Method and device for identifying microelectronic assembly circuit characters

Abstract

The invention discloses a method and device for identifying microelectronic assembly circuit characters, and the method comprises the following steps: carrying out the data collection of a microelectronic assembly circuit tube shell, and processing the collected data, so as to obtain sample data; constructing an identification model comprising a text detection network and a text identification network; training the recognition model by using the sample data; and deploying the trained identification model in hardware equipment for identifying character contents on the microelectronic assembly circuit tube shell. According to the method, the accuracy and the speed of character recognition are improved, the overall efficiency and the intelligent level of microelectronic assembly circuit production are improved through application of automation and intelligent technologies, and the method is of great significance in promoting technical progress and industrial upgrading of the microelectronic industry.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular to a method and a device for recognizing characters of a microelectronic assembly circuit. Background Art

[0002] Microelectronics assembly technology is a technology that assembles microelectronic components through high-density interconnection, micro-welding, 3D tooling and other processes. It is widely used in aerospace, electronic information and other fields. Traditional microelectronics assembly circuits use QR codes as product identification, and scan the code to record product flow information in an assembly line manner. QR codes have the advantages of large information capacity and fast recognition speed, but in certain specific environments, such as metal or ceramic shell surfaces, the printing and recognition of QR codes may be subject to certain restrictions. Therefore, many production systems still use ink or laser marking to generate character batch numbers on the shell as product identification. However, this type of flow record usually needs to be manually filled in by humans in the production system, which is not only inefficient but also prone to errors. With the increasing number of assembled circuit types and outputs, intelligent identification of circuit shell batch numbers and real-time recording of the production status of each circuit have become crucial issues.

[0003] Optical character recognition is a technology that aims to automatically identify text in images. Traditional machine learning algorithms recognize characters based on manually extracted features, which mainly rely on the experience of researchers and lack stability. With the development of deep learning, researchers first applied CNN (convolutional neural network) and RNN (recurrent neural network) to end-to-end text detection and recognition. Today, mainstream OCR technology is divided into two stages: text detection and text recognition. Among them, DBNet and EAST methods are commonly used in text detection models. DBNet is currently the most efficient text detection network. CRNN and SRN methods are commonly used in text recognition. High-precision recognition is achieved through high-capacity models and complex reasoning, but due to its complex model structure, it is less efficient for real-time scene applications. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the technical problem to be solved by the present invention is to provide a method and device for identifying characters on a microelectronic assembly circuit, which can accurately and quickly identify the character batch number on the circuit tube shell.

[0005] In order to solve the above technical problems, a technical solution adopted by the present invention is: to provide a method for identifying characters of a microelectronic assembly circuit, comprising the following steps:

[0006] Collecting data on the microelectronic assembly circuit tube shell and processing the collected data to obtain sample data;

[0007] Build a recognition model including a text detection network and a text recognition network;

[0008] Using the sample data to train the recognition model;

[0009] The trained recognition model is deployed in hardware devices to identify the character content on the microelectronic assembly circuit shell.

[0010] Furthermore, the step of collecting data from the microelectronic assembly circuit shell and processing the collected data to obtain sample data includes the following sub-steps:

[0011] Collecting data of microelectronic assembly circuit tube shells, the data including batch number pictures of a plurality of microelectronic assembly circuit tube shells;

[0012] Preprocess the collected image data to obtain a data set;

[0013] Annotate the text area of ​​the image in the data set and generate corresponding text box coordinates;

[0014] Label the characters in the text box one by one to ensure that each character has an accurate label and obtain sample data.

[0015] Furthermore, the step of preprocessing the collected image data to obtain a data set includes the following sub-steps:

[0016] Unify the size of the collected images to ensure that all images have the same resolution;

[0017] Normalize the pixel values ​​of the images to eliminate the differences in pixel values ​​between different images;

[0018] Image filtering technology is used to remove noise from images, and the obtained image data constitutes a data set.

