CN114332500A

CN114332500A - Image processing model training method and device, computer equipment and storage medium

Info

Publication number: CN114332500A
Application number: CN202111073887.3A
Authority: CN
Inventors: 何必畅; 胡易; 鄢科; 黄飞跃
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2021-09-14
Filing date: 2021-09-14
Publication date: 2022-04-12

Abstract

The embodiment of the application provides an image processing model training method, an image processing model training device, computer equipment and a storage medium, wherein the method comprises the following steps: acquiring an original training data set and an incremental training data set; clustering a plurality of training data included in an original training data set and a plurality of training data included in an incremental training data set to obtain N cluster sets; each cluster set comprises one or more training data, and the one or more training data of each cluster set belong to the same data category; carrying out sampling compression processing on training data included in the N clustering sets to obtain a sampling training data set; training the pre-trained data processing model by using a sampling training data set to obtain a target data processing model; the pre-trained data processing model is obtained by training based on the original training data set, so that the quantity of training data required by the pre-trained model during retraining can be reduced, and the accuracy of the model after retraining can be improved.

Description

Image processing model training method and device, computer equipment and storage medium

Technical Field

The present application relates to the field of computer technologies, and in particular, to a method and an apparatus for training an image processing model, a computer device, and a storage medium.

Background

With the development of computer technology, artificial intelligence is widely concerned worldwide due to the development of deep learning technology, and network models obtained based on deep learning technology training are gradually applied to various industries, for example, a voice recognition model can be applied to an intelligent sound box and an intelligent vehicle, an image recognition model can be applied to monitoring equipment, a text recognition model can be applied to an intelligent customer service product, and the like. When any network model is used for recognition processing, the network model is required to be pre-trained by some original training data to obtain a pre-trained model, and then the pre-trained model can be further re-trained according to a specific application scene.

In specific implementation, some training data can be newly added on the basis of the original training data by combining with a specific application scene, and then the pre-trained model is retrained again by using the original training data and the newly added training data. While this ensures that the pre-trained model learns the information in the new training data and the original training data well, as the training data continues to increase, the resources and time required for training increase proportionally. Therefore, how to train the pre-training model again becomes a research hotspot.

Disclosure of Invention

The embodiment of the application provides an image processing model training method and device, computer equipment and a storage medium, which can reduce the quantity of training data required by retraining a pre-trained model and improve the accuracy of the retrained model.

In one aspect, an embodiment of the present application provides an image processing model training method, where the method includes:

acquiring an original training data set and an incremental training data set;

clustering a plurality of training data included in the original training data set and a plurality of training data included in the incremental training data set to obtain N cluster sets; each cluster set comprises one or more training data, and the one or more training data of each cluster set belong to the same data category;

sampling and compressing the training data included in the N cluster sets to obtain a sampling training data set;

training the pre-trained data processing model by using the sampling training data set to obtain a target data processing model; the pre-trained data processing model is trained based on the original training data set.

In one aspect, an embodiment of the present application provides an image processing model training apparatus, where the apparatus includes:

the acquisition unit is used for acquiring an original training data set and an incremental training data set;

the processing unit is used for clustering a plurality of training data included in the original training data set and a plurality of training data included in the incremental training data set to obtain N clustering sets; each cluster set comprises one or more training data, and the one or more training data of each cluster set belong to the same data category;

the processing unit is further configured to perform sampling compression processing on the training data included in the N cluster sets to obtain a sampled training data set;

the processing unit is further used for training the pre-trained data processing model by using the sampling training data set to obtain a target data processing model; the pre-trained data processing model is trained based on the original training data set.

In one aspect, the present application provides a computer device, which includes an input device, an output device, a processor, and a computer storage medium, where the processor and the computer storage medium are connected to each other, where the computer storage medium is used to store a computer program, and the processor is configured to invoke the computer program to execute the above-mentioned image processing model training method.

In one aspect, the present application provides a computer-readable storage medium, in which a computer program is stored, and when the computer program is executed by a processor, the computer program is used to implement the image processing model training method described above.

In one aspect, embodiments of the present application provide a computer program product or a computer program, where the computer program product includes a computer program; the computer program is stored in a computer readable storage medium, and when executed by a processor of a computer device, performs the image processing model training method described above.

