which contains the name, type, manufacturer and number of pins of the electronic component, and arranges all PIN Names according to the pin number in the text. Then, a data set is produced based on these text sets, and the pre-trained data set and labeled data are migrated and sent to the Soft Masked BERT network for training to obtain a spelling correction language model for performing semantic logic repair on the electronic component PIN Name obtained by the electronic component symbol diagram text recognition module.

[0015] Preferably, the overline extraction algorithm comprises the following steps:

[0016] Step S101: reading a character string image within the coordinate range of the text information detected by the DBNet++ network;

[0017] Step S102: performing binarization processing;

[0018] Step S103: drawing a horizontal projection histogram of the character string image and counting the number of pixels in the horizontal direction;

[0019] Step S104: determining whether there is an obvious pixel-free region in the horizontal projection histogram; if no such region exists, then there is no overline in the character string image; if such region exists, then there is an overline in the character string image;

[0020] Step S105: segmenting the character string image using the pixel-free area, and selecting the narrow part as the overline image;

[0021] Step S106: Perform vertical histogram statistics on the overlined portion of the image to obtain the coordinate range of each overline segment.

[0022] Preferably, the character segmentation algorithm and the single character classification CNN network recognition algorithm include the following steps:

[0023] Step S201: reading the character string image obtained after overline extraction;

[0024] Step S202: Performing a connected domain analysis on the image to obtain N connected domains, and dividing the character string image into N blocks of single-character images based on vertical midlines between adjacent connected domains;

[0025] Step S203: taking a single character image from left to right;

[0026] Step 204: If the aspect ratio of the single character image is greater than 1.5, there are overlapping characters, and the process proceeds to step S205; otherwise, the process proceeds to step S206;

[0027] Step S205: using the connected component method to segment the single character image into single character images without adhesion characters, and then using the single character classification CNN network recognition algorithm to identify the single character images respectively, and then proceeding to step S207;

[0028] Step S206: using a single character classification CNN network recognition algorithm to identify the single character image;

[0029] Step S207: insert the recognized character into the end of the previous recognition result string. If there is still an unrecognized single character image, go to step S203; otherwise, go to step S208.

[0030] Step S208: Match the recognition result character string with the result obtained by the overline extraction algorithm to obtain the final recognition result.

[0031] Preferably, the spelling correction language model training includes the following steps:

[0032] Step S301: Collect and organize symbol information of several electronic components, including device name, manufacturer name, device type, number of pins, and PIN Names of all pins;

[0033] Step S302: Construct the symbol information of the electronic component into an English sentence, which includes the device name, manufacturer, device type, and number of pins, and lists all the PIN names of the electronic component in order as Origin_text;

[0034] Step S303: construct a batch of Random_text based on PIN Name recognition error experience and random method;

[0035] Step S304: creating annotation data;

[0036] Step S305: The origin_text, random_text, and annotated data are fed into the Soft Masked BERT network for training. During fine-tuning, the learning rate is maintained at 2×10^-5, the number of hidden units in the bidirectional GRU is 256, and the batch size used by the model is 320. The spelling correction language model is used to perform semantic repair on the PIN Name obtained by the electronic component symbol and text recognition module.

[0037] Compared with the prior art, the present invention has the following advantages:

[0038] 1. It can quickly and accurately extract various character information from electronic component symbols and circuit schematics of picture type, and accurately obtain electronic component symbol models;

[0039] 2. Based on the electronic component symbol diagram scenario, the electronic component symbol diagram’s own DBNet++ network is trained to accurately detect the text in the electronic component symbol diagram;

[0040] 3. Based on the characteristics that the characters in the electronic component symbol diagram are printed and only contain English letters, numbers and some special characters, the CNN network is used, which has a simple and effective structure and high recognition accuracy;

[0041] 4. Based on the fact that there are a large number of overlined characters in the electronic component symbol diagram, the overline is detected using horizontal and vertical projection histograms, which can accurately locate the overline;

[0042] 5. A common electronic component PIN name database was constructed, and the confidence level of the PIN name obtained by the electronic component symbol and text recognition module was set accordingly;

