CN112748941A

CN112748941A - Feedback information-based target application program updating method and device

Info

Publication number: CN112748941A
Application number: CN202010784873.1A
Authority: CN
Inventors: 张维; 陈卫东; 暴林超
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2020-08-06
Filing date: 2020-08-06
Publication date: 2021-05-04
Anticipated expiration: 2040-08-06
Also published as: CN112748941B

Abstract

The invention discloses a method and a device for updating a target application program based on feedback information. The method comprises the following steps: acquiring actual scene information input in a target application program and result feedback information corresponding to the actual scene information; updating a target neural network model in a target application program according to the actual scene information and the result feedback information, wherein the target neural network model is a neural network model obtained by performing model training on a pre-trained neural network model by using a training sample set, the training sample set comprises at least two groups of sample scene information, different groups of sample scene information correspond to different actual sample recognition results of a target object, loss functions used in model training have different corresponding weights in different groups of sample scene information, and the weights corresponding to different groups of sample scene information are in negative correlation with the number of sample scene information in different groups of sample scene information.

Description

Feedback information-based target application program updating method and device

Technical Field

The invention relates to the field of computers, in particular to a method and a device for updating a target application program based on feedback information, a storage medium and electronic equipment.

Background

At present, in some small sample service scenes, the model can be obtained generally by training in advance.

For the training of the model, the adopted scheme is generally the deployment after off-line training, namely the model is trained by using the marked data, and after the model is formally deployed and on-line, the parameters of the model are all fixed. However, for a small-sample service scene, the number of training samples that can be used is small, and there are cases where the samples are not uniformly distributed, resulting in a poor classification effect of the model obtained for the small-sample service scene.

Aiming at the problem that in the related art, in a small sample business scene, due to the fact that the number of training samples is small and the samples are not uniformly distributed, a model obtained through training in the small sample business scene has a poor classification effect, and an effective solution is not provided.

Disclosure of Invention

The embodiment of the invention provides a method and a device for updating a target application program based on feedback information, a storage medium and electronic equipment, and aims to at least solve the technical problem that in the related technology, in a small sample business scene, a model obtained through training in the small sample business scene has a poor classification effect due to the fact that the number of training samples is small and the samples are unevenly distributed.

According to an aspect of the embodiments of the present invention, there is provided a method for updating a target application based on feedback information, including: acquiring actual scene information input in the target application program and result feedback information corresponding to the actual scene information, wherein the target application program uses a target neural network model, the result feedback information is feedback information of an actual recognition result output by the target neural network model, and the actual recognition result is an estimated recognition result of the target object output by the target neural network model according to the actual scene information; updating the target neural network model in the target application program according to the actual scene information and the result feedback information; the target neural network model is obtained by performing model training on a pre-trained neural network model by using a training sample set, wherein the pre-trained neural network model is used for outputting an estimated recognition result of a target object according to input scene information, the training sample set comprises at least two sets of sample scene information, the at least two sets of sample scene information and at least two sets of actual sample recognition results have a one-to-one correspondence relationship, different sets of sample scene information correspond to different actual sample recognition results of the target object, corresponding weights of loss functions used in model training are different in different sets of sample scene information, and the weights corresponding to the different sets of sample scene information are in negative correlation with the number of sample scene information in the different sets of sample scene information.

According to another aspect of the embodiments of the present invention, there is also provided an apparatus for updating a target application based on feedback information, including: a first acquiring unit configured to acquire actual scene information input in a target application program using a target neural network model and result feedback information corresponding to the actual scene information, the result feedback information being feedback information of an actual recognition result output by the target neural network model, the actual recognition result being an estimated recognition result of the target object output by the target neural network model based on the actual scene information; a first processing unit, configured to update the target neural network model in the target application according to the actual scene information and the result feedback information; the target neural network model is obtained by performing model training on a pre-trained neural network model by using a training sample set, wherein the pre-trained neural network model is used for outputting an estimated recognition result of a target object according to input scene information, the training sample set comprises at least two sets of sample scene information, the at least two sets of sample scene information and at least two sets of actual sample recognition results have a one-to-one correspondence relationship, different sets of sample scene information correspond to different actual sample recognition results of the target object, corresponding weights of loss functions used in model training are different in different sets of sample scene information, and the weights corresponding to the different sets of sample scene information are in negative correlation with the number of sample scene information in the different sets of sample scene information.

According to an aspect of the embodiments of the present invention, there is provided an image recognition method based on a target application, including: inputting image information to a target application program, wherein the target application program uses a target neural network model; identifying the image information through the target neural network model to obtain an image identification result, wherein the image identification result is an estimated identification result of a target object output by the target neural network model according to the image information; wherein the target neural network model is obtained by performing model training on a pre-trained neural network model by using a training sample set, wherein, the pre-training neural network model is used for outputting the pre-estimation recognition result of the target object according to the input scene information, the training sample set is a set obtained by preprocessing an image set to be processed, the training sample set comprises at least two groups of sample scene information, the at least two groups of sample scene information and the at least two groups of actual sample identification results have a one-to-one correspondence relationship, different groups of sample scene information correspond to different actual sample identification results of the target object, the loss function used in the model training has different corresponding weights in different sets of sample scene information, the weights corresponding to the different sets of sample scene information are inversely related to the number of sample scene information in the different sets of sample scene information.

According to another aspect of the embodiments of the present invention, there is also provided an image recognition apparatus based on a target application, including: an input unit configured to input image information to a target application program, wherein the target application program uses a target neural network model; a second processing unit, configured to identify the image information through the target neural network model to obtain an image identification result, where the image identification result is an estimated identification result of a target object output by the target neural network model according to the image information; wherein the target neural network model is obtained by performing model training on a pre-trained neural network model by using a training sample set, wherein, the pre-training neural network model is used for outputting the pre-estimation recognition result of the target object according to the input scene information, the training sample set is a set obtained by preprocessing an image set to be processed, the training sample set comprises at least two groups of sample scene information, the at least two groups of sample scene information and the at least two groups of actual sample identification results have a one-to-one correspondence relationship, different groups of sample scene information correspond to different actual sample identification results of the target object, the loss function used in the model training has different corresponding weights in different sets of sample scene information, the weights corresponding to the different sets of sample scene information are inversely related to the number of sample scene information in the different sets of sample scene information.

According to yet another aspect of the application, a computer program product or computer program is provided, comprising computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions to enable the computer device to execute the method provided in the various optional implementation modes of the feedback information-based target application program updating method.

According to another aspect of the embodiments of the present invention, there is also provided an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the method for updating the target application based on the feedback information through the computer program.

According to the invention, a target neural network model needs to be trained in advance, a pre-trained neural network model suitable for identifying a target object is obtained according to scene information, the pre-trained neural network model can output an estimated identification result of the target object according to the input scene information, then the pre-trained neural network model is subjected to model training by using a training sample set to obtain the target neural network model, wherein the training sample set comprises at least two groups of sample scene information, the at least two groups of sample scene information correspond to at least two groups of actual sample identification results in a one-to-one manner, different groups of sample scene information correspond to different actual sample identification results of the target object, loss functions used in model training have different corresponding weights in different groups of sample scene information, and the weights corresponding to the different groups of sample scene information are in negative correlation with the number of sample scene information in the different groups of sample scene information, then, actual scene information and result feedback information corresponding to the actual scene information may be input in a target application program using a target neural network model, where the result feedback information is feedback information for an actual recognition result output by the target neural network model, and the actual recognition result is an estimated recognition result of the target object output by the target neural network model according to the actual scene information; and finally, updating the target neural network model in the target application program on line through the actual scene information and the result feedback information. By adopting the scheme, the corresponding weights of the loss functions used in the model training in different groups of sample scene information are set to be different, so that the different groups of sample scene information can be distributed in a balanced manner, the recognition effect on the different groups of sample scene information is improved, and the target neural network model in the target application program is updated through the real actual scene information and the result feedback information, so that the training samples can be enriched, the target neural network model can be more accurate, and the recognition accuracy of the target neural network model can be improved. In a small sample service scene, the weights of different types of samples of the small samples are adjusted through the loss function, so that the samples of different types of the small samples are uniformly distributed, and the effect of improving the classification and identification of the corresponding models in the small sample service scene is achieved.

