Substrate carrier identification method, system, and storage medium

Abstract

Disclosed are a substrate carrier identification method, a system, and a storage medium. The method is used for identifying different objects inside a substrate carrier, the objects comprising a substrate, a separator and a cushion, located in a semiconductor substrate carrier. The method comprises: acquiring image data to be tested of the uppermost layer in a semiconductor substrate carrier; sending each piece of image data to be tested to a pre-trained classification neural network model for object identification; and the classification neural network model determining whether a no-load state is reached, and if so, ending the process; if not, then on the basis of an identification result of the classification neural network model, calling a corresponding manipulator to pick up the uppermost object. The method can quickly and accurately identify objects in a semiconductor substrate carrier by means of a neural network model, and call a corresponding manipulator.

Description

A substrate carrier identification method, system and storage medium TECHNICAL FIELD

[0001] The present application relates to the technical field of semiconductor transport equipment, and in particular to a substrate carrier identification method, system and storage medium. BACKGROUND

[0002] A semiconductor substrate carrier is a container for loading substrates and their protective inserts, mainly used in the packaging and transportation process of substrates to prevent the substrates from being scratched and damaged. The semiconductor substrate carrier includes a box body with an open end, the box body defines a placing cavity, and the open end of the box body is covered with an upper cover, and the substrates are stacked in the placing cavity. The bottom of the box body and the lower end of the upper cover are both placed with a buffer pad, which is usually made of soft material such as foam and other dust-free soft materials, used to absorb impact and vibration during transportation. At the same time, in order to prevent the adjacent substrates from contacting each other and causing wear or contamination, a spacer is placed between adjacent substrates. The spacer is used to form a physical barrier between the substrates to prevent direct contact between the substrates, thereby reducing the risk of scratching or other types of damage. Therefore, when the semiconductor substrate carrier carries the substrates, the placement from bottom to top is as follows: box body bottom plate-buffer pad-substrate-spacer-substrate-spacer-…-substrate-buffer pad-upper cover.

[0003] Because the materials or types of substrates, spacers and buffer pads are different, and the types of substrates are also different, if a corresponding mechanical hand is used to take and place them, the damage of the substrates may be caused by the error of the mechanical claws. How to provide a method for quickly identifying the internal objects of the semiconductor substrate carrier to quickly match the corresponding mechanical hand becomes a technical problem to be solved. SUMMARY

[0004] In order to overcome the above-mentioned shortcomings, the purpose of the present application is to provide a substrate carrier identification method, system and storage medium, which can quickly and accurately identify the objects in the semiconductor substrate carrier.

[0005] In order to achieve the above purpose, the technical solution adopted by the present application is: a substrate carrier identification method for identifying different objects in the substrate carrier, the objects including substrates, spacers and buffer pads in the semiconductor substrate carrier, the method comprising:

[0006] acquiring the uppermost layer of the to-be-tested image data in the semiconductor substrate carrier;

[0007] sending each of the to-be-tested image data to a pre-trained classification neural network model for object identification;

[0008] The classification neural network model judges whether the no-load state is reached, if yes, the method ends, if no, corresponding manipulator is called according to the recognition result of the classification neural network model to pick up the uppermost object.

[0009] The application has the advantages that: by pre-training and deploying a classification neural network model, the classification neural network model is used to directly process image data, the category of the object can be quickly judged to call corresponding manipulator. The neural network model has high recognition efficiency and fast response.

[0010] Further, the training of the classification neural network model comprises:

[0011] Physical property data of the object to be recognized is collected;

[0012] A plurality of original image data of each object and in a no-load state are collected to form and save original image data sets corresponding to each object and the no-load state;

[0013] Part or all of the original image data of the original image data set is input into an initial neural network model as a training set to train the initial neural network model;

[0014] The trained initial neural network model forms a classification neural network model, and the classification neural network model is stored and deployed.