[0043] 6. Train a Soft Masked BERT language model specifically for electronic component PIN names, so that it can perform overall semantic and logical repair on previously identified electronic component PIN names based on the characteristics of the electronic component type, and perform effective error correction. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A module diagram of a method for recognizing character information in an electronic component symbol diagram according to the present invention;

[0045] Figure 2 This is a flow chart of the overline extraction algorithm in the electronic component symbol text recognition module of the present invention;

[0046] Figure 3 This is a flow chart of the character segmentation algorithm and single character classification CNN network recognition algorithm in the present invention. DETAILED DESCRIPTION

[0047] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0048] like Figure 1 As shown, a method for recognizing character information in an electronic component symbol diagram includes an electronic component symbol diagram text detection module, an electronic component symbol diagram text recognition module, and an electronic component symbol diagram PIN Name text semantic repair module;

[0049] The electronic component symbol image text detection module uses the segmentation-based scene text detection network DBNet++ to detect image areas containing text information for electronic component symbol images and circuit schematics.

[0050] The electronic component symbol image text recognition module consists of an overline extraction algorithm, a character segmentation algorithm, and a single-character classification CNN network recognition algorithm. The overline extraction algorithm is first used to extract the overlines in the character area. The character segmentation algorithm is then used to segment the characters in the image area obtained by the electronic component symbol image text detection module. Finally, the classification CNN network recognition algorithm is used to identify each character and obtain the text information contained therein.

[0051] The electronic component symbol PIN name semantic logic repair module includes the construction of a PIN name vocabulary and a spelling correction language model. After the PIN name vocabulary is built, the spelling correction language model is used to perform spelling correction on the PIN name information obtained by the electronic component symbol text recognition module and output the final recognition result.

[0052] In this embodiment, when the electronic component symbol diagram text detection module is used, 1,000 electronic component symbol diagrams of picture type are selected from various electronic component manuals of common chip manufacturers around the world, and 100 circuit schematics provided by chip manufacturers are selected. All text information in the electronic component symbol diagrams and circuit schematics are annotated and sent to the pre-trained DBNet++ network for training to obtain a proprietary DBNet++ network, which can achieve accurate detection of text information in electronic component symbol diagrams and circuit schematics.

[0053] In this embodiment, the overline extraction algorithm uses horizontal projection and vertical projection histograms to extract overlines. The horizontal projection histogram is first used to determine whether there is an overline in the character string. If there is an overline, the overline is cut. Then, a vertical projection is performed on the cut part to accurately locate the character corresponding to the overline.

[0054] Specifically, if Figure 2 As shown, the overline extraction algorithm includes the following steps:

[0055] Step S101: reading a character string image within the coordinate range of the text information detected by the DBNet++ network;

[0056] Step S102: performing binarization processing;

[0057] Step S103: drawing a horizontal projection histogram of the character string image and counting the number of pixels in the horizontal direction;

[0058] Step S104: determining whether there is an obvious pixel-free region in the horizontal projection histogram; if no such region exists, then there is no overline in the character string image; if such region exists, then there is an overline in the character string image;

[0059] Step S105: segmenting the character string image using the pixel-free area, and selecting the narrow part as the overline image;

[0060] Step S106: Perform vertical histogram statistics on the overlined portion of the image to obtain the coordinate range of each overline segment.

[0061] In this embodiment, the character segmentation algorithm is the connected domain analysis method, and the single-character classification CNN network recognition algorithm is a three-layer CNN network. 52 uppercase and lowercase English characters, 10 numeric characters and 20 special characters are collected and sorted. Each single-character image contains two types: itself and rotated 180°. Each single character contains 1,000 pictures of commonly printed fonts in various electronic component manuals, totaling 164,000 pictures as a training set, constituting a CNN training set for single-character images. A three-layer CNN network is used, and a total of 70 rounds of training are performed to train a single-character classification CNN network recognition algorithm in electronic component symbol diagrams.