Drawings

The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this application, illustrate embodiment(s) of the invention and together with the description serve to explain the invention without limiting the invention. In the drawings:

fig. 1 is a schematic diagram of an application environment of a feedback information-based target application update method according to an embodiment of the present invention;

FIG. 2 is a flowchart illustrating an alternative feedback information based update method for a target application according to an embodiment of the present invention;

FIG. 3 is a schematic flow diagram of an alternative method for updating a target neural network model, according to an embodiment of the present invention;

FIG. 4 is a schematic flow diagram of an alternative method for updating a target neural network model, according to an embodiment of the present invention;

FIG. 5 is a schematic flow diagram of yet another alternative method of updating a target neural network model, according to an embodiment of the present invention;

FIG. 6 is a schematic diagram illustrating an alternative process for pre-processing a picture according to an embodiment of the present invention;

FIG. 7 is a flowchart illustrating an alternative method for target application based image recognition according to embodiments of the present invention;

FIG. 8 is a flow chart illustrating an alternative optical disc identification based feedback according to an embodiment of the present invention;

FIG. 9 is a schematic diagram of an alternative user feedback information according to an embodiment of the present invention;

FIG. 10 is a schematic flow chart diagram of an alternative feedback information based target application according to an embodiment of the present invention;

FIG. 11 is a schematic structural diagram of an alternative apparatus for updating a target application based on feedback information according to an embodiment of the present invention;

FIG. 12 is a block diagram of an alternative target application based image recognition apparatus according to an embodiment of the present invention;

fig. 13 is a schematic structural diagram of an alternative electronic device according to an embodiment of the invention.

Detailed Description

In order to make the technical solutions of the present invention better understood, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

It should be noted that the terms "first," "second," and the like in the description and claims of the present invention and in the drawings described above are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used is interchangeable under appropriate circumstances such that the embodiments of the invention described herein are capable of operation in sequences other than those illustrated or described herein. Furthermore, the terms "comprises," "comprising," and "having," and any variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, system, article, or apparatus that comprises a list of steps or elements is not necessarily limited to those steps or elements expressly listed, but may include other steps or elements not expressly listed or inherent to such process, method, article, or apparatus.

Technical terms involved in the embodiments of the present invention include:

1. neural Networks (NN) are complex network systems formed by widely interconnecting a large number of simple processing units (called neurons), reflect many basic features of human brain functions, and are highly complex nonlinear dynamical learning systems. The neural network has the capabilities of large-scale parallel, distributed storage and processing, self-organization, self-adaptation and self-learning, and is particularly suitable for processing inaccurate and fuzzy information processing problems which need to consider many factors and conditions simultaneously. For example, a Convolutional Neural Network (CNN) is a feed-forward Neural Network with a deep structure and containing convolution calculation, and its artificial neurons can respond to peripheral cells in a part of coverage range, and has excellent performance for large-scale image processing.

2. Long tail data: labels in the index data are distributed unevenly, most data are concentrated under a few labels, and the deep learning model has poor effect on long-tail data; for example, the number of the image samples is 500, the image samples include samples of three categories, the samples of the three categories include that the tray in the image is empty (the number of samples is 110), the tray in the image is not empty (the number of samples is 90), and the image does not have a tray (the number of samples is 400), accordingly, the recognition result of the deep learning model on the image samples and the corresponding label of the recognition result also include three types, the recognition result is that the tray in the image is not empty, and the corresponding label is 1; the identification result is that the tray in the image is empty, and the corresponding label is 2; the recognition result is that there is no plate in the image and the corresponding label is 3. In this case, the distribution of the labels 1, 2 and 3 is unbalanced, so that the recognition accuracy of the deep learning model for the label 3 is high, and the recognition accuracy for the labels 1 and 2 is low, so that the effect of the deep learning model on long-tail data is poor.

3. Pre-training the model: the method refers to that after a model is trained on a larger task, the feature extraction capability of the model can be transferred to other tasks. For example, the model developed for task A is reused as an initial point in the process of developing the model for task B. For example, a pre-trained model on ImageNet can be used as a pre-trained neural network model in the update method of the target application based on the feedback information, and the capability of the trained neural network model for image recognition can be generalized to the target neural network model according to the embodiment of the present application.

4. Image dataset (a Large-Scale Hierarchical Image Database, ImageNet for short): a large visual database is used for visual target recognition software research, is the largest database for image recognition in the world, and can recognize objects from pictures.

According to an aspect of the embodiments of the present invention, there is provided a method for updating a target application based on feedback information. Alternatively, the above-mentioned update method of the target application based on the feedback information can be applied, but not limited, to the application environment as shown in fig. 1. As shown in fig. 1, acquiring, by a terminal device 102, actual scene information input in a target application program that uses a target neural network model and result feedback information corresponding to the actual scene information, the result feedback information being feedback information for an actual recognition result output by the target neural network model, the actual recognition result being an estimated recognition result of the target object output by the target neural network model based on the actual scene information, and updating, by the terminal device 102, the target neural network model in the target application program based on the actual scene information and the result feedback information; the server 104 needs to pre-train a target neural network model, wherein the target neural network model is a network model obtained by model training a pre-trained neural network model by using a training sample set, wherein, the pre-training neural network model is used for outputting the pre-estimation recognition result of the target object according to the input scene information, the training sample set comprises at least two groups of sample scene information, the at least two groups of sample scene information and at least two groups of actual sample recognition results have one-to-one correspondence, different groups of sample scene information correspond to different actual sample recognition results of the target object, the loss function used in the model training has different corresponding weights in different sets of sample scene information, the weights corresponding to the different sets of sample scene information are inversely related to the number of sample scene information in the different sets of sample scene information. The above is merely an example, and the embodiments of the present application are not limited herein.

Alternatively, the method may be applied to a business scene of image recognition, such as a scene of recognizing whether tableware is empty, recognizing a specific object (e.g. a person or an animal) in an image, and the like, and the embodiment is not limited in any way herein.

The updating of the target application program based on the feedback information in the embodiment of the invention relates to technologies such as artificial intelligence, machine learning and the like, wherein:

artificial Intelligence (AI) is a theory, method, technique and application system that uses a digital computer or a machine controlled by a digital computer to simulate, extend and expand human Intelligence, perceive the environment, acquire knowledge and use the knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technique of computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a manner similar to human intelligence. Artificial intelligence is the research of the design principle and the realization method of various intelligent machines, so that the machines have the functions of perception, reasoning and decision making.

The artificial intelligence technology is a comprehensive subject and relates to the field of extensive technology, namely the technology of a hardware level and the technology of a software level. The artificial intelligence infrastructure generally includes technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technologies, operation/interaction systems, mechatronics, and the like. The artificial intelligence software technology mainly comprises a computer vision technology, a voice processing technology, a natural language processing technology, machine learning/deep learning and the like.

Machine Learning (ML) is a multi-domain cross discipline, and relates to a plurality of disciplines such as probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and the like. The special research on how a computer simulates or realizes the learning behavior of human beings so as to acquire new knowledge or skills and reorganize the existing knowledge structure to continuously improve the performance of the computer. Machine learning is the core of artificial intelligence, is the fundamental approach for computers to have intelligence, and is applied to all fields of artificial intelligence. Machine learning and deep learning generally include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and formal education learning.