[0019] Furthermore, the text detection network includes a feature extraction module, a feature fusion module, a prediction module, a differentiable binarization module and a post-processing module, wherein:

[0020] The feature extraction module is used to extract a multi-scale feature map of the image;

[0021] The feature fusion module is used to sample feature maps of different scales to the same scale and cascade them to generate a fused feature map;

[0022] The prediction module is used to predict the probability map of the text area and the adaptive binarization threshold map based on the fused feature map;

[0023] The differentiable binarization module is used to calculate an approximate binary image according to the probability image and the binarization threshold image;

[0024] The post-processing module is used to generate a text box that can be used for subsequent character recognition based on the binary image and output a text area image.

[0025] Furthermore, when constructing a recognition model including a text detection network and a text recognition network, the SE module in the differentiable binarization module is deleted.

[0026] Furthermore, the text recognition network includes a convolution module, a first hybrid module, a second hybrid module, a third hybrid module and a fully connected layer, wherein:

[0027] The convolution module includes a plurality of convolution blocks, which are used to segment the text region image output by the text detection network into small text blocks, obtain character components and fusion features, and output a first feature map;

[0028] The first mixing module includes three local mixing modules, which are used to preliminarily extract local features of the first feature map and fuse them to form a second feature map;

[0029] The second mixing module includes three local mixing modules and three global mixing modules, which are used to deeply extract local features of the second feature map and span local areas, associate and integrate features at different positions, and output the first feature representation;

[0030] The third mixing module includes three global mixing modules, which are used to further extract and integrate global features according to the first feature representation and output a second feature representation;

[0031] The fully connected layer is used to convert the second feature representation output by the third mixing module into a final character recognition result.

[0032] Furthermore, the step of using the sample data to train the recognition model specifically includes the following sub-steps:

[0033] Set parameters, including training momentum, weight decay, learning rate, and batch size;

[0034] Divide the sample data into training set, validation set and test set according to the preset ratio;

[0035] Use the training set to train the text detection network in the recognition model and use the text region images output by the text detection network to train the text recognition network;

[0036] After each round of training, the recognition model is verified using the validation set, the performance indicators of the model are calculated, and the parameters are adjusted;

[0037] Repeat training and validation until the performance indicators on the validation set meet the preset requirements.

[0038] Furthermore, the performance indicators include accuracy and time, and the time refers to the time required for the recognition model to complete the prediction on the validation set;

[0039] The calculation formula of the accuracy is as follows:

[0040]

[0041] In formula (1), Accuracy represents the accuracy, TP represents the number of instances correctly predicted in true samples, TN represents the number of instances not correctly predicted in true samples, FP represents the number of instances in false samples that are incorrectly predicted as true samples, and FN represents the number of instances not correctly predicted in true samples.

[0042] Furthermore, the hardware device includes a data server, an application server, an AI algorithm server and a job recognition terminal, and the recognition model is deployed in the AI ​​algorithm server;

[0043] The step of deploying the trained recognition model in a hardware device for recognizing the character content on a microelectronic assembly circuit casing includes the following sub-steps:

[0044] The operation recognition terminal is used to collect image data of the microelectronic assembly circuit, transmit it to the AI ​​algorithm server and send a recognition request to the application server;

[0045] After receiving the recognition request, the application server calls the recognition model to process the image data, detect the character area of ​​the input image and recognize the character data;

[0046] The recognized character data is transmitted to the data server for storage.

[0047] In order to solve the above technical problems, another technical solution adopted by the present invention is to provide a device for identifying characters of a microelectronic assembly circuit, comprising:

[0048] A data processing module is used to collect data from the microelectronic assembly circuit tube shell and process the collected data to obtain sample data;

[0049] A model building module, used to build a recognition model including a text detection network and a text recognition network;

[0050] A model training module, used for training the recognition model using the sample data;

[0051] The character recognition module is used to deploy the trained recognition model in the hardware device for recognizing the character content on the microelectronic assembly circuit shell.