[0062] Specifically, in this embodiment, Figure 3 As shown, the character segmentation algorithm and the single character classification CNN network recognition algorithm include the following steps:

[0063] Step S201: reading the character string image obtained after overline extraction;

[0064] Step S202: Performing a connected domain analysis on the image to obtain N connected domains, and dividing the character string image into N blocks of single-character images based on vertical midlines between adjacent connected domains;

[0065] Step S203: taking a single character image from left to right;

[0066] Step 204: If the aspect ratio of the single character image is greater than 1.5, there are overlapping characters, and the process proceeds to step S205; otherwise, the process proceeds to step S206;

[0067] Step S205: using the connected component method to segment the single character image into single character images without adhesion characters, and then using the single character classification CNN network recognition algorithm to identify the single character images respectively, and then proceeding to step S207;

[0068] Step S206: using a single character classification CNN network recognition algorithm to identify the single character image;

[0069] Step S207: insert the recognized character into the end of the previous recognition result string. If there is still an unrecognized single character image, go to step S203; otherwise, go to step S208.

[0070] Step S208: Match the recognition result character string with the result obtained by the overline extraction algorithm to obtain the final recognition result.

[0071] In this embodiment, when building a PIN name vocabulary, 200,000 electronic component symbols of various device types from chip manufacturers were collected and organized. All PIN names that appeared in these symbols were statistically analyzed to build a PIN name vocabulary. The frequency of each PIN name was counted. For each PIN name obtained by the electronic component symbol image text recognition module, the confidence level of the PIN name was set based on whether it appeared in the PIN name vocabulary and its corresponding frequency.

[0072] The spelling correction language model collects and organizes 200,000 electronic component symbols of various device types from chip manufacturers, constructs them into sentences, and corresponds to a text for each electronic component. The text contains the name, type, manufacturer, and number of pins of the electronic component, and arranges all PIN names in the text by pin number. Then, a dataset is created based on these text collections. The pre-trained dataset and annotated data are migrated and sent to the Soft Masked BERT network for training. The resulting spelling correction language model is used to perform semantic logic repair on the electronic component PIN names obtained by the electronic component symbol image text recognition module.

[0073] Specifically, the spelling correction language model training includes the following steps:

[0074] Step S301: Collect and organize symbol information of 200,000 electronic components, including device name, manufacturer name, device type, number of pins, and PIN Name of all pins;

[0075] Step S302: Construct the symbol information of the electronic component into an English sentence, which includes the device name, manufacturer, device type, and number of pins, and lists all the pin names of the electronic component in order as Origin_text. For example, the ADC device ADS1240 of TI company corresponds to the following text sentence: "The ADS1240 is an ADC of TI company. It is composed of 24 pins. The names of the pins from 1 to 24 in anascending order are DVDD, DGND, XIN, XOUT, RESET, DSYNC, PDWN, DGND, VREF+, VREF–, AIN0 / D0, AIN1 / D1, AIN2 / D2, AIN3 / D3, AINCOM, AGND, AVDD, POL, CS, DIN, DOUT, SCLK, DRDY, BUFEN.";

[0076] Step S303: construct a batch of Random_texts based on the experience of PIN Name recognition errors and a random method. For example, the Random_text corresponding to the sample sentence in step S302 is as follows: "ADS1240 is an ADC of TI company. It is composed of 24 pins. The names of the pins from 1 to 24 in an ascending order are DVDD, DGND, XIN, X*, RESET, DSYNC, PDWN, DGND, VREF+, VREF*, AIN0 / D0, AIN1 / D1, AIN* / D2, AIN3 / D3, AINCOM, AGND, AVDD, POL, CS, DIN, DOUT, SCLK, DRDY, BUFEN.";

[0077] Step S304: Create annotation data. The annotation data corresponding to Origin_text in step S302 and Random_text in step S303 is “0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 1 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0”, where 0 represents that the corresponding PIN Name is correct and 1 represents that the corresponding PIN Name is incorrect.

[0078] Step S305: The origin_text, random_text, and annotated data are fed into the Soft Masked BERT network for training. During fine-tuning, the learning rate is maintained at 2×10^-5, the number of hidden units in the bidirectional GRU is 256, and the batch size used by the model is 320. The spelling correction language model is used to perform semantic repair on the PIN Name obtained by the electronic component symbol and text recognition module.

