Garbage sorting robot deep learning model updating method, device and system

Abstract

The application discloses a garbage classification robot deep learning model updating method, device and system, the method comprises the following steps: obtaining a to-be-updated data set, the to-be-updated data set comprises a plurality of classification failure data and a plurality of untrained data; screening and labeling first sample data from the to-be-updated data set, and adding the first sample data to a historical category training set; using the historical category training set to update the model through a preset deep learning model and generate an update key value; and sending the update key value to a garbage classification robot, so that the garbage classification robot uses the update garbage classification robot to update the model. The application can construct a training data set by using garbage image data that cannot be recognized by the robot when the robot cannot recognize the garbage, use the training data set to train the model and generate update parameters, and quickly update through the update parameters, so that the time consumption of model updating is shortened, and the model updating efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of updating of deep learning models, in particular to a deep learning model updating method, device and system for a garbage classification robot. BACKGROUND

[0002] With the development of economy, people's living conditions are gradually improving, and more and more garbage is generated in daily life. In order to avoid pollution of the environment due to too much garbage, garbage disposal is a very important problem. In order to uniformly and efficiently classify and dispose of different garbage, one of the garbage classification and disposal methods is to install a pre-trained deep learning model on an intelligent device (such as a garbage robot), to quickly identify the type of garbage through the intelligent device, so as to improve the efficiency of garbage classification.

[0003] In actual application, when a new garbage image data that is not used in the training process of the model is encountered, the classification is often wrong. In order to improve the classification ability, there are two commonly used updating learning methods at present. One is to integrate the new garbage image data into the original training data, and then use the re-integrated training data to train the learning model, so that the model can classify the new garbage. The other is to select appropriate samples from unannotated samples and give them to data engineers for re-annotation and addition to the training set, so that the model can continuously learn from the annotated samples to improve the accuracy of classification.

[0004] However, the current commonly used updating learning method has the following technical problems: annotating all new data and adding them to the training set makes the number of re-integrated training sets larger, increases the training model time, and reduces the updating learning efficiency. If the distribution of new data does not match the old data, the trained deep learning model will still tend to ignore the new data, and the updating effect is not good. If the re-annotation method is used for updating learning, it can only be performed in the state of class distribution matching. However, in actual scenarios, the probability of class distribution mismatch is high, which causes the sample selector to misselect, resulting in failure of active learning. SUMMARY

[0005] The present application provides a garbage classification robot classification learning method and device. When the deep learning model cannot classify new garbage, the method uses new garbage image data to edit and generate a key value of an updateable model, so that the model can be iteratively updated according to the key value, thereby shortening the time consumption of updating and improving the updating efficiency.

[0006] A first aspect of an embodiment of the present application provides a deep learning model updating method for a garbage classification robot, which comprises:

[0007] obtaining a to-be-updated data set, the to-be-updated data set comprising: a plurality of classification failure data uploaded by a garbage classification robot and a plurality of untrained data uploaded by a user, wherein the classification failure data is obtained after the garbage classification robot performs a classification operation on garbage image data;

[0008] filtering and labeling first sample data from the to-be-updated data set, and adding the first sample data to a historical category training set;

[0009] updating a model by using the historical category training set through a preset deep learning model, and generating an updated key value;

[0010] sending the updated key value to the garbage classification robot, so that the garbage classification robot updates the model by using the updated garbage classification robot.

[0011] In a possible implementation manner of the first aspect, the filtering and labeling first sample data from the to-be-updated data set comprises:

[0012] calculating a data score value of each data contained in the to-be-updated data set by using a query function, to obtain a plurality of data score values;

[0013] filtering a plurality of target data score values greater than a preset value from the plurality of data score values, and taking data corresponding to each target data score value as first sample data;

[0014] obtaining data labeling uploaded by the user, and adding the data labeling to the first sample data.

[0015] In a possible implementation manner of the first aspect, the calculating a data score value of each data contained in the to-be-updated data set by using a query function comprises:

[0016] calculating a semantic score value and a unique score value of each data contained in the to-be-updated data set by using a query function;

[0017] adding the semantic score value and the unique score value to obtain a data score value.