[0015] The classification neural network model can identify the physical properties of the object and the no-load state, find the corresponding object or no-load state according to the identified physical properties, and then identify the object and judge the no-load state.

[0016] Further, the initial neural network model is a model that has been trained on a data set, the initial neural network model uses a ResNet neural network, and the data set is an ImageNet data set.

[0017] Since the initial neural network model has been trained on a large data set, the initial neural network model has a certain reliability, which is convenient for subsequent training. Avoiding the problem of starting from scratch to train a new model, which requires a large amount of data set, a long training time and expensive computing hardware.

[0018] Further, before the trained initial neural network model forms a classification neural network model, it further comprises:

[0019] A test set is collected, the test set is input into the trained initial neural network model for testing, and the initial neural network model that passes the test is a classification neural network model.

[0020] Only the initial neural network model that has passed the test can meet the use requirements and be fixed as the classification neural network model.

[0021] Further, the to-be-tested image data is stored in the training set and the training set is updated after the identification is completed. The updated training set can repeatedly retrain the initial neural network model to continuously improve the accuracy.

[0022] Further, before each of the to-be-tested image data is sent to the pre-trained classification neural network model for object identification, it further includes:

[0023] The to-be-tested image data is preprocessed, and the preprocessing is cutting the center region of the to-be-tested image data, and updating the to-be-tested image data to the preprocessed to-be-tested image data.

[0024] The preprocessing removes part of the to-be-tested image data. Through this step, the calculation amount of the classification neural network model image recognition processing is reduced, the recognition efficiency is improved, and the recognition efficiency is improved.

[0025] Further, the plurality of original image data of each of the objects and the empty state are collected to form and save the original image data set corresponding to each of the objects and the empty state.

[0026] The plurality of original image data of different object combination states are collected to form the original image data set corresponding to different object combination states.

[0027] The original image data of different object combinations, that is, the original image data of different substrates and different spacers, are collected to quickly identify the topmost substrate or spacer through early training, so as to improve the accuracy and speed of identification.

[0028] Further, before part or all of the original image data of the original image data set is input into the initial neural network model as a training set for training, it further includes:

[0029] The original image data is preprocessed, and the preprocessing is cutting the center region of the original image data, and updating the original image data to the preprocessed original image data.

[0030] Before the model training, a preprocessing step is added to cut part of the original image data, which reduces the image processing calculation amount of the classification initial neural network model and speeds up the image processing rate.

[0031] Further, the center region is a region with a preset diameter size, which is centered on the center of the original image data or the image data to be detected.

[0032] Further, the isolation member includes an isolation ring and an isolation member, and the preset diameter size is smaller than the diameter of the substrate and the inner diameter of the isolation ring, or the preset diameter size is smaller than the diameter of the substrate and the diameter of the isolation member.

[0033] The selection of the center region ensures that the combination of the object will not be affected after the center region is cut, and the identification will not be affected.

[0034] The application further discloses a substrate carrier recognition system, and the system comprises:

[0035] An imaging module is used to collect image data to be detected.

[0036] A database module is used to store original image data sets corresponding to different objects and empty load states, and one original image data set is saved in one folder.

[0037] A learning model processing module is used to train an initial neural network model by using the original image data set to form a classification neural network model, store the classification neural network model, and process the image data to be detected by using the classification neural network model to identify the object and judge the empty load state.

[0038] A robot communication module is used to receive the recognition result of the learning model processing module, and call the corresponding robot according to the recognition result.

[0039] The system does not need an accurate imaging module, and only needs to collect the original image data set in the database module to train the model, so that the object can be quickly identified by using a simple structure.