[0052] The method and device for identifying characters in a microelectronic assembly circuit of the present invention have at least the following beneficial effects: the present invention constructs an advanced recognition model including a text detection network and a text recognition network, and utilizes deep learning technology to significantly improve the accuracy of character recognition, with the circuit character recognition accuracy reaching 98.69%; when constructing the recognition model, the model is made lightweight by streamlining the model structure, which not only reduces the consumption of computing resources, but also makes the model easier to deploy and integrate in various hardware devices, thereby lowering the threshold and cost of technical application; by optimizing the model structure and algorithm flow, rapid character recognition is achieved, with a recognition time of 1.068 seconds per image, It greatly improves production efficiency, reduces manual intervention and waiting time, and enables the production line to run more smoothly and efficiently; the method and device of the present invention can identify multiple circuits at the same time, and have good adaptability to the size and color of the circuit tube shell, and can play an excellent role in the production of microelectronic assembly circuits of different specifications and types, thereby improving the versatility and flexibility of the technology; by deploying the trained recognition model in the hardware device, the automation of circuit character recognition is realized, avoiding the inefficiency and error-prone problems of manually filling in circulation records in the traditional method, improving the accuracy and traceability of the data, and providing strong support for the quality control and production management of microelectronic assembly circuits. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0054] Figure 1 The present invention is a flowchart of a method for identifying characters of a microelectronic assembly circuit according to an embodiment of the present invention.

[0055] Figure 2 for Figure 1 Flow chart of step S1 in FIG.

[0056] Figure 3 This is an example diagram of part of the data collected in one embodiment of the present invention.

[0057] Figure 4 Schematic diagram of the structure of a text detection network in one embodiment of the present invention.

[0058] Figure 5 This is a schematic diagram of the structure of the text recognition network before improvement.

[0059] Figure 6 Schematic diagram of the structure of an improved text detection network in one embodiment of the present invention.

[0060] Figure 7 for Figure 1Flow chart of step S3 in FIG.

[0061] Figure 8 It is a schematic diagram of the structure of hardware equipment in one embodiment of the present invention.

[0062] Fig. 9 for Figure 1 Flow chart of step S4 in FIG.

[0063] Fig.10 Schematic diagram of the process of character recognition by a recognition model in one embodiment of the present invention.

[0064] Fig.11 It is a structural block diagram of an embodiment of a device for identifying characters of a microelectronic assembly circuit according to the present invention. DETAILED DESCRIPTION

[0065] The present invention will be further described below in conjunction with the accompanying drawings.

[0066] See also Figure 1 , is a flow chart of an embodiment of a method for identifying characters in a microelectronic assembly circuit of the present invention. This embodiment specifically includes the following steps:

[0067] S1. Obtain sample data.

[0068] Specifically, data is collected from the microelectronic assembly circuit tube shell and the collected data is processed to obtain sample data. Figure 2 , this step S1 may include the following sub-steps:

[0069] S11. Data collection.

[0070] Specifically, data of microelectronic assembly circuit tube shells is collected, and the data includes batch number pictures of several microelectronic assembly circuit tube shells. Figure 3 The microelectronic assembly circuit uses a tube shell made of gold, nickel, or ceramic as a carrier, and its batch number is generated by laser or ink marking. In this embodiment, more than 6,000 batch number images of microelectronic assembly circuit tube shells were collected.

[0071] S12. Data preprocessing.

[0072] The collected image data is preprocessed to obtain a data set. Specifically, the collected images are resized to ensure that all images have the same resolution; the pixel values ​​of the images are normalized to eliminate the pixel value differences between different images; image filtering technology is used to remove noise in the images, and the obtained image data constitutes a data set.

[0073] S13. Mark the text area.

[0074] Specifically, the text area of ​​the image in the data set is annotated to generate corresponding text box coordinates, providing a basis for subsequent character recognition.

[0075] S14. Mark characters.

[0076] Specifically, the characters in the text box are labeled one by one to ensure that each character has an accurate label, and sample data is obtained. The sample data can be used to train the recognition model.

[0077] S2. Build a recognition model.