[0018] In a possible implementation manner of the first aspect, the semantic score value is calculated as shown in the following formula:

[0019]

[0020] The semantic score value is σ (·), cos (x, y) is used to calculate the similarity degree of x and y, and z s (·) represents a feature output by deep learning, The sample of the original data set is x, a sample of the to-be-updated data set;

[0021] The calculation of the unique score value is as follows:

[0022]

[0023] The unique score value is σ (·), cos (x, y) is used to calculate the similarity of x and y, z s (·) represents the features of the deep learning output, and respectively represent the samples with the first and second ranking of the nearest semantic similarity in the original data set and the sample of the to-be-updated data set.

[0024] In a possible implementation manner of the first aspect, after the step of screening and labeling the first sample data from the to-be-updated data set, the method further comprises:

[0025] screening a plurality of unlabeled second sample data from the to-be-updated data set to form an unlabeled to-be-classified training set;

[0026] extracting a model backbone component from a preset deep learning model, performing model training on the model backbone component by using the to-be-classified training set, and outputting and storing a plurality of class labels.

[0027] In a possible implementation manner of the first aspect, the garbage classification robot performs a classification operation on garbage image data, comprising:

[0028] After obtaining the garbage image data, the data confidence of the garbage image data is identified and calculated;

[0029] If the data confidence is lower than the preset confidence, it is determined that the garbage image data is classification failure data.

[0030] The second aspect of the embodiment of the application provides a deep learning model updating device of a garbage classification robot, comprising:

[0031] An acquisition module is configured to acquire a to-be-updated data set, wherein the to-be-updated data set comprises a plurality of classification failure data uploaded by a garbage classification robot and a plurality of untrained data uploaded by a user, and the classification failure data is obtained after the garbage classification robot performs a classification operation on garbage image data.

[0032] A screening and labeling module is configured to screen and label first sample data from the to-be-updated data set, and add the first sample data to a historical class training set.

[0033] ​The training module is configured to update the model by using the preset deep learning model and the historical category training set and generate an updated key value;

[0034] The updating module is configured to send the updated key value to the garbage classification robot, so that the garbage classification robot updates the model by using the updated garbage classification robot.

[0035] The third aspect of the embodiment of the present application provides a deep learning model updating system of a garbage classification robot, which comprises a background server and a garbage classification robot that interact with each other.

[0036] The garbage classification robot is configured to collect and identify garbage image data, and send the garbage image data that fails to be classified to the background server when the garbage image data fails to be classified.

[0037] The background server is configured to perform model training and updating according to the garbage image data that fails to be classified sent by the garbage classification robot, and generate an updated key value, and send the updated key value to the garbage classification robot, so that the garbage classification robot updates the model by using the updated key value.

[0038] Compared with the prior art, the deep learning model updating method, device and system of the garbage classification robot provided by the embodiment of the present application have the beneficial effects that when the robot fails to identify garbage, the garbage image data that cannot be identified is used to construct a new training data set, the model is trained by using the new training data set and the updated parameters are generated, the model is quickly updated by using the updated parameters, the time consumption of model updating is shortened, and the updating efficiency of the model is improved. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 is a flowchart of a deep learning model updating method of a garbage classification robot provided by an embodiment of the present application;

[0040] Figure 2 is a flowchart of a classification operation of garbage image data by a garbage classification robot provided by an embodiment of the present application;

[0041] Figure 3 is an operation flowchart of a deep learning model updating method of a garbage classification robot provided by an embodiment of the present application;

[0042] Figure 4 is a structural diagram of a deep learning model updating device of a garbage classification robot provided by an embodiment of the present application;

[0043] Figure 5 is a structural diagram of a deep learning model updating system of a garbage classification robot provided by an embodiment of the present application. DETAILED DESCRIPTION

[0044] The technical solutions in the embodiments of the present application will be apparently and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0045] The current commonly used updating learning method has the following technical problems: all new data are labeled and then added to the training set, so that the quantity of the re-integrated training set is increased, the training model time is increased, and the updating learning efficiency is reduced. If the distribution of new data and old data does not match, the trained deep learning model will still tend to old data and ignore new data, and the updating effect is not good. If the re-labeling method is used for updating learning, it can only be performed in the state of class distribution matching. However, in actual scenarios, the probability of class distribution mismatch is high, which causes the sample selector to misselect, resulting in the failure of active learning.