[0040] The application further provides a computer readable storage medium, and the computer readable storage medium stores instructions, and the instructions are executed by a processor to implement the substrate carrier recognition method. BRIEF DESCRIPTION OF DRAWINGS

[0041] FIG. 1 is a flowchart one of an embodiment of the application;

[0042] FIG. 2 is a flowchart two of the embodiment one of the application;

[0043] FIG. 3 is a flowchart one of an embodiment two of the application;

[0044] FIG. 4 is a flowchart two of the embodiment two of the application;

[0045] Figure 5a is a top view of a single substrate, a spacer, or a larger diameter object above in the second embodiment of the present application;

[0046] Figure 5b is a top view of a spacer as a spacer ring, the outer diameter of the spacer ring is larger than the diameter of the substrate, and the substrate is below the spacer ring in the second embodiment of the present application;

[0047] Figure 5c is a top view of a spacer as a spacer ring, the outer diameter of the spacer ring is larger than the diameter of the substrate, and the substrate is above the spacer ring in the second embodiment of the present application;

[0048] Figure 5d is a top view of a spacer as a spacer sheet, the diameter of the spacer sheet is larger than the diameter of the substrate, and the substrate is above the spacer sheet in the second embodiment of the present application;

[0049] Figure 5e is a top view of a spacer as a spacer ring, the outer diameter of the spacer ring is smaller than the diameter of the substrate, and the substrate is above the spacer ring in the second embodiment of the present application;

[0050] Figure 5f is a top view of a spacer as a spacer sheet, the diameter of the spacer sheet is smaller than the diameter of the substrate, and the substrate is below the spacer sheet in the second embodiment of the present application;

[0051] Figure 6 is a schematic diagram of a pretreated initial image data in the second embodiment of the present application.

[0052] Figure 7 is a system block diagram of the third embodiment of the present application;

[0053] Figure 8 is a schematic diagram of the structure of an imaging module and a semiconductor substrate carrier in the third embodiment of the present application.

[0054] In the figure: 1, imaging module; 11, camera; 12, lens; 13, illuminating element; 2, interface interaction module; 3, learning model processing module; 4, database module; 5, robot communication module; 6, buffer pad; 7, substrate; 8, spacer; 81, spacer ring; 82, spacer sheet; 9, central region. DETAILED DESCRIPTION

[0055] The preferred embodiments of the present application will be described in detail below with reference to the accompanying drawings, so that the advantages and features of the present application can be more easily understood by those skilled in the art, and the scope of protection of the present application can be more clearly defined.

[0056] Embodiment I

[0057] A substrate carrier recognition method of the present application is used for the recognition of different objects inside a substrate carrier. Referring to Figure 8, the objects include a substrate 7, a spacer 8, and a buffer pad 6 in a semiconductor substrate carrier. The buffer pad 6 is placed at the lowermost part of the placement cavity of the semiconductor substrate carrier, and then the spacer 8 and the substrate 7 are placed in sequence and at intervals.

[0058] Referring to Figure 1, the method comprises:

[0059] S100, acquire the to-be-tested image data of the uppermost layer in the semiconductor substrate carrier.

[0060] The to-be-tested image data is collected by the imaging module 1, as shown in FIG. 8, the imaging module 1 includes a camera 11, a lens 12 and an illuminating element 13, the illuminating element 13 is used for illuminating the object in the semiconductor substrate carrier, the lens 12 is connected with the camera 11, and the camera 11 is used for photographing and collecting image data.

[0061] The positions of the camera 11 and the illuminating element 13 are flexible, which can be arranged above or beside the semiconductor substrate carrier, as long as the camera 11 can capture the to-be-tested image data of the uppermost layer.

[0062] S200, send each to-be-tested image data to the pre-trained classification neural network model for object recognition.

[0063] The classification neural network model can quickly recognize the physical characteristics of the image data, and recognize the object according to the corresponding relationship between the physical characteristics and the object. That is, the physical characteristics and the object are in a corresponding mapping relationship, and after the physical characteristics are recognized, only one of the substrate 7, the spacer 8 and the buffer pad 6 can be obtained, so as to recognize the object. At the same time, the corresponding physical characteristics of different types of substrates 7, spacers 8 or buffer pads 6 are also different, so the specific type and model of the substrate 7, the spacer 8 or the buffer pad 6 can be located according to the physical characteristics.