Optionally, in this embodiment, the terminal device may be a terminal device configured with a target client, and may include, but is not limited to, at least one of the following: mobile phones (such as Android phones, iOS phones, etc.), notebook computers, tablet computers, palm computers, MID (Mobile Internet Devices), PAD, desktop computers, smart televisions, etc. The target client may be a video client, an instant messaging client, a browser client, an educational client, etc. Such networks may include, but are not limited to: a wired network, a wireless network, wherein the wired network comprises: a local area network, a metropolitan area network, and a wide area network, the wireless network comprising: bluetooth, WIFI, and other networks that enable wireless communication. The server may be a single server, a server cluster composed of a plurality of servers, or a cloud server. The above is only an example, and the present embodiment is not limited to this.

Optionally, in this embodiment, as an optional implementation manner, the method may be executed by a server, or may be executed by a terminal device, or may be executed by both the server and the terminal device, and in this embodiment, the description is given by taking an example that the server (for example, the server 104) executes. As shown in fig. 2, the flow of the method for updating the target application based on the feedback information may include the steps of:

step S202 is to acquire actual scene information input in a target application program using a target neural network model and result feedback information corresponding to the actual scene information, where the result feedback information is feedback information of an actual recognition result output by the target neural network model, and the actual recognition result is an estimated recognition result of the target object output by the target neural network model according to the actual scene information.

Step S204, updating the target neural network model in the target application program according to the actual scene information and the result feedback information.

The target neural network model is obtained by performing model training on a pre-trained neural network model by using a training sample set, wherein the pre-trained neural network model is used for outputting an estimated recognition result of a target object according to input scene information, the training sample set comprises at least two sets of sample scene information, the at least two sets of sample scene information and at least two sets of actual sample recognition results have a one-to-one correspondence relationship, different sets of sample scene information correspond to different actual sample recognition results of the target object, corresponding weights of loss functions used in model training are different in different sets of sample scene information, and the weights corresponding to the different sets of sample scene information are in negative correlation with the number of sample scene information in the different sets of sample scene information.

Optionally, before the target application is formally released online, the target neural network model used by the target application needs to be trained in advance, and the method further includes: acquiring the pre-trained neural network model, wherein the pre-trained neural network model is used for outputting a pre-estimated recognition result of a target object according to input scene information; and performing model training on the pre-trained neural network model by using the training sample set to obtain the target neural network model.

The target application may be a sort program, such as an optical disc identification applet (identifying whether the tableware is empty), an animal identification program, a flower and plant identification program, a clothing identification program, and the like, which are not limited herein.

In general, the number of training samples of a target application program is insufficient, and the generalization capability of a model obtained by a direct training mode is far from being sufficient. The pre-training model on ImageNet can be used as the pre-training neural network model, the generalization capability of the algorithm can be greatly improved by using the pre-training neural network model, and a relatively ideal effect can be obtained only by needing less fine tuning times. Through the pre-training neural network model, the estimation recognition result of the target object (such as tableware, flowers and plants, clothes, animals and the like) can be output according to the input scene information (such as images and the like).

And performing model training on the pre-trained neural network model by using the training sample set to obtain a target neural network model.

For example, for optical disc identification, the actual sample identification result corresponding to one group of sample scene information (pictures) is that the current sample scene information is an empty disc, the actual sample identification result corresponding to one group of sample scene information (pictures) is that the current sample scene information is a non-empty disc, and the actual sample identification result corresponding to one group of sample scene information (pictures) is that no disc exists in the current sample scene information.

During model training, a loss function is used, the corresponding weights of the loss function in different sets of sample scene information are different, and the weights corresponding to the different sets of sample scene information are in negative correlation with the number of the sample scene information in the different sets of sample scene information.

After the target neural network model is obtained in the above manner, the target neural network model may be applied to the target application program at an internal test stage of the target application program, or in other special cases, or after the target application program is formally released and brought online, and the actual scene information and the result feedback information corresponding to the actual scene information may be obtained by the target application program.

The target application uses the target neural network model, the result feedback information is feedback information of an actual recognition result output by the target neural network model, the actual recognition result is an estimated recognition result of the target object output by the target neural network model according to the actual scene information, in other words, the result feedback information is information fed back by a user according to actual scene information, and the actual recognition result is an estimated recognition result of prediction of the target neural network model.

Because the recognition result of the target neural network model may have an error, and the credibility of the real result feedback information fed back by the user is relatively high, the target neural network model in the target application program can be updated on line based on the actual scene information and the result feedback information. That is, the target neural network model is updated based on the real result feedback information fed back by the user, so that the recognition effect of the target neural network model can be more accurate, and the recognition accuracy is improved.

According to the embodiment, a target neural network model needs to be trained in advance, a pre-trained neural network model suitable for identifying a target object is obtained according to scene information, the pre-trained neural network model can output an estimated identification result of the target object according to input scene information, then model training is performed on the pre-trained neural network model by using a training sample set to obtain the target neural network model, wherein the training sample set comprises at least two groups of sample scene information, the at least two groups of sample scene information correspond to at least two groups of actual sample identification results in a one-to-one manner, different groups of sample scene information correspond to different actual sample identification results of the target object, corresponding weights of loss functions used in model training are different in different groups of sample scene information, and the weights corresponding to the different groups of sample scene information are in negative correlation with the number of sample scene information in the different groups of sample scene information, then, actual scene information and result feedback information corresponding to the actual scene information may be input in a target application program using a target neural network model, where the result feedback information is feedback information for an actual recognition result output by the target neural network model, and the actual recognition result is an estimated recognition result of the target object output by the target neural network model according to the actual scene information; and finally, updating the target neural network model in the target application program on line through the actual scene information and the result feedback information. By adopting the scheme, the corresponding weights of the loss functions used in the model training in different groups of sample scene information are set to be different, so that the different groups of sample scene information can be distributed in a balanced manner, the recognition effect on the different groups of sample scene information is improved, and the target neural network model in the target application program is updated through the real actual scene information and the result feedback information, so that the training samples can be enriched, the target neural network model can be more accurate, and the recognition accuracy of the target neural network model can be improved. In a small sample service scene, the weights of different types of samples of the small samples are adjusted through the loss function, so that the samples of different types of the small samples are uniformly distributed, and the effect of improving the classification and identification of the corresponding models in the small sample service scene is achieved.

Optionally, in this embodiment, the updating the target neural network model in the target application includes: and updating the target neural network model in the target application program according to a preset time interval.

Alternatively, the target neural network model in the target application may be updated at preset time intervals (e.g., every other day, every third day).

Through this embodiment, along with the continuous transform of user's demand, update target neural network model according to preset time interval, can make target neural network model follow closely customer's demand, not only improved user experience, make target neural network model more accurate moreover.

Optionally, in this embodiment, the acquiring actual scene information input in the target application and result feedback information corresponding to the actual scene information includes: acquiring a group of information pairs from an information pair set, wherein each information pair in the information pair set comprises actual scene information and result feedback information which are input in the target application program and have corresponding relations; acquiring a group of sample scene information from the training sample set, and acquiring a group of actual sample identification results which have one-to-one correspondence with the group of sample scene information; the updating the target neural network model in the target application program according to the actual scene information and the result feedback information includes: and updating the target neural network model in the target application program according to the group of information pairs, the group of sample scene information and the group of actual sample identification results.

Alternatively, the information pair set may be understood as a database for storing actual scene information and result feedback information. Each information pair in the information pair set comprises actual scene information and result feedback information which are input in the target application program and have corresponding relations.

And acquiring a group of information pairs from the information pair set, acquiring a group of sample scene information from the training sample set, and acquiring a group of actual sample identification results which have one-to-one correspondence with the group of sample scene information.

Then, a set of information pairs fed back by the user, the set of sample scene information, and the set of actual sample recognition results are combined to jointly update the target neural network model in the target application program.

Through the embodiment, the target neural network model is updated in a mode of combining the group of information pairs fed back by the user, the group of sample scene information and the group of actual sample identification results, so that the target neural network model can be updated based on the real-time feedback information of the user, the problem of poor model training effect caused by the small number of training samples in the traditional mode is solved, and in addition, the parameters of the target neural network model can be more accurate in the mode of updating based on the information fed back by the user, so that the target neural network model achieves better identification effect.