In the embodiment of the application, after acquiring an original training data set and an incremental training data set, a computer device may perform clustering processing on a plurality of training data included in the original training data set and a plurality of training data included in the incremental training data set to obtain N cluster sets; then, carrying out sampling compression processing on training data included in the N clustering sets to obtain a sampling training data set; and training the pre-trained data processing model by utilizing a sampling training data set to obtain a target data processing model. The training data included in the N clustering sets are sampled and compressed, the obtained sampled data set includes the training data in the original training data set and the training data in the incremental training data set, the problem of information redundancy caused by the training data is solved, all the incremental training data and the original training data do not need to be input into the pre-trained data processing model, and the quantity of the training data required by the pre-trained data processing model during retraining is reduced. The sampled training data set obtained by sampling can comprise training data in the original training data set and training data in the incremental training data set, so that the pre-trained data processing model is trained by the original training data and the incremental training data in a coordinated manner, and the accuracy of the model after retraining is improved.

Drawings

In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained according to the drawings without creative efforts.

FIG. 1 is a schematic flow chart diagram illustrating an image processing model training method according to an embodiment of the present disclosure;

fig. 2a is a schematic flowchart of a clustering network sampling and compressing training data in spatial feature distribution according to an embodiment of the present disclosure;

fig. 2b is a schematic diagram of a training of a clustering network according to an embodiment of the present application;

FIG. 2c is a schematic structural diagram of a pre-trained data processing model according to an embodiment of the present disclosure;

FIG. 3 is a schematic flowchart of an image processing model training method according to an embodiment of the present disclosure;

FIG. 4 is a schematic diagram of a model alternation training provided by an embodiment of the present application;

FIG. 5 is a schematic structural diagram of an image processing model training apparatus according to an embodiment of the present disclosure;

fig. 6 is a schematic structural diagram of a computer device according to an embodiment of the present application.

Detailed Description

The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are only a part of the embodiments of the present application, 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 application.

The embodiment of the application provides a model training scheme, which can acquire each training data, wherein each training data can be training data in an original training data set and training data in an incremental training data set; the original training data set is a set of training data used for training an initial data processing model; the incremental training data set is a set of newly added training data relative to the original training data set; then, clustering processing is carried out on each training data to obtain similarity distribution corresponding to each data category under a plurality of data categories (the subsequent similarity distribution is called a clustering set); then, sampling and compressing the cluster set corresponding to each data category to obtain a sampling training data set for model training; the pre-trained data processing model is then trained using the sampled training data set. In an embodiment, a metric learning method may be used to learn the similarity relationship between the training data, and perform clustering processing on the training data according to the similarity relationship between the training data to obtain a cluster set corresponding to each data category under multiple data categories. The metric learning may also be referred to as similarity learning, that is, the similarity between training data is calculated. After obtaining a plurality of cluster sets, a density sampling function (e.g., a logarithmic function) may be used to perform sampling compression on the cluster set corresponding to each data category, so as to obtain a sampling training data set for model training. Wherein, the density sampling function can satisfy the condition that the higher the density, the higher the sampling number of the cluster set, but the lower the sampling proportion.

In the prior art, when the pre-trained model is trained again, a large amount of training data is adopted to train the model again, so that the problem of information redundancy caused by the large amount of training data exists, namely when the large amount of training data participates in the training of the pre-trained model, the model needs to repeatedly learn repeated and similar information due to the distribution characteristic of natural data, so that the training data is unbalanced, and the model obtained after the re-training is poor in performance when facing difficult training data. In the embodiment of the application, each training data is clustered in the model processing scheme, and then the training data in the cluster set is sampled and compressed, so that the problem of training data redundancy can be solved to a large extent, all incremental training data and original training data are not required to be input into the pre-trained model, and the quantity of training data required by the pre-trained model during retraining is reduced.

In addition, incremental learning in the prior art, such as using distillation learning means, or memory-based incremental learning means, is a relatively rigid replica of the original model; in the process of retraining the pre-trained model, a new branch of the pre-trained model is trained only through new training data, and then an old branch in the pre-trained model is subjected to fine tuning by using a new branch training result. In the embodiment of the application, the original training data set and the incremental training data set are actually sampled by sampling and compressing the training data included in the plurality of cluster sets, and the sampled training data set can include the training data in the original training data set and the training data in the incremental training data set, so that the original training data and the incremental training data are cooperatively trained, the accuracy of the model after retraining is improved, and a paradigm capable of continuously learning the incremental training data set is also provided.

Based on the model training scheme provided above, please refer to fig. 1, fig. 1 is a schematic flowchart of a method for training an image processing model provided in an embodiment of the present application, where the method for training an image processing model is executable by a computer device, the computer device may be a terminal device or a server, and the terminal device may be a smart phone, a tablet computer, a laptop computer, a desktop computer, a vehicle-mounted device, or the like; the server may be an independent physical server, a server cluster or a distributed system formed by a plurality of physical servers, a cloud server providing basic cloud computing services such as cloud service, a cloud database, cloud computing, a cloud function, cloud storage, Network service, cloud communication, middleware service, domain name service, security service, Content Delivery Network (CDN), big data and an artificial intelligence platform, and the like. The image processing model training method may include the following steps S101 to S104:

s101, acquiring an original training data set and an incremental training data set.