[0064] S300, the classification neural network model judges whether the empty state is reached, if yes, the process is ended, if not, step 400 is executed.

[0065] When the empty state is reached, it indicates that there is no object to be transported and recognized in the semiconductor substrate carrier, at this time, the recognition of the object in the semiconductor substrate carrier can be ended.

[0066] S400, according to the recognition result in step S200, a corresponding mechanical hand is called to pick up the uppermost object.

[0067] Different objects may need different mechanical hands to pick up, that is:

[0068] When the uppermost object is recognized as the substrate 7, a mechanical hand suitable for clamping the substrate 7 is called to pick up the uppermost substrate 7;

[0069] When the uppermost object is recognized as the spacer 8, a mechanical hand suitable for clamping the spacer 8 is called to pick up the uppermost spacer 8;

[0070] When the topmost object is identified as the cushion 6, the mechanical hand adapted to the cushion 6 is called to pick up the topmost cushion 6.

[0071] In one embodiment, the cushion 6 can also be directly picked up manually. However, the substrate 7 and the spacer 8 must be picked up by the corresponding mechanical hand due to the process requirements.

[0072] After step S400 is completed, step S100 can be repeatedly executed until the classification neural network model in S300 determines that the empty state has been reached. At this time, the objects in a semiconductor substrate carrier have all been transported.

[0073] The identification of the substrate, the spacer, and the cushion can use photoelectric sensors, color photoelectric sensors, machine vision systems, and systems combining sensors and machine vision. However, photoelectric sensors and color photoelectric sensors do not have the ability to distinguish materials; machine vision systems can integrate multiple functions, but they require higher professional capabilities for detection and are more expensive; and systems combining sensors and machine vision may not be able to fully solve the inherent limitations of a single method. Therefore, in this embodiment, a classification neural network model is pre-trained and deployed to directly process image data, which can quickly determine the category of the object to call the corresponding mechanical hand. The neural network model has high recognition efficiency and fast response.

[0074] In one embodiment, the image data to be tested can completely cover the topmost layer of the semiconductor substrate carrier or partially cover the topmost layer of the semiconductor substrate carrier. That is, the image data to be tested can be an image of a complete object or an empty state, or an image of a partial object or an empty state. Because the physical characteristics of the object do not change whether it is complete or partial, the object can be identified and the empty state can be determined.

[0075] In one embodiment, referring to FIG. 2, the training process of the classification neural network model includes:

[0076] S11, collect physical characteristic data of the object to be identified.

[0077] The physical characteristics include size, color, surface flatness, surface granularity, and surface texture, etc. The physical characteristics of different objects are different, but the physical characteristics of one object are fixed and single. The physical characteristics and the object have a one-to-one mapping relationship, that is, after knowing the size, color, surface flatness, surface granularity, and surface texture, the specific object (including the type and model of the object) can be determined.

[0078] The substrate 7 has a size of 6 inches, 8 inches, 12 inches, etc. The surface of the substrate 7 has different colors due to different coating layers. The surface of the substrate 7 has different particle sizes due to different coating materials. The surface flatness and surface texture of the substrate 7 are different due to different process requirements. The spacer 8 has different colors, sizes, textures, and surface smoothness. The surface roughness, color, and size of the cushion 6 are also different. Therefore, after identifying all the physical characteristics of an object, the object is determined.

[0079] S12, collect a plurality of original image data of each object and an empty state to form and save an original image data set corresponding to each object and the empty state.

[0080] The original image data set corresponds to each object and the empty state one by one, that is, an object has an original image data set, and the empty state has an original image data set. The original image data belonging to the same original image data set is saved in the same folder, and the folder is named differently. For example, the folder substrate 1, the folder substrate 2, the folder spacer, the folder cushion, and the folder empty state. The original image data of the substrate 1 is saved in the folder substrate 1, the original image data of the substrate 2 is saved in the folder substrate 2, the original image data of the spacer is saved in the folder spacer, the original image data of the cushion is saved in the folder cushion, and the original image data of the empty state is saved in the folder empty state.