Optionally, in this embodiment, the obtaining a group of information pairs from the information pair set includes: acquiring a first group of information pairs from the information pair set; the acquiring of a set of sample scene information from the training sample set includes: acquiring a first group of sample scene information from the training sample set; wherein a first number of information pairs in the first set of information pairs is less than a second number of sample scene information in the first set of sample scene information; or, a ratio between the first number and a target total number is smaller than a predetermined ratio threshold, wherein the target total number is a sum of the first number and the second number.

Optionally, the information pair set may be information actually fed back by the user, and in an actual scene, there may be a situation of malicious feedback of the user, and to avoid this, when the first set of information pairs is obtained from the information pair set and the first set of sample scene information is obtained from the training sample set, the following manner may be adopted:

for example, a first number of information pairs in a first set of information pairs may be made smaller than a second number of sample scene information in a first set of sample scene information; alternatively, the first and second electrodes may be,

the ratio between the first number and a target total number, which is the sum of the first number and the second number, may be made smaller than a predetermined ratio threshold.

By the embodiment, if the result of the malicious feedback of the user exists in the information pair set, the updating effect of the target neural network model may be deteriorated, and by limiting the number or the proportion of the first group of information pairs fed back by the user, the adverse effect caused by the malicious feedback of the user can be reduced to a certain extent,

optionally, in this embodiment, as shown in fig. 3, an updating method of a target neural network model is provided in this embodiment, and the specific steps are as follows:

step S302, inputting the actual scene information to the target neural network model, and obtaining an estimated recognition result corresponding to the actual scene information and output by the target neural network model.

After the target application program is formally released and on-line, the actual scene information input by a user can be acquired in real time, and then the actual scene information can be input into the target neural network model, and the actual scene information is predicted and identified by the target neural network model to obtain a predicted identification result corresponding to the actual scene information.

Step S304, inputting the estimated recognition result corresponding to the actual scene information and the result feedback information into the loss function to obtain a first loss value, where the loss function includes a balance factor, and the balance factor is used to adjust weights corresponding to different pieces of scene information in the actual scene information.

And then inputting the estimated recognition result and real result feedback information fed back by the user into a loss function, and obtaining a first loss value through the loss function.

In the process of updating the target neural network model, if the number of different pieces of scene information in the actual scene information is different, the obtained first loss value may be larger, and at this time, the weights corresponding to the different pieces of scene information in the actual scene information may be adjusted by adjusting the balance factor in the loss function, so as to adjust the size of the first loss value.

Step S306, determining that the target neural network model is updated when the first loss value is less than or equal to a first preset threshold, so as to obtain an updated target neural network model.

And continuously adjusting the balance factor in the loss function until the first loss value is less than or equal to the first preset threshold value, which indicates that the target neural network model is completely updated, and at this time, obtaining the updated target neural network model.

By means of the embodiment, the mode of adjusting the balance factors in the loss function is adopted, the scene information with different quantities can be relatively balanced, the problem of poor recognition effect caused by unbalanced quantity of the scene information is avoided, and the recognition accuracy of the target neural network model is improved.

Optionally, in this embodiment, as shown in fig. 4, another method for updating a target neural network model is further provided in this embodiment, and the specific steps are as follows:

step S402, inputting the actual scene information in the set of information pairs into the target neural network model, and obtaining a first set of estimated recognition results output by the target neural network model and corresponding to the actual scene information in the set of information pairs.

Optionally, after the target application program is formally released and brought online, the actual scene information in the group of acquired information pairs is input to the target neural network model, the actual scene information is predicted through the target neural network model, and a first group of estimated recognition results corresponding to the actual scene information in the group of information pairs is output.

Step S404, inputting the set of sample scene information to the target neural network model, and obtaining a second set of pre-estimated recognition results output by the target neural network model and corresponding to the set of sample scene information.

Optionally, the acquired set of sample scene information is input to the target neural network model, the set of sample scene information is predicted by the target neural network model, and a second set of predicted identification results corresponding to the set of sample scene information is output.

Step S406, inputting the first group of estimated recognition results and the result feedback information in the group of information pairs into the loss function to obtain a second loss value, where the loss function includes a balance factor, and the balance factor is used to adjust weights corresponding to different pieces of scene information in the actual scene information.

Then, the first group of estimated recognition results and the result feedback information in the real group of information pairs fed back by the user are input into a loss function, and a second loss value can be obtained through the loss function.

In the process of updating the target neural network model, if the number of different pieces of scene information in the actual scene information is different, the obtained second loss value may be larger, and at this time, the weights corresponding to the different pieces of scene information in the actual scene information may be adjusted by adjusting the balance factor in the loss function, so as to adjust the size of the second loss value.

Step S408, inputting the second group of estimated recognition results and the group of actual sample recognition results into the loss function to obtain a third loss value.

And inputting the second group of estimated recognition results and a group of actual sample recognition results which have one-to-one correspondence with the group of sample scene information into a loss function, and obtaining a third loss value through the loss function.

And step S410, determining a fourth loss value corresponding to the target neural network model according to the second loss value and the third loss value.

And then, adding the second loss value and the third loss value to obtain a fourth loss value corresponding to the target neural network model, or weighting and summing the second loss value and the third loss value to obtain a fourth loss value corresponding to the target neural network model.

Step S412, determining that the target neural network model is updated when the fourth loss value is less than or equal to a second preset threshold, so as to obtain an updated target neural network model.

And continuously adjusting the balance factor in the loss function until the fourth loss value is less than or equal to the second preset threshold value, which indicates that the target neural network model is completely updated, and at this time, obtaining the updated target neural network model.

Optionally, as shown in fig. 5, in this embodiment, a method for training a target neural network model is provided, which includes the following specific steps:

step S502, inputting the at least two groups of sample scene information into the pre-training neural network model to obtain at least two groups of estimated sample identification results which are output by the pre-training neural network model and correspond to the at least two groups of sample scene information.

Optionally, at least two sets of sample scene information are input into a pre-trained neural network model, the at least two sets of sample scene information are predicted through the pre-trained neural network model, and at least two sets of predicted sample identification results corresponding to the at least two sets of sample scene information are output.

Step S504, inputting the at least two groups of estimated sample identification results and the at least two groups of actual sample identification results into the loss function to obtain a fifth loss value, where the loss function includes a balance factor, and the balance factor is used to adjust weights corresponding to the at least two groups of sample scene information.

Optionally, the at least two groups of estimated sample identification results and the at least two groups of actual sample identification results having a one-to-one correspondence with the at least two groups of sample scene information are input into a loss function, and a fifth loss value can be obtained through the loss function.

In the process of updating the target neural network model, if the number of different scene information in the at least two groups of sample scene information is different, the obtained fifth loss value may be larger, and at this time, the weights corresponding to the different scene information in the at least two groups of sample scene information may be adjusted by adjusting the balance factor in the loss function, so as to adjust the size of the fifth loss value.

Step S506, determining that the training of the pre-trained neural network model is completed to obtain the target neural network model when the fifth loss value is less than or equal to a third preset threshold value.

And continuously adjusting the balance factor in the loss function until the fifth loss value is less than or equal to the third preset threshold value, which indicates that the training of the target neural network model is completed, and at this time, obtaining the target neural network model.

Through the embodiment, the mode of adjusting the balance factors in the loss function is adopted, so that at least two groups of sample scene information with different quantities can be relatively balanced, the problem of poor recognition effect caused by unbalanced quantity of the scene information is avoided, and the recognition precision of the target neural network model is improved.