[0078] Specifically, a recognition model including a text detection network and a text recognition network is constructed. Figure 4 The text detection network includes a feature extraction module, a feature fusion module, a prediction module, a differentiable binarization module and a post-processing module, wherein: the feature extraction module is used to extract a multi-scale feature map of an image; the feature fusion module is used to sample feature maps of different scales to the same scale and cascade them to generate a fused feature map; the prediction module is used to predict a probability map and an adaptive binarization threshold map of a text area based on the fused feature map; the differentiable binarization module is used to calculate an approximate binary map based on the probability map and the binarization threshold map; the post-processing module is used to generate a text box that can be used for subsequent character recognition based on the binary map, and output a text area image.

[0079] The specific process of calculating the approximate binary image using the differentiable binarization module is as follows:

[0080]

[0081] in, represents the approximate step function of the binary image, T represents the network learning threshold image, and k represents the gradient gain factor. In the differentiable binarization process, the k gradient gain factor is used to adjust the gradient size during back propagation. By adjusting this factor, the sensitivity of the network to the loss function can be controlled, thereby affecting the training speed and effect. As the k value increases, the larger the back propagation gradient, the higher the accuracy of text area detection. i,j -T i,j Represents the DB function, which refers to the function used in the differentiable binarization process. It is usually an approximate step function that can be learned and optimized by the network during the training process.

[0082] In order to improve the network detection speed, in this embodiment, when constructing a recognition model including a text detection network and a text recognition network, the SE module in the differentiable binarization module is deleted. The SE module is based on the channel attention mechanism, and can establish dependencies between channels to improve the accuracy. However, in the case of high image pixels and large channel sizes, the SE module will result in a large computational time cost, affecting the network detection speed. The present invention deletes the SE module in the differentiable binarization module, sacrificing part of the accuracy in exchange for a faster detection speed.

[0083] See also Figure 5 and Figure 6 , the text recognition network includes a convolution module, a first mixing module, a second mixing module, a third mixing module and a fully connected layer, wherein: the convolution module includes a plurality of convolution blocks, which are used to divide the text area image output by the text detection network into small text blocks, obtain character components and fusion features, and output a first feature map; the first mixing module includes three local mixing modules, which are used to preliminarily extract local features of the first feature map and fuse them to form a second feature map; the second mixing module includes three local mixing modules and three global mixing modules, which are used to deeply extract local features of the second feature map and cross local areas, associate and integrate features at different positions, and output a first feature representation; the third mixing module includes three global mixing modules, which are used to further extract and integrate global features according to the first feature representation, and output a second feature representation; the fully connected layer is used to convert the second feature representation output by the third mixing module into a final character recognition result.

[0084] This implementation introduces a single-vision text recognition model based on Transformer decoding as a text recognition network, which can omit the sequence modeling process and effectively mine image text information. The global hybrid module can establish long-distance dependencies between character components and weaken the influence of non-text parts in the image. The local hybrid module encodes the character morphology and recognizes the correlation between different parts within a character and the character stroke structure.

[0085] Since the batch number of the circuit does not contain Chinese characters, has no complex stroke structure and has a single background during recognition, in order to further improve the recognition speed, this network reduces the global mixing modules in the first mixing module and the local mixing modules in the third mixing module of the original visual recognition network, reduces repeated merging, and reduces the number of local mixing modules in the first mixing module to three, the number of local mixing modules and global mixing modules in the second mixing module to three respectively, and the number of global mixing modules in the third mixing module to three.

[0086] S3. Use sample data to train the recognition model.

[0087] See also Figure 7 , this step S3 includes the following sub-steps:

[0088] S31. Set parameters, including training momentum, weight decay, learning rate, and batch size.

[0089] Specifically, reasonable parameters are set for the training process, which mainly include the training impulse value used to accelerate gradient descent and reduce oscillation, weight decay used to prevent model overfitting and reduce the weight size through regularization method, learning rate used to control the parameter update step size, and batch size indicating the number of samples used in each training.

[0090] S32, divide the set into training set, validation set and test set.