[0081] The original image data of each object is collected when the object is outside the semiconductor substrate carrier. The original image data contains the physical characteristics of the corresponding object, that is, all the physical characteristics of the object can be analyzed from the original image data.

[0082] The empty state refers to a state in which no object is placed in the semiconductor substrate carrier. The original image data of the object and the empty state can be collected by the imaging module 1 or other image collection devices. The empty state also has corresponding physical characteristics.

[0083] S13, input part or all of the original image data of the original image data set as a training set into an initial neural network model for training.

[0084] In an embodiment, the initial neural network model can be a model that has been trained on a large dataset, and using an already trained initial neural network model avoids the problem of training a new model from scratch, which requires a large dataset, a long training time, and expensive computing hardware. At this time, since the initial neural network model has been trained on a large dataset, the initial neural network model already has a certain reliability, which facilitates subsequent training. That is, in this embodiment, we use transfer learning to use the knowledge obtained from previous tasks to improve the generalization ability for another task, which is a widely used method in the field of image processing.

[0085] S14, the initial neural network model trained forms a classification neural network model, and the classification neural network model is stored and deployed.

[0086] In an embodiment, the initial neural network model uses a ResNet neural network, which has achieved success in image classification tasks due to its skip connections. The ResNet neural network alleviates the problem of gradient disappearance by promoting direct information flow between layers, improves the accuracy of image recognition and other tasks, and realizes the creation of an extremely deep neural network with hundreds of layers. The ResNet neural network is initially trained on a large dataset such as ImageNet. Then, through specific training sets, careful fine-tuning and optimization can obtain the final classification neural network model. At this time, the ResNet neural network only needs about 30 original image data for each object to meet the training requirements.

[0087] In another embodiment, the initial neural network model can also use an Inception V3 convolutional neural network.

[0088] The classification neural network model can identify the physical characteristics of the object and the empty state, and according to the identified physical characteristics, it can find the corresponding object or empty state, and then identify the object and judge the empty state.

[0089] In an embodiment, after the image data to be tested is identified, it can be updated and stored in the training set as a training set, and the initial neural network model can be repeatedly retrained to continuously improve the accuracy.

[0090] Before storing and deploying the classification neural network model, the trained initial neural network model needs to be tested, and only the initial neural network model that passes the test can be fixed as the classification neural network model. When the test fails, the initial neural network model needs to be adjusted and retrained and tested until the trained initial neural network model passes the test.

[0091] In an embodiment, the original image data in the original image data set that is not used as the training set can be used as a test set, and the test set is input into the initial neural network model that has been trained to test the model. Because the original image data in the original image data set has been manually classified, the result of the model classification can be compared with the result of the manual classification to test the model.

[0092] In an embodiment, test image data can also be additionally collected as a test set for storage, and the test set corresponds to each object and the empty state one by one, that is, one object has one test set, and the empty state has one test set. One test set is stored in one folder, and the folder is named differently.

[0093] The classification neural network model can be reused, and can be copied to other devices for use when needed.

[0094] When the number of objects is increased, the classification neural network model is retrained, that is, new training sets and test sets are added according to the increased objects, the fixed classification neural network model is retrained using the new training sets, and a new classification neural network model is formed after the new test set is passed. At this time, the model training is avoided from the beginning each time.

[0095] In an embodiment, the classification neural network model can also perform surface defect recognition to obtain the defect type of the substrate 7 to be recognized. The original image data in the training set includes substrates 7 with different defects, and the defect type is manually labeled on the original image data. The initial network model is trained by the training set, and the obtained classification network model can recognize the defect type. The surface defects of the substrate 7 include small notches, scratches, or cracks on the surface of the substrate 7.