Optionally, in this embodiment, the at least two groups of estimated sample identification results and the at least two groups of actual sample identification results are combinedObtaining a fifth loss value if the input is to the loss function, comprising: determining the fifth loss value according to the following formula:

wherein the Loss is the fifth Loss value, the number of the at least two groups of estimated sample identification results and the number of the at least two groups of actual sample identification results are both N, N is a natural number, and p is a natural number_kQ is one of the at least two groups of actual sample recognition results_kThe gamma is the balance factor for one of the at least two sets of predicted sample identification results.

Optionally, assuming that the number of the at least two groups of predicted sample identification results and the number of the at least two groups of actual sample identification results are N, where N is a natural number, the at least two groups of predicted sample identification results and the at least two groups of actual sample identification results may be input into the following formula to obtain the fifth loss value:

wherein the Loss is the fifth Loss value, and p is the maximum Loss value_kQ is one of the at least two groups of actual sample recognition results_kThe gamma is the balance factor for one of the at least two sets of predicted sample identification results.

It is understood that, for the calculation manners of the first loss value, the second loss value and the third loss value, the calculation manner of the fifth loss value can be referred to, and details are not repeated herein.

Through the embodiment, the formula is adopted, the fifth loss value can be accurately calculated, whether the pre-training neural network model is trained or not is evaluated through the fifth loss value, and the training accuracy of the pre-training neural network model is improved.

Optionally, in this embodiment, in a case that the fifth loss value is greater than a third preset threshold, the method further includes: increasing said balance factor in said loss function; and training the pre-trained neural network model by using the training sample set until the fifth loss value is less than or equal to the third preset threshold value, and determining that the pre-trained neural network model is trained to obtain the target neural network model.

Optionally, if the obtained fifth loss value is greater than a third preset threshold, a mode of a balance factor in the loss function may be increased, and the pre-trained neural network model is trained using the training sample set until the fifth loss value is less than or equal to the third preset threshold, it is determined that the training of the pre-trained neural network model is completed, so as to obtain the target neural network model.

Optionally, in this embodiment, before the model training of the pre-trained neural network model by using the training sample set to obtain the target neural network model, the method further includes: determining the size of the short edge of each image in the image set to be processed as a first size to obtain a first image set; cutting out each image in the first image set by taking the height as a second size and the width as a third size to obtain a second image set, and determining the second image set as the training sample set; or turning over the images with the first probability in the second image set to obtain a third image set, and determining the third image set as the training sample set; or rotating the images with the second probability in the third image set within a preset angle range to obtain a fourth image set, and determining the fourth image set as the training sample set.

Optionally, before starting model training on the pre-trained neural network model, a training sample set needs to be obtained, for example, by preprocessing the image set to be processed.

It should be noted that each to-be-processed image in the to-be-processed image set may be an RGB picture.

Specifically, the size of the short side of each image in the set of images to be processed is determined to be a first size (e.g., 256), resulting in a first set of images.

Then, each image in the first image set is cut out with a height of a second size (e.g. 224) and a width of a third size (e.g. 224) to obtain a second image set, and the second image set is determined as the training sample set. The intercepting mode is random intercepting, namely any area in each image can be intercepted, and only the height is ensured to be the second size, and the width is ensured to be the third size.

Alternatively, the training sample set may be obtained as follows:

and inverting the images with the first probability (such as 0.5, 0.6 and the like) in the second image set to obtain a third image set, and determining the third image set as the training sample set. For example, half of the images in the second image set are arbitrarily selected to be subjected to front-back inversion to obtain a third image set, and the third image set is determined as the training sample set.

Alternatively, the training sample set may be obtained as follows:

and rotating the images with the second probability (such as 0.5, 0.6 and the like) in the third image set within a preset angle range to obtain a fourth image set, and determining the fourth image set as the training sample set. For example, arbitrarily selecting half of the images in the third image set to rotate within [ -45, 45] degrees to obtain a fourth image set, and determining the fourth image set as the training sample set.

In a possible embodiment, as shown in fig. 6, the specific steps of the method for preprocessing the image set to be processed are as follows:

in step S601, an RGB picture is input.

Step S602, standardizing the size of the short edge of each RGB picture in the image set to be processed to 256.

Step S603, randomly intercepting each RGB picture in the normalized to-be-processed image set with a size of 224 × 224.

In step S604, the picture with the size of 224 × 224 is randomly flipped with a probability of 0.5.

In step S605, the reversed picture with the size of 224 × 224 is randomly rotated by [ -45, 45] degrees with a probability of 0.5.

The pictures that have undergone the above steps S601-S605 are determined as the above training sample set.

It is understood that the above is only an example, and the present embodiment is not limited thereto.

Through the embodiment, the image set to be processed can be expanded by carrying out standardization, overturning operation and rotation operation on the image in the image set to be processed, so that the obtained training sample set is richer, and the sample diversity of the training sample set is improved.

In the above manner, after the target application is formally released online, the target application can be used to identify an image, and in the embodiment of the present invention, an image identification method based on the target application is provided, as shown in fig. 7, the specific steps are as follows:

step S702, inputting image information to a target application program, wherein the target application program uses a target neural network model;

step S704, recognizing the image information through the target neural network model to obtain an image recognition result, where the image recognition result is an estimated recognition result of the target object output by the target neural network model according to the image information;

wherein the target neural network model is obtained by performing model training on a pre-trained neural network model by using a training sample set, wherein, the pre-training neural network model is used for outputting the pre-estimation recognition result of the target object according to the input scene information, the training sample set is a set obtained by preprocessing an image set to be processed, the training sample set comprises at least two groups of sample scene information, the at least two groups of sample scene information and the at least two groups of actual sample identification results have a one-to-one correspondence relationship, different groups of sample scene information correspond to different actual sample identification results of the target object, the loss function used in the model training has different corresponding weights in different sets of sample scene information, the weights corresponding to the different sets of sample scene information are inversely related to the number of sample scene information in the different sets of sample scene information.

Optionally, the target neural network model needs to be trained in advance, and for a specific training mode of the target neural network model, the detailed description is given in the above embodiments, and is not repeated here.

After the target neural network model is trained, the target neural network model can be used in the target application program, a user can use the target application program to input image information to the target application program, and the target neural network model in the target application program identifies the image information to obtain a corresponding image identification result.

For example, taking the target application as a program for identifying a tray, the user inputs an image without a tray to the target application, and can obtain an image identification result of "no tray is identified", the user inputs an image without a tray being empty to the target application, and can obtain an image identification result of "tray being empty", and the user inputs an image without a tray being empty to the target application, and can obtain an image identification result of "tray being not empty".

Optionally, the user may feed back a user feedback result corresponding to the image information through the target application program according to the accuracy of the image recognition result, and then the target application program may use the image information and the corresponding user feedback result as a sample for updating the target application program, so as to improve the recognition rate of the target application program.

According to the embodiment, the target neural network model trained in advance is used in the target application program, and the target neural network model can be updated in real time according to user feedback, so that the identification precision of the target neural network model is more and more accurate, the identification precision page of the image information by the target application program is more accurate, and the identification accuracy of the target application program is improved.

It should be noted that, the current deep learning scheme is generally deployed after offline training, that is, a labeled data training model is used, and after formal deployment and online deployment, parameters of the model are fixed.

The above scheme has the following disadvantages:

1. the situation that the sample data amount is small and the sample distribution is not uniform is poor in performance.

2. After formal deployment is carried out, the parameters of the model are fixed, so that errors in the operation of the deep learning model cannot be sensed.

In order to solve the disadvantages of the above solutions, the following describes a flow of an update method of a target application based on feedback information with reference to an alternative example.

Optionally, the general idea of the embodiment of the present application is to adopt a transfer learning scheme on existing data, use a pre-trained neural network model as an initial parameter, dynamically adjust sample weights for long tail distribution, train the pre-trained neural network model, and obtain a target neural network model; after the target neural network model is formally deployed on line, an interface fed back by a user is provided, feedback information (such as the actual scene information and the result feedback information) of the user is collected through the interface to serve as a label, data fed back by the user is used as training data, the target neural network model is trained and updated on line in the background, and the recognition effect of the target neural network model is improved. By adopting the technical scheme of the invention, the problems of insufficient data and long tail distribution in a deep learning scene can be alleviated by effectively utilizing the migration learning and the online feedback of the user.