The original training data set may include a plurality of training data, and may be referred to as an old training data set, where the training data in the original training data set is used to train an initial data processing model to obtain data of a pre-trained data processing model. The incremental training data set may include a plurality of training data, which may be referred to as a new training data set, which may be understood as newly generated training data (i.e., training data generated in addition to the plurality of training data in the original training data set). The original training data set and the incremental training data set are generally used for the same training task (e.g., training the same data processing model), for example, the training task is a training image processing model, the plurality of training data included in the original training data set and the plurality of training data included in the incremental training data set may all belong to training data of an image class, that is, the plurality of training data included in the original training data set are all a plurality of training images, the plurality of training data included in the incremental training data set are all a plurality of training images, and the plurality of training images may be different image classes; the plurality of training images are each used to train an image processing model. For another example, the training task is a training text processing model, the plurality of training data included in the original training data set and the plurality of training data included in the incremental training data set may all belong to training data of a text class, that is, the plurality of training data included in the original training data set are all a plurality of training texts, the plurality of training data included in the incremental training data set are all a plurality of training texts, the plurality of training texts may be different text classes, and the plurality of training texts are all used for training the text processing model.

As an embodiment, when the model needs to be trained or the pre-trained model needs to be trained again, a model training request can be submitted; the computer device may obtain an original training data set and an incremental training data set in response to the model training request. As another example, a computer device may retrieve an original training data set from a first memory space and an incremental training data set from a second memory space, where the first memory space and the second memory space may each be a local memory space of the computer device, a blockchain network, a device dedicated to data storage, and so on. The first storage space and the second storage space may be the same or different, and this is not limited in this embodiment of the application.

In one embodiment, the computer device may first obtain an initial original training data set and an initial incremental training data set, and then pre-process training data in the initial original training data set and training data in the initial incremental training data set to obtain an original training data set and an incremental training data set. Wherein the preprocessing may include resizing the training data in the initial raw training data set and the training data in the initial incremental training data set to a uniform size, and the like.

In the embodiment of the present application, there is a certain difference in the preprocessing of the training data. For example, when the training data is a training image, the pre-processing of the plurality of training images by the computer device may include one or more of: carrying out image standardization on a plurality of training images, namely adjusting the sizes of the training images to be uniform; and denoising the plurality of training images. For another example, when the sample data is a training text, the pre-processing of the plurality of training texts by the computer device may include one or more of: carrying out text standardization on each training text, namely adjusting the length of each training text to be uniform; and removing invalid characters in each training text, wherein the invalid characters can be set according to experience values or business requirements, and characters such as numbers, punctuations and the like can be set as the invalid characters.

S102, clustering a plurality of training data included in an original training data set and a plurality of training data included in an incremental training data set to obtain N clustering sets; each cluster set comprises one or more training data, and the one or more training data of each cluster set belong to the same data category.

Wherein N is an integer of 1 or more. Taking training data as an example of a training image, the data category may belong to an animal category, a person category, and the like. When N is 1, there is only one cluster set; the set of clusters may include one or more training images, each of the one or more training images in the set of clusters belonging to an animal class or a human class. When N is 2, a cluster set 1 and a cluster set 2 may be obtained, where each of the one or more training images included in the cluster set 1 belongs to an animal class; the cluster set 2 includes one or more training images that all belong to the person category.

In a specific implementation, the computer device may perform a similar comparison between every two training data on a plurality of training data included in the original training data set and a plurality of training data included in the incremental training data set, and perform a clustering process on the original training data set and the incremental training data set according to a similar comparison result between every two training data, to obtain N cluster sets.