[0096] Embodiment two

[0097] This embodiment discloses a substrate carrier recognition method, as shown in FIG. 3, which includes the following steps:

[0098] S101, obtaining the image data to be measured of the uppermost layer in the semiconductor substrate carrier.

[0099] S201, preprocessing the image data to be measured, which is the cutting of the center region 9 of the image data to be measured.

[0100] S301, sending each preprocessed image data to be measured to the pre-trained classification neural network model for object recognition.

[0101] S401, the classification neural network model judges whether the empty state is reached, if yes, the process is ended, if not, step 501 is executed.

[0102] S501, call the corresponding mechanical hand to pick up the uppermost object according to the recognition result in step S301.

[0103] The difference between this embodiment and embodiment one is that step S200 is added to pre-process the to-be-tested image data collected by the visual acquisition module, and the pre-processed image data is classified by the classification neural network model. Because the pre-processing is the cutting of the center region 9 of the to-be-tested image data, that is, removing part of the to-be-tested image data, therefore, by adding this step, the calculation amount of the image recognition processing of the classification neural network model is reduced, the recognition efficiency is improved, and the recognition efficiency is improved.

[0104] In this embodiment, referring to FIG. 4, the training process of the classification neural network model includes:

[0105] S21, collect the physical characteristic data of the object to be identified.

[0106] S22, collect a plurality of original image data of each object, different object combination state and empty load state to form an original image data set corresponding to each object, different object combination state and empty load state.

[0107] S23, pre-process each original image data in the original image data set and save the pre-processed original image data set, the pre-processing is cutting the center region 9 of the original image data.

[0108] The cutting range of the original image data is the same as the cutting range of the to-be-tested image data.

[0109] S24, input part or all of the original image data of the original image data set into the initial neural network model as a training set for training.

[0110] S25, the trained initial neural network model forms a classification neural network model, and the classification neural network model is stored and deployed.

[0111] In this embodiment, on the one hand, a pre-processing step is added before model training, part of the original image data is cut, which reduces the image processing calculation amount of the classification initial neural network model and speeds up the image processing rate. On the other hand, because the imaging module 1 collects the uppermost to-be-tested image data, the to-be-tested image data may include the substrate 7 and the spacer 8. In embodiment one, there may be a problem of accuracy when identifying. Therefore, the original image data of different object combinations, that is, the original image data of different substrate and different spacer combinations, is collected to quickly identify the uppermost substrate 7 or spacer 8 through pre-training, so as to improve the accuracy and speed of identification.

[0112] The spacer 8 can be a spacer sheet 82 or a spacer ring 81. The spacer sheet 82 is generally circular, and the diameter of the spacer sheet 82 is smaller than the diameter of the substrate 7, or in some special cases, the diameter of the spacer sheet 82 can be larger than the diameter of the substrate 7. The inner diameter of the spacer ring 81 is generally smaller than the diameter of the substrate 7, and the outer diameter of the spacer ring 81 can be larger or smaller than the diameter of the substrate 7. Therefore, when the spacer 8 and the substrate 7 are stacked, the imaging module 1 generally captures an image of the substrate 7 and the spacer 8 at the same time, and the captured image is a ring-shaped image. At this time, it is necessary to quickly identify whether the uppermost object is the spacer 8 or the substrate 7.