Taking the application scenario as the optical disc identification in the AI commonweal applet as an example, the convolutional neural network CNN is used to discriminate the current picture input by the user into the AI commonweal applet, which is a three-classification problem, for example, whether the disc in the current picture is an optical disc or not, the specific discrimination result is: the target category is 1, and the current picture is an empty plate; the target category is 2, and the plate in the current picture is not empty; the target category is 3 and there is no plate in the current picture.

As shown in fig. 7, the user uploads the RGB image, the server inputs the RGB image uploaded by the user to the CNN, and the CNN returns the recognition result and feeds back the recognition result to the user through the server.

The problem of the disc recognition in the AI commonweal applet is that the existing training data is very deficient, the number of samples per category is less than 500, and the scene is very limited, most of the discs are photos taken by some specific canteens, the disc types are single, and the generalization of the algorithm model is seriously affected. In the embodiment, the generalization capability of the model is improved and the effect of the model is obviously improved by introducing the methods of transfer learning and user feedback.

The method for updating the target application program based on the feedback information mainly comprises the following steps:

step one, using a pre-training neural network model.

In consideration of the limitation of data distribution and the shortage of the number of samples, the generalization capability of the model obtained by direct training is far from insufficient, a pre-trained neural network model on ImageNet can be used as an initialization model, the generalization capability of the algorithm can be greatly improved by using the pre-trained neural network model, and a relatively ideal effect can be obtained only by few fine tuning times.

And step two, preprocessing the data.

Alternatively, the data preprocessing may refer to the method shown in fig. 6, which is not described herein.

And step three, weighting the self-adaptive cross entropy function.

Optionally, in the training samples, the distribution among different classes is extremely unbalanced, a few classes contain most pictures in the sample set, and a weight adaptive cross entropy function can be set to alleviate the problem of imbalance among the classes.

Wherein p in the above formula is a correct label, q is a predicted value, and q can be understood as a result processed by the softmax function. By increasing the loss weight of the error sample, the model is more concentrated on the samples which are difficult to classify during training, so that the phenomenon of sample imbalance is relieved.

And step four, updating the model on line.

1. User feedback information is collected.

After the optical disc identification software in the AI public welfare applet is formally deployed online, as shown in fig. 8, a feedback window may be provided for the user, if the user considers that the current optical disc identification software is classified incorrectly, the user may feed back the error and simultaneously provide a correct label, and meanwhile, the optical disc identification software background may collect the picture and the label provided by the user, and store the picture label pair in Images Pool (e.g., ImageNet database).

2. And (4) performing on-line training.

At regular intervals, the target neural network model extracts pictures from the original training set and Images Pool to train the target neural network model, and updates parameters of the target neural network model. However, as shown in fig. 9, considering that the user may feedback maliciously, that is, feedback a large number of wrong labels, thereby causing online training errors, as shown in fig. 10, the influence of the malicious feedback labels of the user may be reduced by restricting the proportion of samples in the Images Pool in one batch (a group of samples) during training. The loss used in the training is the normal cross entropy loss, and the model is generally trained on line every 1 day.

Through this embodiment, firstly, can use the user feedback interface to collect user's feedback information, regard user feedback information as exact label, online update model, reduce the reappearance of this type of mistake, secondly, utilize the combination of user feedback and migration study, can effectively alleviate the not enough problem of training data, promote the generalization ability of model, moreover, through the dynamic adjustment of sample weight, can effectively promote the effect of algorithm on long tail data (the sample is unbalanced), extremely low the recognition effect who has improved the model.

It should be noted that, for simplicity of description, the above-mentioned method embodiments are described as a series of acts or combination of acts, but those skilled in the art will recognize that the present invention is not limited by the order of acts, as some steps may occur in other orders or concurrently in accordance with the invention. Further, those skilled in the art should also appreciate that the embodiments described in the specification are preferred embodiments and that the acts and modules referred to are not necessarily required by the invention.

According to another aspect of the embodiments of the present invention, there is also provided an apparatus for updating a target application based on feedback information, as shown in fig. 11, the apparatus including:

a first obtaining unit 1002, configured to obtain actual scene information input in a target application program and result feedback information corresponding to the actual scene information, where the target application program uses a target neural network model, the result feedback information is feedback information of an actual recognition result output by the target neural network model, and the actual recognition result is an estimated recognition result of the target object output by the target neural network model according to the actual scene information;

a first processing unit 1004, configured to update the target neural network model in the target application according to the actual scene information and the result feedback information;

As an optional technical solution, the first obtaining unit is further configured to obtain a set of information pairs from an information pair set, where each information pair in the information pair set includes actual scene information and result feedback information having a corresponding relationship, which are input in the target application program; acquiring a group of sample scene information from the training sample set, and acquiring a group of actual sample identification results which have one-to-one correspondence with the group of sample scene information; the first processing unit is further configured to update the target neural network model in the target application program according to the set of information pairs, the set of sample scene information, and the set of actual sample identification results.

As an optional technical solution, the first obtaining unit is further configured to obtain a first group of information pairs from the information pair set; the first obtaining unit is further configured to obtain a first set of sample scene information from the training sample set; wherein a first number of information pairs in the first set of information pairs is less than a second number of sample scene information in the first set of sample scene information; or, a ratio between the first number and a target total number is smaller than a predetermined ratio threshold, wherein the target total number is a sum of the first number and the second number.

As an optional technical solution, the first processing unit: the method comprises the following steps: the first processing module is used for inputting the actual scene information to the target neural network model to obtain an estimated recognition result which is output by the target neural network model and corresponds to the actual scene information; a second processing module, configured to input the pre-estimated recognition result corresponding to the actual scene information and the result feedback information into the loss function to obtain a first loss value, where the loss function includes a balance factor, and the balance factor is used to adjust weights corresponding to different pieces of scene information in the actual scene information; and a first determining module, configured to determine that the target neural network model is updated completely to obtain the updated target neural network model when the first loss value is less than or equal to a first preset threshold.

As an optional technical solution, the first processing unit includes: a third processing module, configured to input the actual scene information in the group of information pairs to the target neural network model, so as to obtain a first group of pre-estimated recognition results output by the target neural network model and corresponding to the actual scene information in the group of information pairs; the fourth processing module is used for inputting the group of sample scene information to the target neural network model to obtain a second group of estimated recognition results which are output by the target neural network model and correspond to the group of sample scene information; a fifth processing module, configured to input the first group of pre-estimated recognition results and the result feedback information in the group of information pairs into the loss function to obtain a second loss value, where the loss function includes a balance factor, and the balance factor is used to adjust weights corresponding to different pieces of scene information in the actual scene information; a sixth processing module, configured to input the second group of estimated recognition results and the group of actual sample recognition results into the loss function to obtain a third loss value; a second determining module, configured to determine a fourth loss value corresponding to the target neural network model according to the second loss value and the third loss value; and a third determining module, configured to determine that the target neural network model is updated completely to obtain the updated target neural network model when the fourth loss value is less than or equal to a second preset threshold.

As an optional technical solution, the apparatus further includes: a second obtaining unit, configured to obtain the pre-trained neural network model, where the pre-trained neural network model is used to output an estimated recognition result of a target object according to input scene information; a third processing unit, configured to perform model training on the pre-trained neural network model by using the training sample set to obtain the target neural network model

As an optional technical solution, the first processing unit includes: a seventh processing module, configured to input the at least two sets of sample scene information to the pre-trained neural network model, so as to obtain at least two sets of estimated sample recognition results, which are output by the pre-trained neural network model and correspond to the at least two sets of sample scene information; an eighth processing module, configured to input the at least two groups of estimated sample identification results and the at least two groups of actual sample identification results into the loss function to obtain a fifth loss value, where the loss function includes a balance factor, and the balance factor is used to adjust weights corresponding to the at least two groups of sample scene information; and a fourth determining module, configured to determine that the pre-trained neural network model is trained completely to obtain the target neural network model when the fifth loss value is less than or equal to a third preset threshold.