[0046] In order to solve the above problems, the deep learning model updating method of the garbage classification robot provided by the embodiments of the present application will be described and explained in detail below through the following specific embodiments.

[0047] Reference Figure 1 Fig. 1 shows a flowchart of a deep learning model updating method of a garbage classification robot according to an embodiment of the present application.

[0048] In an embodiment, the method is applicable to a background server which can be in communication connection with the garbage classification robot, and the background server is provided with the same deep learning model as the garbage classification robot. The model training can be performed in the background server, and then the training update parameters are sent to the robot, so that the robot can be quickly updated to improve the updating efficiency.

[0049] As an example, the deep learning model updating method of the garbage classification robot can include:

[0050] S11, obtaining a to-be-updated data set, the to-be-updated data set including a plurality of classification failure data uploaded by the garbage classification robot and a plurality of untrained data uploaded by a user, wherein the classification failure data is obtained by the garbage classification robot after performing a classification operation on garbage image data.

[0051] In order to enable the garbage classification robot to classify more different garbage and improve its classification ability, in a preferred embodiment, the garbage classification robot uploaded multiple classification failure data and the user uploaded multiple untrained data can be received at the same time. The model is updated and trained using the two kinds of data, so that the garbage classification robot can update the model using the updated data, thereby being able to identify and classify more different garbage.

[0052] Among them, the untrained data can be the original data obtained by the user shooting different garbage in the real scene. Among them, it can contain several different categories, which can not be involved in the previous training. In order to distinguish, the user can mark the data when uploading, and the marking information can be the category of the object.

[0053] In an optional embodiment, only the garbage classification robot uploaded multiple classification failure data can also be received, and the multiple classification failure data is used to construct the to-be-updated data set. Specifically, a certain number of classification failure data can be received to construct the to-be-updated data set.

[0054] Reference Figure 2 Fig. 1 shows a flowchart of the garbage classification robot provided by an embodiment of the present application performing classification operation on garbage image data.

[0055] Among them, in an embodiment, the garbage classification robot performs classification operation on garbage image data, including:

[0056] Substep S111, after obtaining the garbage image data, the data confidence of the garbage image data is identified and calculated.

[0057] After the garbage classification robot obtains the garbage image data (for example, the image data of the garbage), it can identify the garbage image data through its built-in deep learning model. In the process of identifying the picture containing the garbage, if the target is identified, the deep learning model can give the probability of inputting a certain category of the target. Using this probability information, the confidence level of the identification result can be rated to obtain the data confidence.

[0058] Substep S112, if the data confidence is lower than the preset confidence, it is determined that the garbage image data is classification failure data.

[0059] Compare the data confidence with the preset confidence. If the data confidence is lower than the preset confidence, it can be determined that the garbage image data is classification failure data. The garbage classification robot can upload the classification failure data to the background server, so that the background server can collect and perform model training.

[0060] S12, filter and mark the first sample data from the to-be-updated data set, and add the first sample data to the historical category training set.

[0061] In an embodiment, since the data to be updated in the data set contains a plurality of different data, there are data uploaded by the garbage classification robot and new garbage image data uploaded by the user without labeling and classification learning, some unlabeled data in the prior training can be screened from the plurality of data contained in the data set to be updated, and the model training using these new data can enable the robot to recognize these new data.

[0062] Specifically, after screening the first sample data, the first sample data can be stored in the category set set or labeled by the user, forming a historical category training set.

[0063] In order to accurately screen some new data which has not been trained by the model, in an optional embodiment, step S12 can include the following sub-steps:

[0064] Sub-step S121, calculating the data score value of each data contained in the data set to be updated by using the query function, obtaining a plurality of data score values.