[0113] FIGS. 5a-5f are schematic diagrams of different combinations of objects, in which FIG. 5a is an image of a single substrate 7, a spacer 8, or an object with a larger diameter above, which is the original image of each object collected at this time. FIG. 5b is an original image collected when the spacer 8 is a spacer ring 81, the outer diameter of the spacer ring 81 is larger than the diameter of the substrate 7, and the substrate 7 is below the spacer ring 81. FIG. 5c is an original image collected when the spacer 8 is a spacer ring 81, the outer diameter of the spacer ring 81 is larger than the diameter of the substrate 7, and the substrate 7 is above the spacer ring 81. FIG. 5d is an original image collected when the spacer 8 is a spacer sheet 82, the diameter of the spacer sheet 82 is larger than the diameter of the substrate 7, and the substrate 7 is above the spacer sheet 82. FIG. 5e is an original image collected when the spacer 8 is a spacer ring 81, the outer diameter of the spacer ring 81 is smaller than the diameter of the substrate 7, and the substrate 7 is above the spacer ring 81. FIG. 5e is an original image collected when the spacer 8 is a spacer sheet 82, the diameter of the spacer sheet 82 is smaller than the diameter of the substrate 7, and the substrate 7 is below the spacer sheet 82. Collecting original image data of different combinations of objects increases the collection of such original image data and training, which can quickly and accurately identify whether the substrate 7 or the spacer 8 is the uppermost object, thereby improving the identification effect.

[0114] Of course, different combinations of objects can not only include the six cases mentioned above. It is necessary to collect original image data of all combinations of substrates 7 and spacer sheets 82 for model training to improve the recognition efficiency and accuracy of the model.

[0115] In one embodiment, referring to FIG. 6, the center region 9 is a region cut out with a center of original image data or image data to be tested as a center point and a preset diameter size. The initial neural network model and the classification neural network model process are both image data after the center region 9 is removed.

[0116] The preprocessing steps of the original image data and the image data to be tested are the same, including:

[0117] Determining the center point of the original image data and the image data to be tested;

[0118] Cutting out a region with the center point as the center and a preset diameter size.

[0119] The preset diameter size is smaller than the diameter of the substrate 7 and the inner diameter of the isolation ring 81, or smaller than the smaller one of the diameter of the substrate 7 and the diameter of the isolation piece 8. That is, it is ensured that after cutting, the combination of the objects will not be affected, and the identification will not be affected.

[0120] Of course, the substrate 7 and the isolation piece 8 can be polygonal structures, and the polygonal structures are also applicable to the above method.

[0121] Embodiment three

[0122] The embodiment discloses a substrate carrier identification system, which adopts the method described in embodiment one or embodiment two, and refers to FIG. 7. The identification system includes an imaging module 1, an interface interaction module 2, a learning model processing module 3, a database module 4, and a robot communication module 5.

[0123] The imaging module 1 is used to collect image data to be tested. The imaging module 1 includes a camera 11, a lens 12, and an illuminating piece 13. The illuminating piece 13 is used to illuminate the objects in the semiconductor substrate carrier. The lens 12 is connected with the camera 11. The camera 11 is used to take pictures and collect image data.

[0124] The interface interaction module 2 includes a result output interface display and a parameter configuration interface display. The parameter configuration interface display can input the physical characteristic data of the objects, such as the corresponding size, color, surface flatness, surface granularity, and surface texture of the substrate 7. When there is a new object to be identified, the physical characteristic data of the object is manually input in the parameter configuration interface. The result output interface display can display the identification result.

[0125] The database module 4 is used to store the original image data set corresponding to the objects and the empty state. The original image data belonging to the same original image data set is saved in the same folder, and different folders are named differently.

[0126] The learning model processing module 3 stores a pre-trained classification neural network model. The learning model processing module 3 processes the image data to be tested to identify the objects.

[0127] In one embodiment, the learning model processing module 3 includes a training unit, a preprocessing unit, and a model storage unit. The preprocessing unit is used to preprocess the image data to be tested and the initial image data. The model storage unit stores an initial neural network model and a classification neural network model after training. The training unit is used to call the initial neural network model and the original image data set to train the model and form the classification neural network model.

[0128] The robot communication module 5 is used to receive the identification result of the learning model processing module 3 and call the corresponding robot according to the identification result.

[0129] Embodiment Four

[0130] The application further provides a computer readable storage medium, and the computer readable storage medium stores instructions, and the instructions are executed by a processor to implement the substrate carrier identification method.