As an optional technical solution, the eighth processing module is further configured to determine the fifth loss value according to the following formula:

As an optional technical solution, the apparatus further includes: a ninth processing module, configured to increase the balance factor in the loss function when the fifth loss value is greater than a third preset threshold; and training the pre-trained neural network model by using the training sample set until the fifth loss value is less than or equal to the third preset threshold value, and determining that the pre-trained neural network model is trained to obtain the target neural network model.

As an optional technical solution, the apparatus further includes: the image processing device comprises a first determining unit, a second determining unit and a processing unit, wherein the first determining unit is used for determining the size of the short side of each image in an image set to be processed as a first size to obtain a first image set; a second determining unit, configured to cut out each image in the first image set with a height as a second size and a width as a third size to obtain a second image set, and determine the second image set as the training sample set; or, a third determining unit, configured to flip the images with the first probability in the second image set to obtain a third image set, and determine the third image set as the training sample set; or, the fourth determining unit is configured to rotate the image with the second probability in the third image set by a preset angle range to obtain a fourth image set, and determine the fourth image set as the training sample set.

As an optional technical solution, the first processing unit is further configured to update the target neural network model in the target application program according to a preset time interval.

According to another aspect of the embodiments of the present invention, there is also provided an apparatus for updating a target application based on feedback information, as shown in fig. 12, the apparatus including:

an input unit 1102 configured to input image information to a target application program that uses a target neural network model;

a second processing unit 1104, configured to recognize the image information through the target neural network model to obtain an image recognition result, where the image recognition result is an estimated recognition result of a target object output by the target neural network model according to the image information;

According to a further aspect of embodiments of the present invention, there is also provided a storage medium having a computer program stored therein, wherein the computer program is arranged to perform the steps of any of the above-mentioned method embodiments when executed.

Alternatively, in the present embodiment, the storage medium may be configured to store a computer program for executing the steps of:

s1, inputting image information to a target application program, wherein the target application program uses a target neural network model;

s2, recognizing the image information through the target neural network model to obtain an image recognition result, wherein the image recognition result is an estimated recognition result of the target object output by the target neural network model according to the image information;

Alternatively, in this embodiment, a person skilled in the art may understand that all or part of the steps in the methods of the foregoing embodiments may be implemented by instructing hardware related to the terminal device through a program, where the program may be stored in a computer-readable storage medium, and the storage medium may include: flash disks, ROM (Read-Only Memory), RAM (Random Access Memory), magnetic or optical disks, and the like.

According to another aspect of the embodiment of the present invention, there is also provided an electronic device for implementing the method for updating a target application based on feedback information, where the electronic device may be a terminal device or a server shown in fig. 1. The present embodiment takes the electronic device as a server as an example for explanation. As shown in fig. 13, the electronic device comprises a memory 1202 and a processor 1204, the memory 1202 having stored therein a computer program, the processor 1204 being arranged to perform the steps of any of the above-described method embodiments by means of the computer program.

Optionally, in this embodiment, the processor may be configured to execute the following steps by a computer program:

Alternatively, it is understood by those skilled in the art that the structure shown in fig. 13 is only an illustration and is not a limitation to the structure of the electronic device. For example, the electronic device may also include more or fewer components (e.g., network interfaces, etc.) than shown in FIG. 13, or have a different configuration than shown in FIG. 13.

The memory 1202 may be used to store software programs and modules, such as program commands/modules corresponding to the method and apparatus for updating a target application based on feedback information in the embodiment of the present invention, and the processor 1204 executes various functional applications and data processing by running the software programs and modules stored in the memory 1202, that is, implements the method for updating a target application based on feedback information. The memory 1202 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 1202 can further include memory located remotely from the processor 1204, which can be connected to a terminal over a network. Examples of such networks include, but are not limited to, the internet, intranets, local area networks, mobile communication networks, and combinations thereof. As an example, as shown in fig. 13, the memory 1202 may include, but is not limited to, a first obtaining unit 1002 and a first processing unit 1004 in the update apparatus of the target application based on the feedback information, or the memory 1202 may include, but is not limited to, an input unit 1102 and a second processing unit 1104 in the image recognition apparatus of the target application based on the feedback information. In addition, the image recognition device may further include, but is not limited to, other module units in the update device of the target application based on the feedback information, or may further include, but is not limited to, other module units in the image recognition device based on the target application, which is not described in detail in this example.

Optionally, the transmitting device 1206 is configured to receive or transmit data via a network. Examples of the network may include a wired network and a wireless network. In one example, the transmitting device 1206 includes a Network adapter (NIC) that can be connected to a router via a Network cable to communicate with the internet or a local area Network. In one example, the transmitting device 1206 is a Radio Frequency (RF) module, which is used to communicate with the internet in a wireless manner.

In addition, the electronic device further includes: a connection bus 1208 for connecting the various module components in the electronic device.

In other embodiments, the terminal device or the server may be a node in a distributed system, where the distributed system may be a blockchain system, and the blockchain system may be a distributed system formed by connecting a plurality of nodes through a network communication. Nodes can form a Peer-To-Peer (P2P, Peer To Peer) network, and any type of computing device, such as a server, a terminal, and other electronic devices, can become a node in the blockchain system by joining the Peer-To-Peer network.

Alternatively, in this embodiment, a person skilled in the art may understand that all or part of the steps in the methods of the foregoing embodiments may be implemented by instructing hardware related to the terminal device through a program, where the program may be stored in a computer-readable storage medium, and the storage medium may include: flash disks, Read-Only memories (ROMs), Random Access Memories (RAMs), magnetic or optical disks, and the like.

The above-mentioned serial numbers of the embodiments of the present invention are merely for description and do not represent the merits of the embodiments.

The integrated unit in the above embodiments, if implemented in the form of a software functional unit and sold or used as a separate product, may be stored in the above computer-readable storage medium. Based on such understanding, the technical solution of the present invention may be embodied in the form of a software product, which is stored in a storage medium and includes a plurality of commands for enabling one or more computer devices (which may be personal computers, servers, network devices, etc.) to execute all or part of the steps of the above methods according to the embodiments of the present invention.

Optionally, in this embodiment, the computer software product or the software product may perform the following steps:

In the above embodiments of the present invention, the descriptions of the respective embodiments have respective emphasis, and for parts that are not described in detail in a certain embodiment, reference may be made to related descriptions of other embodiments.

In the several embodiments provided in the present application, it should be understood that the disclosed client may be implemented in other manners. The above-described embodiments of the apparatus are merely illustrative, and for example, the above-described division of the units is only one type of division of logical functions, and other divisions may be realized in practice, for example, a plurality of units or components may be combined or integrated into another system, or some features may be omitted, or not executed. In addition, the shown or discussed mutual coupling or direct coupling or communication connection may be an indirect coupling or communication connection through some interfaces, units or modules, and may be in an electrical or other form.

The units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of the embodiment.

In addition, functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may exist alone physically, or two or more units are integrated into one unit. The integrated unit can be realized in a form of hardware, and can also be realized in a form of a software functional unit.

The foregoing is only a preferred embodiment of the present invention, and it should be noted that, for those skilled in the art, various modifications and decorations can be made without departing from the principle of the present invention, and these modifications and decorations should also be regarded as the protection scope of the present invention.