[0065] In an embodiment, sub-step S121 can include the following sub-steps:

[0066] S1211, calculating the semantic score value and the unique score value of each data contained in the data set to be updated by using the query function.

[0067] Specifically, the calculation of the semantic score value is as follows:

[0068]

[0069] For the semantic score value, σ(·) is a normalization function, cos(x, y) is used to calculate the similarity of x and y, z s (·) represents the feature output by deep learning, is the sample of the original data set, is the sample of the data set to be updated;

[0070] The calculation of the unique score value is as follows:

[0071]

[0072] For the unique score value, σ(·) is a normalization function, cos(x, y) is used to calculate the similarity of x and y, z s (·) represents the feature output by deep learning, and respectively represent the sample of the original data set and the sample of the data set to be updated The samples ranked first and second with the latest semantic similarity.

[0073] S1212, add the semantic score value and the unique score value to obtain a data score value.

[0074] After addition, S quety = S sem + S dis .

[0075] Sub-step S122, filter a plurality of target data score values greater than a preset value from the plurality of data score values, and take data corresponding to each target data score value as first sample data.

[0076] In an embodiment, a plurality of target data score values greater than a preset value can be filtered from the plurality of data score values, and then data corresponding to the target data score values is taken as first sample data.

[0077] Alternatively, a plurality of target data score values can also be filtered from the plurality of data score values according to a preset number based on the size of the data score values. For example, if 5 first sample data are needed, the first 5 target data score values with the largest values can be filtered from the plurality of data score values.

[0078] Sub-step S123, obtain data annotation uploaded by a user, and add the data annotation to the first sample data.

[0079] If the filtered first sample data is untrained data uploaded by a user, the user can have added annotation to the data, and the data can not be annotated. If the user does not annotate the untrained data, the untrained data can also be annotated. If the filtered first sample data is classification failure data uploaded by a robot, the classification failure data can be displayed to the user, and then the user can annotate the classification failure data, and then the annotation is added to the first sample data.

[0080] Specifically, the annotated content or information can be a garbage category corresponding to the data, for example, a pop-top can, a plastic bag, paper scraps, etc.

[0081] The samples worth annotating are filtered out, but the remaining samples also have very low semantic similarity and very low unique score, i.e., S query very low score. These data can also be used for model training. In order to effectively use these untrained data and improve the classification ability of the model, in an embodiment, after step S12, the method can further include:

[0082] S21, filter a plurality of unannotated second sample data from the to-be-updated data set to form an unannotated to-be-classified training set.

[0083] Specifically, after screening a plurality of first sample data from the data set to be updated, the remaining non-first sample data is taken as second sample data, and a plurality of unlabeled second sample data is taken as a training set to be classified.

[0084] S22, extracting a model backbone component from a preset deep learning model, using the training set to be classified to train the model backbone component, and outputting and storing a plurality of class labels.

[0085] The preset deep learning model can be a model set in a background server. The CNN basic backbone excluding the classifier can be taken out from the preset deep learning model to obtain the model backbone component, and then the model backbone component is trained using the training set to be classified, so that it can recognize and output the class labels for the plurality of classified garbage, and each class label can correspond to a classification category.

[0086] The class labels are stored, and when receiving garbage image data corresponding to the class labels, the garbage image data can be used to form a historical class training set for subsequent model training.

[0087] S13, using the historical class training set to update the model through the preset deep learning model and generating an updated key value.

[0088] In an embodiment, the historical class training set and the old training set can be used to iteratively train the deep learning model, so as to obtain the model parameters of the deep learning model after training, and then integrate each model parameter into a key-value pair to obtain the updated key value.

[0089] S14, sending the updated key value to the garbage classification robot, so that the garbage classification robot uses the updated garbage classification robot to update the model.

[0090] The updated key value can be sent to the garbage classification robot. Since the deep learning model used by the garbage classification robot is the same as the deep learning model of the background server, the deep learning model of the garbage classification robot can directly update the key value to update the model, thereby quickly completing the update operation.