[0131] The program code contained in the computer readable medium can be transmitted by any suitable medium, including but not limited to wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0132] Those skilled in the art should understand that the embodiments of the application can be provided as a method, device, or computer program product. Therefore, the application can adopt a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the application can adopt a computer program product in the form of one or more computer usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) containing computer usable program code.

[0133] The above embodiments are only for illustrating the technical concept and characteristics of the application, and the purpose is to enable those skilled in the art to understand the content of the application and implement it, and cannot limit the protection scope of the application. Any equivalent changes or modifications made according to the spirit and principle of the application should be covered within the protection scope of the application.

Claims

1. A substrate carrier identification method for identifying different objects inside a substrate carrier, wherein the objects include a substrate, a spacer, and a buffer pad located within the semiconductor substrate carrier, characterized in that: The method comprises: acquiring the top layer of the semiconductor substrate carrier to be tested image data; each of the test image data is sent to the pre-trained classification neural network model for object recognition; the classification neural network model judges whether the empty state is reached, if yes, it is ended; if not, according to the recognition result of the classification neural network model, the corresponding mechanical hand is called to pick up the top layer object.

2. The substrate carrier identification method of claim 1, wherein: The training of the classification neural network model comprises: collecting the physical characteristics data of the objects to be identified; collecting a plurality of original image data of each of the objects and the empty state to form and save the original image data set corresponding to each of the objects and the empty state; part or all of the original image data of the original image data set is input into the initial neural network model as a training set for training; the trained initial neural network model forms a classification neural network model, and the classification neural network model is stored and deployed.

3. The substrate carrier identification method of claim 2, wherein: The initial neural network model is a model that has been trained on a data set, the initial neural network model uses ResNet neural network, and the data set is ImageNet data set.

4. The substrate carrier identification method of claim 2, wherein: Before the trained initial neural network model forms a classification neural network model, it further comprises: collecting a test set, inputting the test set into the trained initial neural network model for testing, and the initial neural network model that passes the test is a classification neural network model.

5. The method of claim 2, wherein: After the test image data is identified, it is stored in the training set and the training set is updated.

6. A method of identifying a substrate carrier according to any one of claims 2-5, characterized in that: Before each of the test image data is sent to the pre-trained classification neural network model for object recognition, it further comprises: preprocessing the test image data, which is cutting the center area of the test image data, updating the test image data to the preprocessed test image data.

7. The method of claim 6, wherein: While collecting a plurality of original image data of each of the objects and the empty state to form and save the original image data set corresponding to each of the objects and the empty state, it further comprises: collecting a plurality of original image data of different object combination states to form an original image data set corresponding to different object combination states.

8. The method of claim 7, wherein: Before part or all of the original image data of the original image data set is input into the initial neural network model as a training set for training, it further comprises: preprocessing the original image data, which is cutting the center area of the original image data, updating the original image data to the preprocessed original image data.

9. The method of claim 8, wherein: The center area is a region cut out with a center of the original image data or the test image data as a center point and a preset diameter size.

10. The method of claim 9, wherein: The isolation piece comprises an isolation ring and an isolation piece, and the preset diameter size is smaller than the substrate diameter and the inner diameter of the isolation ring, or the preset diameter size is smaller than the smaller one of the substrate diameter and the diameter of the isolation piece.

11. A substrate carrier identification system, characterized by: The system comprises: an imaging module, the imaging module is used for collecting test image data; A database module is configured to store original image data sets corresponding to different objects and empty states, and one original image data set is saved in one folder; A learning model processing module is configured to train an initial neural network model by the original image data sets to form a classification neural network model, store the classification neural network model, and process a to-be-tested image data by the classification neural network model to identify the object and determine the empty state; A robot communication module is configured to receive the identification result of the learning model processing module and call a corresponding robot according to the identification result.

12. A computer-readable storage medium, characterized in that: The computer readable storage medium stores instructions, and the instructions are executed by the processor to implement the substrate carrier identification method in any one of claims 1-10.

Images