The specific implementation manner of the step S102 may include the following three types:

(1) each training data in the original training data set corresponds to a data type, and each training data in the incremental training data set corresponds to a data type; the computer equipment can directly compare the data types of every two training data, and adds the two training data with the same data type into the cluster set corresponding to the data type, and after the two training data are compared, N cluster sets can be obtained. For example, the original training data set includes 2 training data, respectively training data 1 and training data 2; the training data 1 corresponds to the data type 1, and the training data 2 corresponds to the data type 2; the incremental training data set comprises 2 training data which are respectively training data 3 and training data 4, wherein the training data 3 corresponds to the data category 1; training data 4 corresponds to data category 2; the computer equipment compares the data type of the training data 1 with the data type of the training data 2, and finds that the data type of the training data 1 is different from the data type of the training data 2; then the computer equipment compares the data type of the training data 1 with the data type of the training data 3, and finds that the data type of the training data 1 is the same as the data type of the training data 3; adding training data 1 and training data 3 into a cluster set corresponding to the data category 1; the computer device then compares the data class of training data 1 with the data class of training data 4, finding that the data class of training data 1 is different from the data class of training data 2. Then comparing the data type of the training data 2, the data type of the training data 3 and the data type of the training data 4 according to the data type comparison process of the training data 1, and finally adding the training data 2 and the training data 4 into a cluster set corresponding to the data type 2; finally generating 2 (N-2) cluster sets; respectively, a collection of clusters corresponding to data category 1 and a collection of clusters corresponding to data category 2.

(2) Performing similarity calculation on every two training data in the plurality of training data included in the original training data set and the plurality of training data included in the incremental training data set by adopting a similarity calculation method to obtain the similarity of every two training data; and clustering every two training data with similarity greater than a similarity threshold to obtain N cluster sets. Wherein, the similarity algorithm can be cosine similarity algorithm, Pearson correlation coefficient, etc.; the similarity threshold may be set according to experience or requirements, which is not limited in the embodiment of the present application.

For example, the original training data set includes training data 1, the incremental training data set includes training data 2 and training data 3; the similarity algorithm is a cosine similarity algorithm, and the similarity threshold is a threshold value 1; the computer device may perform cosine similarity calculation on a vector corresponding to the training data 1 and a vector corresponding to the training data 2 by using a cosine similarity calculation method to obtain cosine similarity between the training data 1 and the training data 2, and then the computer device determines whether the cosine similarity between the training data 1 and the training data 2 is greater than a threshold 1; the computer equipment determines that the cosine similarity between the training data 1 and the training data 2 is greater than a threshold value 1, and carries out clustering processing on the training data 1 and the training data 2 to obtain a clustering set; similarly, the computer device may determine the cosine similarity between the training data 1 and the training data 3, determine that the cosine similarity between the training data 1 and the training data 3 is smaller than the threshold 1, and add the training data 3 to another clustering set without performing clustering on the training data 1 and the training data 3, thereby finally obtaining two clustering sets.

(3) In the above (1) and (2), because the number of the training data is large, the time spent in comparing one by one is long, therefore, the embodiment of the present application may adopt the clustering network to perform feature extraction on a plurality of training data included in the original training data set and a plurality of training data included in the incremental training data set, so as to achieve that the features of the training data having the same data category are as consistent as possible, the features of the training data having different data categories are as different as possible, and the clustering process may be performed on the plurality of training data included in the original training data set and the plurality of training data included in the incremental training data set more quickly and accurately. At this time, the computer device performs clustering processing on a plurality of training data included in the original training data set and a plurality of training data included in the incremental training data set, and the specific implementation manner of obtaining N clustering sets is as follows: calling a clustering network to perform feature extraction on a plurality of training data included in an original training data set and a plurality of training data included in an incremental training data set to obtain a feature vector corresponding to each training data; the cluster network may be various types of deep neural networks, for example, the cluster network may be a recurrent neural network model, and when the training data is a training image, the cluster network may also be ResNet50, and so on; when the training data is training text, the clustering network may also be Word2vec (Word vector model) and so on. After passing through the clustering network, the feature vector of each training data is a high-dimensional feature vector with the same size and dimension, so that each training data can be conveniently clustered according to the feature vector of each training data in the follow-up process. And the computer equipment carries out clustering processing on each training data based on the characteristic vector corresponding to each training data to obtain N clustering sets.

As an implementation manner, the clustering network is configured to extract a high-dimensional feature vector of each training data, and the computer device may perform clustering processing on each training data based on the feature vector corresponding to each training data by using a clustering algorithm to obtain N cluster sets. Wherein the clustering algorithm may be mean shift clustering, density-based clustering methods, and the like. As another embodiment, the clustering network may not only extract the high-dimensional feature vector of each training data, but also include a clustering algorithm, and the computer device may invoke the clustering algorithm in the clustering network to process the feature vector corresponding to each training data, so as to directly obtain N cluster sets.