[0091] The application uses the newly built training data set to update and train the model, and directly sends the updated parameters to the robot, so that the robot can quickly update the model, effectively shorten the time consumption of the update operation, and improve the efficiency of the model update.

[0092] Referring to Figure 3 , a flowchart of a garbage classification robot deep learning model updating method provided by an embodiment of the application is shown.

[0093] Specifically, the background server can receive the instructions of the garbage classification robot and receive the classification failure data sent by the garbage classification robot, then build the classification failure data and the un-labeled un-trained data uploaded by the user into a to-be-updated data set, then select appropriate first sample data from the to-be-updated data set, let the first sample data build a new training data set T1, and build a new data set T2 from the second sample data which is not selected. The T1 is used for iterative updating and training of the model to generate model update parameters (key-value pairs), and the model update parameters are sent to the robot to make the robot update the model; the T2 is used for update learning to generate new category labels for subsequent classification of new data.

[0094] In the embodiment, the garbage classification robot deep learning model updating method has the beneficial effects that when the robot cannot identify garbage, the garbage image data that cannot be identified is used to build a new training data set, the new training data set is used for model training to generate update parameters, and the update parameters are used for rapid updating of the model, thereby shortening the time consumption of model updating and improving the updating efficiency of the model.

[0095] The embodiment of the application also provides a garbage classification robot deep learning model updating device, referring to Figure 4 , a structure schematic diagram of a garbage classification robot deep learning model updating device provided by an embodiment of the application is shown.

[0096] For example, the garbage classification robot deep learning model updating device can include:

[0097] The acquisition module 501 is configured to acquire a to-be-updated data set, and the to-be-updated data set includes a plurality of classification failure data uploaded by a garbage classification robot and a plurality of un-trained data uploaded by a user, wherein the classification failure data is obtained after the garbage classification robot performs a classification operation on garbage image data.

[0098] The screening and labeling module 502 is configured to screen and label first sample data from the to-be-updated data set, and add the first sample data to a historical category training set.

[0099] The training module 503 is configured to use the historical category training set to update the model by a preset deep learning model and generate update keys and values.

[0100] The updating module 504 is configured to send the update keys and values to the garbage classification robot, so that the garbage classification robot uses the update garbage classification robot to update the model.

[0101] Optionally, the screening and labeling module is further configured to:

[0102] calculating a data score value of each data contained in the to-be-updated data set by using a query function, to obtain a plurality of data score values;

[0103] filtering a plurality of target data score values greater than a preset value from the plurality of data score values, and taking data corresponding to each target data score value as first sample data;

[0104] obtaining data labeling uploaded by a user, and adding the data labeling to the first sample data.

[0105] Optionally, the screening and labeling module is further configured to:

[0106] calculating a semantic score value and a unique score value of each data contained in the to-be-updated data set by using a query function;

[0107] adding the semantic score value and the unique score value to obtain a data score value.

[0108] Optionally, the semantic score value is calculated according to the following formula:

[0109]

[0110] wherein the semantic score value is σ, σ(·) is a normalization function, cos(x, y) is used to calculate the similarity between x and y, and z s (·) represents a feature output by deep learning, is a sample of the original data set, is a sample of the to-be-updated data set;

[0111] The unique score value is calculated according to the following formula:

[0112]

[0113] wherein the unique score value is σ, σ(·) is a normalization function, cos(x, y) is used to calculate the similarity between x and y, and z s (·) represents a feature output by deep learning, and respectively represent samples in the original data set and the to-be-updated data set having the closest semantic similarity.

[0114] Optionally, the apparatus further comprises:

[0115] an unlabeled module configured to filter a plurality of unlabeled second sample data from the to-be-updated data set to form an unlabeled to-be-classified training set;

[0116] The category training module is used for extracting a model backbone component from a preset deep learning model, performing model training on the model backbone component by using the to-be-classified training set, and outputting and storing multiple category labels.