[0091] Specifically, the sample data is divided into a training set, a validation set, and a test set according to a preset ratio. In the present embodiment, the sample data is divided into a training set, a validation set, and a test set according to a ratio of 6:2:2. The training set is used to train the recognition model so that the model can learn the characteristics of the sample data. The validation set is used to verify the performance of the model after each round of training to help adjust the parameters and model structure to ensure that the model performs well on unseen data. The test set is used to finally evaluate the performance of the model to ensure the stability and accuracy of the model in practical applications.

[0092] S33. Training model.

[0093] Specifically, the training set is used to train the text detection network in the recognition model so that it can accurately detect the text area in the image, and the text area image output by the text detection network is used to train the text recognition network so that it can recognize the specific character content in the text area.

[0094] S34. Calculate performance indicators.

[0095] Specifically, after each round of training, the recognition model is verified using the validation set, the performance indicators of the model are calculated, and the parameters are adjusted. The performance indicators include accuracy and time, and the time refers to the time required for the recognition model to complete the prediction on the validation set.

[0096] The accuracy is obtained by the following formula:

[0097]

[0098] Among them, Accuracy represents the accuracy rate, TP represents the number of instances correctly predicted in true samples, TN represents the number of instances not correctly predicted in true samples, FP represents the number of instances in false samples that are incorrectly predicted as true samples, and FN represents the number of instances not correctly predicted in true samples.

[0099] S35. Repeat training and verification until the performance indicators meet the requirements.

[0100] Specifically, the parameters are adjusted according to the performance indicators (including accuracy and time) on the validation set, and training and validation are repeated until the performance indicators on the validation set meet the preset requirements (such as the accuracy reaches the set threshold, etc.). Then, the model can be considered to have been trained and can be put into practical application.

[0101] S4. Identify characters.

[0102] Specifically, the trained recognition model is deployed in a hardware device to recognize the character content on the microelectronic assembly circuit casing. Figure 8 The hardware device includes a data server, an application server, an AI algorithm server and a job recognition terminal, and the recognition model is deployed in the AI ​​algorithm server.

[0103] See also Fig. 9 , this step S4 includes the following sub-steps:

[0104] S41, collecting image data.

[0105] Specifically, the job identification terminal is used to collect image data of the microelectronic assembly circuit. The job identification terminal transmits the collected image data to the AI ​​algorithm server and sends a recognition request to the application server at the same time.

[0106] S42. Recognize character data.

[0107] Specifically, after receiving the recognition request, the application server calls the recognition model to process the received image data, detects the character area of ​​the input image and recognizes the character data in the character area. For an example, please refer to Fig.10 , which is the specific process of character recognition by the recognition model in this implementation manner.

[0108] S43. Store character data.

[0109] Specifically, the AI ​​algorithm server transmits the recognized character data to the data server for storage.

[0110] See also Fig.11 , is a structural block diagram of an embodiment of the device for identifying characters of a microelectronic assembly circuit of the present invention. The device for identifying characters of a microelectronic assembly circuit of this embodiment is used to implement the method for identifying characters of a microelectronic assembly circuit as described in the above embodiment. Specifically, the device for identifying characters of a microelectronic assembly circuit of this embodiment includes a data processing module 100, a model building module 200, a model training module 300 and a character recognition module 400. Among them:

[0111] The data processing module 100 is used to collect data from the microelectronic assembly circuit shell and process the collected data to obtain sample data.

[0112] The model building module 200 is used to build a recognition model including a text detection network and a text recognition network. The text detection network is used to detect text areas, and includes a feature extraction module, a feature fusion module, a prediction module, a differentiable binarization module, and a post-processing module. Among them: the feature extraction module is used to extract the multi-scale feature map of the image; the feature fusion module is used to sample the feature maps of different scales to the same scale and cascade them to generate a fused feature map; the prediction module is used to predict the probability map and the adaptive binarization threshold map of the text area based on the fused feature map; the differentiable binarization module is used to calculate an approximate binary map based on the probability map and the binarization threshold map; the post-processing module is used to generate a text box that can be used for subsequent character recognition based on the binary map, and output a text area image.