[0079] The present invention provides a method for recognizing character information in electronic component symbol diagrams, which can quickly and accurately extract various character information from image-type electronic component symbol diagrams and circuit schematics, and precisely obtain electronic component symbol diagram models. Based on the electronic component symbol diagram scenario, the method trains the electronic component symbol diagram's own DBNet++ network, enabling accurate detection of text within the electronic component symbol diagram. Taking into account the characteristics of printed characters in electronic component symbol diagrams, which consist of only English letters, numbers, and some special characters, a CNN network is employed, which has a simple and effective structure and high recognition accuracy. Taking into account the presence of a large number of overlined characters in electronic component symbol diagrams, the method uses horizontal and vertical projection histograms to detect overlines, enabling accurate positioning of overlines. A common electronic component PIN name vocabulary is constructed, and confidence levels are set for PIN names obtained by the electronic component symbol diagram text recognition module. A dedicated SoftMasked BERT language model for electronic component PIN names is trained, enabling it to perform overall semantic and logical repair of previously recognized electronic component PIN names based on the characteristics of the electronic component type, effectively correcting errors.

[0080] At the same time, it should be noted that the description of the above technical solutions is exemplary, and this specification can be embodied in different forms and should not be construed as being limited to the technical solutions set forth herein. On the contrary, providing these descriptions will make the disclosure of the present invention thorough and complete, and will fully convey the scope disclosed in this specification to those skilled in the art. In addition, the technical solutions of the present invention are limited only by the scope of the claims. The features of the various embodiments of the present invention may be combined or spliced with each other in part or in whole, and may be performed in various different ways as those skilled in the art will fully understand. The embodiments of the present invention may be performed independently of each other, or may be performed together in a mutually dependent relationship.

[0081] For ordinary technicians in the technical field to which the present invention belongs, they can make several simple deductions or substitutions without departing from the concept of the present invention, which should be regarded as falling within the scope of protection of the present invention.

Claims

1. A method for recognizing character information in an electronic component symbol diagram, characterized in that: It includes electronic component symbol image text detection module, electronic component symbol image text recognition module and electronic component symbol image PIN Name text semantic repair module; The electronic component symbol diagram text detection module uses a segmentation-based scene text detection network DBNet++ to detect image areas containing text information for electronic component symbol diagrams and circuit schematic diagrams of the picture type. When the electronic component symbol diagram text detection module is used, several electronic component symbol diagrams of the picture type are selected from various electronic component manuals of chip manufacturers, and several circuit schematic diagrams provided by chip manufacturers are selected. All text information in the electronic component symbol diagrams and circuit schematic diagrams is annotated and sent to a pre-trained DBNet++ network for training to obtain a proprietary DBNet++ network, thereby achieving accurate detection of text information in electronic component symbol diagrams and circuit schematic diagrams; The electronic component symbol image text recognition module is composed of an overline extraction algorithm, a character segmentation algorithm, and a single-character classification CNN network recognition algorithm. The overline extraction algorithm is first used to extract overlines in the character area. The overline extraction algorithm uses horizontal projection and vertical projection histograms to extract overlines. The horizontal projection histogram is first used to determine whether there are overlines in the character string. If there are overlines, the overlines are cut. The cut part is then vertically projected to accurately locate the character corresponding to the overline. The character segmentation algorithm is then used to segment the characters in the image area obtained by the electronic component symbol image text detection module. Finally, the classification CNN network recognition algorithm is used to identify each character to obtain the text information therein. The electronic component symbol PIN Name semantic logic repair module includes PIN Name vocabulary construction and spelling correction language model. After the PIN Name vocabulary is constructed, the spelling correction language model is used to perform spelling correction on the PIN Name information obtained by the electronic component symbol text recognition module, and the final recognition result is output.

2. The method for recognizing character information in an electronic component symbol diagram according to claim 1, wherein: The character segmentation algorithm is a connected component analysis method.