[0117] Optionally, the garbage classification robot performs a classification operation on garbage image data, including:

[0118] After obtaining the garbage image data, data confidence of the garbage image data is identified and calculated.

[0119] If the data confidence is lower than a preset confidence, it is determined that the garbage image data is classification failure data.

[0120] The embodiment of the application further provides a garbage classification robot deep learning model updating system, referring to Figure 5 , a structural schematic diagram of a garbage classification robot deep learning model updating system provided by an embodiment of the application is shown.

[0121] For example, the garbage classification robot deep learning model updating system can include:

[0122] A background server and a garbage classification robot for interaction;

[0123] The garbage classification robot is used for collecting and identifying garbage image data, and sending the garbage image data of classification failure to the background server when the garbage image data fails to be classified.

[0124] The background server is used for performing model training and updating according to the garbage image data of classification failure sent by the garbage classification robot, and generating an update key value, and sending the update key value to the garbage classification robot, so that the garbage classification robot performs model updating by using the update key value.

[0125] Specifically, the background server can include a server module, a server deep learning model, a model data storage module, a server communication module and an image data storage module.

[0126] The server module is connected with the server deep learning model and the server communication module respectively, the server communication module is connected with the model data storage module and the image data storage module respectively, and the server deep learning model is connected with the model data storage module.

[0127] Specifically, the garbage classification robot can include a central control module, a robot deep learning model, a robot data storage module, a robot communication module and a camera module for collecting images.

[0128] The central control module is connected with the robot deep learning model, the robot data storage module, the robot communication module and the camera module for collecting images respectively, the robot deep learning model is connected with the camera module for collecting images, and the robot communication module is connected with the server communication module, so as to transmit garbage image data and update parameters between the background server and the garbage classification robot.

[0129] In an embodiment, the camera module can accept the instruction of the central control module, complete the parameter setting of the camera device such as height and inclination, obtain image data, complete target detection and identification by the robot deep learning model, store in the robot data storage module after the instruction of the central control module, and feed back the current working state information to the central control module.

[0130] The robot communication module is a communication interface between the robot system and the outside world, can complete bidirectional communication with the central control module, can accept the control instruction from the outside world, and can feed back the current working state information of the robot to the outside world.

[0131] The robot data storage module accepts the instruction of the central control module, feeds back the current working state information to the central control module, and stores large-scale data from the data collection module.

[0132] The central control module is the decision center of the robot system, can obtain the working state information of the remaining modules, and can send instructions to the remaining modules, so as to coordinate the work of the modules.

[0133] The robot deep learning model is a logical model for identifying data of the robot, can identify the garbage target picture in the image captured by the camera, and can receive the model parameter update instruction from the central control module, update the model parameters that need to be updated, and realize online update.

[0134] Those skilled in the art can clearly understand that, for the convenience of description and brevity, the specific working process of the above-mentioned device or system can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0135] Further, the embodiment of the present application also provides an electronic device, which comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor implements the deep learning model updating method of the garbage classification robot according to the above-mentioned embodiment when executing the program.

[0136] Further, the embodiment of the present application also provides a computer readable storage medium, which stores computer executable instructions, and the computer executable instructions are used to make the computer execute the deep learning model updating method of the garbage classification robot according to the above-mentioned embodiment.

[0137] The above is the preferred embodiment of the present application, it should be noted that for those skilled in the art, without departing from the principles of the present application, can also make several improvements and refinements, these improvements and refinements are also considered the scope of protection of the present application.