[0113] The text recognition network is used to recognize character data, including a convolution module, a first mixing module, a second mixing module, a third mixing module and a fully connected layer. Among them: the convolution module includes a number of convolution blocks, which are used to divide the text area image output by the text detection network into small text blocks, obtain character components and fusion features, and output the first feature map. The first mixing module includes three local mixing modules, which are used to preliminarily extract the local features of the first feature map and fuse them to form a second feature map. The second mixing module includes three local mixing modules and three global mixing modules, which are used to deeply extract the local features of the second feature map and cross the local area, associate and integrate the features at different positions, and output the first feature representation. The third mixing module includes the global mixing module, which is used to further extract and integrate the global features according to the first feature representation and output the second feature representation. The fully connected layer is used to convert the second feature representation output by the third mixing module into the final character recognition result.

[0114] The model training module 300 is used to train the recognition model using the sample data.

[0115] The character recognition module 400 is used to recognize the character content on the microelectronic assembly circuit shell, including hardware equipment and a recognition model deployed in the hardware equipment. The hardware equipment includes a data server, an application server, an AI algorithm server, and a job recognition terminal, and the recognition model is deployed in the AI ​​algorithm server.

[0116] The invention significantly improves the accuracy of character recognition by constructing an advanced recognition model including a text detection network and a text recognition network, and utilizes deep learning technology, and the accuracy of circuit character recognition is as high as 98.69%. When constructing the recognition model, the model is made lightweight by streamlining the model structure, which not only reduces the consumption of computing resources, but also makes the model easier to deploy and integrate in various hardware devices, thereby reducing the threshold and cost of technology application. By optimizing the model structure and algorithm flow, rapid character recognition is achieved, reaching a recognition time of 1.068 seconds per image, greatly improving production efficiency, reducing manual intervention and waiting time, and enabling the production line to run more smoothly and efficiently. The method and device of the invention can simultaneously recognize multiple circuits, and have good adaptability to the size and color of the circuit shell, and can play an excellent role in the production of microelectronic assembly circuits of different specifications and types, thereby improving the versatility and flexibility of the technology. By deploying the trained recognition model in the hardware device, the automation of circuit character recognition is achieved, the inefficiency and error-prone problems of manually filling in the circulation record in the traditional method are avoided, the accuracy and traceability of the data are improved, and strong support is provided for the quality control and production management of the microelectronic assembly circuit.

[0117] The above content only expresses the preferred embodiments of the present invention, and its description is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention patent. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be based on the attached claims.

Claims

1. A method for identifying characters of a microelectronic assembly circuit, characterized in that: The following steps are involved: Collecting data on the microelectronic assembly circuit tube shell and processing the collected data to obtain sample data; Build a recognition model including a text detection network and a text recognition network; Using the sample data to train the recognition model; The trained recognition model is deployed in hardware devices to identify the character content on the microelectronic assembly circuit shell.

2. The method for identifying characters of a microelectronic assembly circuit as claimed in claim 1, characterized in that: The step of collecting data from the microelectronic assembly circuit shell and processing the collected data to obtain sample data includes the following sub-steps: Collecting data of microelectronic assembly circuit tube shells, the data including batch number pictures of a plurality of microelectronic assembly circuit tube shells; Preprocess the collected image data to obtain a data set; Annotate the text area of ​​the image in the data set and generate corresponding text box coordinates; Label the characters in the text box one by one to ensure that each character has an accurate label and obtain sample data.

3. The method for identifying characters of a microelectronic assembly circuit as claimed in claim 2, characterized in that: The step of preprocessing the collected image data to obtain a data set includes the following sub-steps: Unify the size of the collected images to ensure that all images have the same resolution; Normalize the pixel values ​​of the images to eliminate the differences in pixel values ​​between different images; Image filtering technology is used to remove noise from images, and the obtained image data constitutes a data set.