Claims

1. A method for updating a target application program based on feedback information is characterized by comprising the following steps:

acquiring actual scene information input in a target application program and result feedback information corresponding to the actual scene information, wherein the target application program uses a target neural network model, the result feedback information is feedback information of an actual recognition result output by the target neural network model, and the actual recognition result is an estimated recognition result of the target object output by the target neural network model according to the actual scene information;

updating the target neural network model in the target application program according to the actual scene information and the result feedback information;

the target neural network model is obtained by performing model training on a pre-trained neural network model by using a training sample set, wherein the pre-trained neural network model is used for outputting an estimated recognition result of a target object according to input scene information, the training sample set comprises at least two sets of sample scene information, the at least two sets of sample scene information and at least two sets of actual sample recognition results have a one-to-one correspondence relationship, different sets of sample scene information correspond to different actual sample recognition results of the target object, corresponding weights of loss functions used in model training in different sets of sample scene information are different, and the weights corresponding to different sets of sample scene information are in negative correlation with the number of sample scene information in different sets of sample scene information.

2. The method of claim 1,

the acquiring actual scene information input in the target application program and result feedback information corresponding to the actual scene information includes: acquiring a group of information pairs from an information pair set, wherein each information pair in the information pair set comprises actual scene information and result feedback information which are input in the target application program and have a corresponding relation; acquiring a group of sample scene information from the training sample set, and acquiring a group of actual sample identification results which have one-to-one correspondence with the group of sample scene information;

the updating the target neural network model in the target application program according to the actual scene information and the result feedback information includes: updating the target neural network model in the target application program according to the set of information pairs, the set of sample scene information and the set of actual sample identification results.

3. The method of claim 2,

the obtaining a set of information pairs from the set of information pairs comprises: obtaining a first group of information pairs from the set of information pairs;

the obtaining a set of sample scene information from the training sample set includes: acquiring a first group of sample scene information from the training sample set;

wherein a first number of information pairs in the first set of information pairs is less than a second number of sample scene information in the first set of sample scene information; or, a ratio between the first number and a target total number is smaller than a predetermined ratio threshold, wherein the target total number is a sum of the first number and the second number.

4. The method of claim 1, wherein the updating the target neural network model in the target application according to the actual context information and the result feedback information comprises:

inputting the actual scene information into the target neural network model to obtain a pre-estimated recognition result which is output by the target neural network model and corresponds to the actual scene information;

inputting the estimated identification result corresponding to the actual scene information and the result feedback information into the loss function to obtain a first loss value, wherein the loss function comprises a balance factor, and the balance factor is used for adjusting weights corresponding to different scene information in the actual scene information;

and under the condition that the first loss value is smaller than or equal to a first preset threshold value, determining that the target neural network model is updated, and obtaining the updated target neural network model.

5. The method of claim 2, wherein the updating the target neural network model in the target application according to the set of information pairs, the set of sample scene information, and the set of actual sample recognition results comprises:

inputting the actual scene information in the group of information pairs into the target neural network model to obtain a first group of estimated recognition results output by the target neural network model and corresponding to the actual scene information in the group of information pairs;

inputting the group of sample scene information into the target neural network model to obtain a second group of estimated recognition results which are output by the target neural network model and correspond to the group of sample scene information;

inputting the first group of pre-estimated recognition results and the result feedback information in the group of information pairs into the loss function to obtain a second loss value, wherein the loss function comprises a balance factor, and the balance factor is used for adjusting weights corresponding to different scene information in the actual scene information;

inputting the second group of estimated recognition results and the group of actual sample recognition results into the loss function to obtain a third loss value;

determining a fourth loss value corresponding to the target neural network model according to the second loss value and the third loss value;

and under the condition that the fourth loss value is smaller than or equal to a second preset threshold value, determining that the target neural network model is updated, and obtaining the updated target neural network model.

6. The method of claim 1, further comprising:

acquiring the pre-training neural network model, wherein the pre-training neural network model is used for outputting a pre-estimated recognition result of a target object according to input scene information;

and performing model training on the pre-trained neural network model by using the training sample set to obtain the target neural network model.

7. The method of claim 6, wherein model training the pre-trained neural network model using a set of training samples to obtain a target neural network model comprises:

inputting the at least two groups of sample scene information into the pre-training neural network model to obtain at least two groups of estimated sample identification results which are output by the pre-training neural network model and correspond to the at least two groups of sample scene information;

inputting the at least two groups of estimated sample identification results and the at least two groups of actual sample identification results into the loss function to obtain a fifth loss value, wherein the loss function comprises a balance factor, and the balance factor is used for adjusting weights corresponding to the at least two groups of sample scene information;

and under the condition that the fifth loss value is smaller than or equal to a third preset threshold value, determining that the training of the pre-trained neural network model is completed, and obtaining the target neural network model.

8. The method of claim 7, wherein inputting the at least two sets of predicted sample identifications and the at least two sets of actual sample identifications to the loss function to obtain a fifth loss value comprises:

determining the fifth loss value according to the following equation:

wherein the Loss is the fifth Loss value, the number of the at least two groups of estimated sample identification results and the number of the at least two groups of actual sample identification results are both N, N is a natural number, and p is_kFor one actual sample recognition result in the at least two groups of actual sample recognition results, the q_kAnd determining a balance factor of the at least two groups of estimated sample identification results, wherein gamma is the balance factor.

9. The method of claim 7, wherein in case the fifth loss value is greater than a third preset threshold, the method further comprises:

increasing the balance factor in the loss function;

and training the pre-trained neural network model by using the training sample set until the fifth loss value is less than or equal to the third preset threshold value, and determining that the pre-trained neural network model is trained to obtain the target neural network model.

10. The method of any one of claims 1 to 9, wherein prior to the model training of the pre-trained neural network model using a set of training samples to obtain a target neural network model, the method further comprises:

determining the size of the short edge of each image in the image set to be processed as a first size to obtain a first image set;

intercepting each image in the first image set by taking the height as a second size and the width as a third size to obtain a second image set, and determining the second image set as the training sample set; or

Turning over the images with the first probability in the second image set to obtain a third image set, and determining the third image set as the training sample set; or

And rotating the images with the second probability in the third image set within a preset angle range to obtain a fourth image set, and determining the fourth image set as the training sample set.

11. The method of any one of claims 1 to 9, wherein the updating the target neural network model in the target application comprises:

and updating the target neural network model in the target application program according to a preset time interval.

12. An image recognition method based on a target application program is characterized by comprising the following steps:

inputting image information to a target application, wherein the target application uses a target neural network model;

identifying the image information through the target neural network model to obtain an image identification result, wherein the image identification result is an estimated identification result of a target object output by the target neural network model according to the image information;

wherein the target neural network model is a network model obtained by model training of a pre-trained neural network model by using a training sample set, wherein the pre-training neural network model is used for outputting the pre-estimated recognition result of the target object according to the input scene information, the training sample set is a set obtained by preprocessing an image set to be processed, the training sample set comprises at least two groups of sample scene information, the at least two groups of sample scene information and the at least two groups of actual sample identification results have one-to-one correspondence, different groups of sample scene information correspond to different actual sample identification results of the target object, the loss function used in the model training has different corresponding weights in different sets of sample scene information, the weights corresponding to the different sets of sample scene information are inversely related to the number of sample scene information in the different sets of sample scene information.

13. An apparatus for updating a target application based on feedback information, comprising:

a first obtaining unit, configured to obtain actual scene information input in the target application program and result feedback information corresponding to the actual scene information, where the target application program uses a target neural network model, the result feedback information is feedback information of an actual recognition result output by the target neural network model, and the actual recognition result is an estimated recognition result of the target object output by the target neural network model according to the actual scene information;

the first processing unit is used for updating the target neural network model in the target application program according to the actual scene information and the result feedback information;

14. A computer-readable storage medium comprising a stored program, wherein the program when executed performs the method of any of claims 1 to 11, or claim 12.

15. An electronic device comprising a memory and a processor, wherein the memory has stored therein a computer program and the processor is arranged to execute the method of any of claims 1 to 11, or claim 12, by means of the computer program.