3. The method for recognizing character information in an electronic component symbol diagram according to claim 1, wherein: The single-character classification CNN network recognition algorithm is a three-layer CNN network. It collects and organizes 52 uppercase and lowercase English characters, 10 numeric characters and several special characters. A total of several single-character images in common fonts in electronic component manuals and circuit schematics constitute a CNN training set of single-character images, and trains a single-character classification CNN network recognition algorithm for electronic component symbol images.

4. The method for recognizing character information in an electronic component symbol diagram according to claim 1, wherein: When building the PIN Name lexicon, we collect and organize electronic component symbol information of various device types from chip manufacturers, statistically analyze all the PIN Names that appear in it, build a PIN Name lexicon, and count the word frequency of each PIN Name. For the PIN Name obtained by the electronic component symbol image text recognition module, we set the confidence level of the PIN Name based on whether it appears in the PIN Name lexicon and its corresponding word frequency.

5. The method for recognizing character information in an electronic component symbol diagram according to claim 1, wherein: The spelling correction language model collects and organizes electronic component symbol information of various device types from chip manufacturers, constructs sentences, and corresponds to a text for each electronic component. The text contains the name, type, manufacturer, and number of pins of the electronic component, and arranges all PIN names in the text by pin number. Then, a dataset is created based on these text collections. The pre-trained dataset and annotated data are migrated and fed into the Soft Masked BERT network for training. The resulting spelling correction language model is used to perform semantic logic repair on the electronic component PIN names obtained by the electronic component symbol graph text recognition module.

6. The method for recognizing character information in an electronic component symbol diagram according to claim 1, wherein: The overline extraction algorithm includes the following steps: Step S101: reading a character string image within the coordinate range of the text information detected by the DBNet++ network; Step S102: performing binarization processing; Step S103: drawing a horizontal projection histogram of the character string image and counting the number of pixels in the horizontal direction; Step S104: determining whether there is an obvious pixel-free region in the horizontal projection histogram; if no such region exists, then there is no overline in the character string image; if such region exists, then there is an overline in the character string image; Step S105: segmenting the character string image using the pixel-free area, and selecting the narrow part as the overline image; Step S106: Perform vertical histogram statistics on the overlined portion of the image to obtain the coordinate range of each overline segment.

7. The method for recognizing character information in an electronic component symbol diagram according to claim 1, wherein: The character segmentation algorithm and the single character classification CNN network recognition algorithm include the following steps: Step S201: reading the character string image obtained after overline extraction; Step S202: Performing a connected domain analysis on the image to obtain N connected domains, and dividing the character string image into N blocks of single-character images based on vertical midlines between adjacent connected domains; Step S203: taking a single character image from left to right; Step 204: If the aspect ratio of the single character image is greater than 1.5, there are overlapping characters, and the process proceeds to step S205; otherwise, the process proceeds to step S206; Step S205: using the connected component method to segment the single character image into single character images without adhesion characters, and then using the single character classification CNN network recognition algorithm to identify the single character images respectively, and then proceeding to step S207; Step S206: using a single character classification CNN network recognition algorithm to identify the single character image; Step S207: insert the recognized character into the end of the previous recognition result string. If there is still an unrecognized single character image, go to step S203; otherwise, go to step S208. Step S208: Match the recognition result character string with the result obtained by the overline extraction algorithm to obtain the final recognition result.

8. The method for recognizing character information in an electronic component symbol diagram according to claim 1, wherein: Spelling correction language model training includes the following steps: Step S301: Collect and organize symbol information of several electronic components, including device name, manufacturer name, device type, number of pins, and PIN Names of all pins; Step S302: Construct the symbol information of the electronic component into an English sentence, which includes the device name, manufacturer, device type, and number of pins, and lists all the PIN names of the electronic component in order as Origin_text; Step S303: construct a batch of Random_text based on PIN Name recognition error experience and random method; Step S304: creating annotation data; Step S305: Origin_text, Random_text and annotated data are fed into the Soft Masked BERT network for training. During fine-tuning, the learning rate is maintained at 2×10^-5, the number of hidden units in the bidirectional GRU is 256, and the batch size used by the model is 320. The spelling correction language model is used to perform semantic repair on the PINName obtained by the electronic component symbol and text recognition module.

Images