4. The method for identifying characters of a microelectronic assembly circuit as claimed in claim 1, characterized in that: The text detection network includes a feature extraction module, a feature fusion module, a prediction module, a differentiable binarization module and a post-processing module, wherein: The feature extraction module is used to extract a multi-scale feature map of the image; The feature fusion module is used to sample feature maps of different scales to the same scale and cascade them to generate a fused feature map; The prediction module is used to predict the probability map of the text area and the adaptive binarization threshold map based on the fused feature map; The differentiable binarization module is used to calculate an approximate binary image according to the probability image and the binarization threshold image; The post-processing module is used to generate a text box that can be used for subsequent character recognition based on the binary image and output a text area image.

5. The method for identifying characters of a microelectronic assembly circuit as claimed in claim 4, characterized in that: When constructing a recognition model including a text detection network and a text recognition network, the SE module in the differentiable binarization module is deleted.

6. The method for identifying characters of a microelectronic assembly circuit as claimed in claim 4, characterized in that: The text recognition network includes a convolution module, a first hybrid module, a second hybrid module, a third hybrid module and a fully connected layer, wherein: The convolution module includes a plurality of convolution blocks, which are used to segment the text region image output by the text detection network into small text blocks, obtain character components and fusion features, and output a first feature map; The first mixing module includes three local mixing modules, which are used to preliminarily extract local features of the first feature map and fuse them to form a second feature map; The second mixing module includes three local mixing modules and three global mixing modules, which are used to deeply extract local features of the second feature map and span local areas, associate and integrate features at different positions, and output the first feature representation; The third mixing module includes three global mixing modules, which are used to further extract and integrate global features according to the first feature representation and output a second feature representation; The fully connected layer is used to convert the second feature representation output by the third mixing module into a final character recognition result.

7. The method for identifying characters of a microelectronic assembly circuit as claimed in claim 1, characterized in that: The step of using the sample data to train the recognition model specifically includes the following sub-steps: Set parameters, including training momentum, weight decay, learning rate, and batch size; Divide the sample data into training set, validation set and test set according to the preset ratio; Use the training set to train the text detection network in the recognition model and use the text region images output by the text detection network to train the text recognition network; After each round of training, the recognition model is verified using the validation set, the performance indicators of the model are calculated, and the parameters are adjusted; Repeat training and validation until the performance indicators on the validation set meet the preset requirements.

8. The method for identifying characters of a microelectronic assembly circuit as claimed in claim 7, characterized in that: The performance indicators include accuracy and time, where the time refers to the time required for the recognition model to complete the prediction on the validation set; The calculation formula of the accuracy is as follows: In formula (1), Accuracy represents the accuracy, TP represents the number of instances correctly predicted in true samples, TN represents the number of instances not correctly predicted in true samples, FP represents the number of instances in false samples that are incorrectly predicted as true samples, and FN represents the number of instances not correctly predicted in true samples.

9. The method for identifying characters of a microelectronic assembly circuit as claimed in claim 1, characterized in that: The hardware device includes a data server, an application server, an AI algorithm server and a job recognition terminal, and the recognition model is deployed in the AI ​​algorithm server; The step of deploying the trained recognition model in a hardware device for recognizing the character content on a microelectronic assembly circuit casing includes the following sub-steps: The operation recognition terminal is used to collect image data of the microelectronic assembly circuit, transmit it to the AI ​​algorithm server and send a recognition request to the application server; After receiving the recognition request, the application server calls the recognition model to process the image data, detect the character area of ​​the input image and recognize the character data; The recognized character data is transmitted to the data server for storage.

10. A device for identifying characters on a microelectronic assembly circuit, characterized in that: include: A data processing module is used to collect data from the microelectronic assembly circuit tube shell and process the collected data to obtain sample data; A model building module, used to build a recognition model including a text detection network and a text recognition network; A model training module, used for training the recognition model using the sample data; The character recognition module is used to deploy the trained recognition model in the hardware device for recognizing the character content on the microelectronic assembly circuit shell.

Images