Claims

1. A method for updating a deep learning model of a waste sorting robot, the method comprising: The method comprises: obtaining a to-be-updated data set, the to-be-updated data set comprising: a plurality of classification failure data uploaded by a garbage classification robot and a plurality of untrained data uploaded by a user, wherein the classification failure data is obtained after the garbage classification robot performs a classification operation on garbage image data; screening and labeling first sample data from the to-be-updated data set, and adding the first sample data to a historical category training set; wherein the screening and labeling of the first sample data from the to-be-updated data set comprises: calculating a data score value of each data included in the to-be-updated data set by using a query function to obtain a plurality of data score values; screening a plurality of target data score values greater than a preset value from the plurality of data score values, taking data corresponding to each target data score value as first sample data; obtaining data labeling uploaded by a user, and adding the data labeling to the first sample data; the calculation of the data score value of each data included in the to-be-updated data set by using the query function comprises: calculating a semantic score value and a unique score value of each data included in the to-be-updated data set by using the query function; and adding the semantic score value and the unique score value to obtain a data score value; updating a model by using the historical category training set through a preset deep learning model to generate an updated key value; sending the updated key value to the garbage classification robot, so that the garbage classification robot updates the model by using the updated key value. 2.The method of claim 1, wherein The calculation of the semantic score value is as shown in the following formula: is a semantic score value, σ(·) is a normalization function, cos(x, y) is used to calculate the similarity of x and y, z s (·) represents the features of the deep learning output, is a sample of the original data set, is a sample of the data set to be updated; The calculation of the unique score value is as shown in the following formula: is a unique score value, σ(·) is a normalization function, cos(x, y) is used to calculate the similarity of x and y, z s (·) represents the features of the deep learning output, and respectively represent the samples in the original data set and the sample to be updated in the data set with the nearest semantic similarity. The rankings of the first and second samples are first and second. 3.The method of claim 1, wherein the garbage classification robot is a robot that classifies garbage into a plurality of types, and the plurality of types of garbage include at least one of paper, plastic, metal, and organic material. After the step of screening and labeling first sample data from the to-be-updated data set, the method further comprises: screening a plurality of unlabeled second sample data from the to-be-updated data set to form an unlabeled to-be-classified training set; extracting a model backbone component from a preset deep learning model, and performing model training on the model backbone component by using the to-be-classified training set to output and store a plurality of category labels.

4. The deep learning model update method for the garbage sorting robot according to claim 1, characterized in that, The classification operation of the garbage classification robot on garbage image data comprises: after obtaining garbage image data, identifying and calculating a data confidence of the garbage image data; if the data confidence is lower than a preset confidence, determining that the garbage image data is classification failure data. 5.A device for updating a deep learning model of a waste sorting robot, characterized by, The device comprises: an obtaining module configured to obtain a to-be-updated data set, the to-be-updated data set comprising: a plurality of classification failure data uploaded by a garbage classification robot and a plurality of untrained data uploaded by a user, wherein the classification failure data is obtained after the garbage classification robot performs a classification operation on garbage image data; a screening and labeling module configured to screen and label first sample data from the to-be-updated data set, and add the first sample data to a historical category training set; wherein the screening and labeling of the first sample data from the to-be-updated data set comprises: The query function is used to calculate a data score value of each data included in the to-be-updated data set, to obtain a plurality of data score values; a plurality of target data score values greater than a preset value are screened from the plurality of data score values, and data corresponding to each target data score value is first sample data; a data label uploaded by a user is obtained, and the data label is added to the first sample data; the query function is used to calculate a semantic score value and a unique score value of each data included in the to-be-updated data set; the semantic score value and the unique score value are added to obtain a data score value; The training module is configured to update the model by using the preset deep learning model and the historical category training set, and generate an update key value; The updating module is configured to send the update key value to the garbage classification robot, so that the garbage classification robot updates the model by using the update key value. 6.A deep learning model updating system of a waste sorting robot, characterized by, The system comprises a background server and a garbage classification robot that interact with each other, and the background server performs the garbage classification robot deep learning model updating method according to any one of claims 1-4. The garbage classification robot is configured to collect and identify garbage image data, and send the garbage image data that fails to be classified to the background server when the garbage image data fails to be classified. The background server is configured to perform model training and updating according to the garbage image data that fails to be classified sent by the garbage classification robot, and generate an update key value, and send the update key value to the garbage classification robot, so that the garbage classification robot updates the model by using the update key value.

7. An electronic device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the garbage classification robot deep learning model updating method according to any one of claims 1-4 when executing the program.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer executable instructions for causing a computer to execute the garbage classification robot deep learning model updating method according to any one of claims 1-4.

Images