It should be noted that, each training data is subjected to cluster analysis through a cluster network and a cluster algorithm, so that a feature space distribution can be actually obtained, and the feature space distribution may include N cluster sets (the N cluster sets may be understood as mapping the feature space distribution into N similarity distribution regions, and clustering the training data of the same data category in the same similarity distribution region). For example, in fig. 2a, the computer device invokes the clustering network to perform feature extraction on each training data to obtain a feature vector corresponding to each training data, and then performs clustering processing on each training data according to the feature vector corresponding to each training data by using a clustering algorithm to obtain 3 similarity distribution areas in the feature spatial distribution, which are a similarity distribution area 11, a similarity distribution area 12, and a similarity distribution area 13, respectively.

Before calling a clustering network to perform feature extraction on a plurality of training data included in an original training data set and a plurality of training data included in an incremental training data set, the deep neural network needs to be trained to obtain the clustering network. The specific objective of the trained clustering network is to make the features of the training data with the same data category as consistent as possible, and make the features of the training data with different data categories as different as possible greatly different, so as to facilitate the accuracy and success rate of clustering the training data by using a clustering algorithm subsequently. The whole clustering network training process can be seen in fig. 2 b: optionally, in fig. 2b, the sample data set may also include an original sample data set and an incremental sample training set, where the original sample data set includes the sample data and a data category tag corresponding to the sample data, and the incremental sample data set includes the sample data and a data category tag corresponding to the sample data; the clustering network can be obtained by training based on a sample data set; then, the computer equipment calls a clustering network to extract the features of each sample data to obtain the feature vector of each sample data; similarity calculation is carried out on the feature vectors of every two sample data to obtain similarity measurement features between every two sample data; and then determining the value of a contrast loss function according to the similarity measurement characteristic between every two sample data and the data category label of each sample data, and optimizing the network parameters of the clustering network according to the direction of reducing the value of the contrast loss function.

The similarity calculation is performed on the feature vectors of every two sample data, and the similarity measurement features between every two sample data are obtained in the following two modes: (1) the computer equipment calls the clustering network to extract the features of each sample data to obtain the feature vector of each sample data in the feature space; then, the similarity between every two sample data is measured by adopting the Euclidean distance in the feature space, that is, the computer device can calculate the Euclidean distance of the feature vector of every two sample data in the feature space, and at this time, the similarity measurement feature between every two sample data is actually the Euclidean distance corresponding to every two sample data. (2) The computer equipment calls the clustering network to extract the features of each sample data to obtain the feature vector of each sample data in the feature space; and then, calculating the mean square error between every two sample data according to the feature vectors of every two sample data in the feature space to obtain the similarity measurement feature between every two sample data. It should be understood that the mean square error between every two sampled data is a similarity metric characteristic of every two sampled data.

In one embodiment, the specific implementation manner of the computer device determining the value of the contrast loss function according to the similarity metric feature between every two sample data and the data category label of each sample data is as follows: obtaining a contrast loss value between every two sample data according to the similarity measurement characteristics between every two sample data and the data category label of every sample data; and carrying out preset operation processing on the obtained contrast loss value to obtain a value of a contrast loss function. As an implementation, the preset operation may be a direct summation; after obtaining the contrast loss value between every two sample data, the computer device may directly perform summation processing on the contrast loss value between every two sample data in the plurality of sample data to obtain the value of the contrast loss function. For example, the computer equipment obtains a contrast loss value between sample data 1 and sample data 2 as 10; the contrast loss value between sample data 3 and sample data 4 is 3; and then the computer equipment directly sums the contrast loss value 10 between the sample data 1 and the sample data 2 and the contrast loss value 3 between the sample data 3 and the sample data 4 to obtain the value of the contrast loss function. As another embodiment, the preset operation may be weighted summation; after obtaining the contrast loss value between every two sample data, the computer device may perform weighting processing on the contrast loss value between every two sample data to obtain a weighted value between every two sample data; and then summing weighted values between every two sample data in the plurality of sample data to obtain a value of the contrast loss function.

In order to realize higher similarity of the same data category label, the Euclidean distance in the corresponding feature space is smaller; the similarity of different data category labels is low, and the Euclidean distance in the corresponding feature space is large. Based on the idea, a contrast loss function of two parameters can be introduced in the process of training the clustering network, and the two parameters can be marked as a first parameter and a second parameter. By the contrast loss function provided by the embodiment of the application, for sample data of the same data category, the contrast loss function aggregates the sample data in the soft boundary of the first parameter as much as possible, and for sample data of different data category labels, the contrast loss function enables the distance to be in the soft boundary higher than the second parameter as much as possible. In one embodiment, the set of sample data includes first sample data and second sample data; the specific implementation manner of obtaining the comparison loss value between every two sample data according to the similarity measurement feature between every two sample data and the data category label of every sample data is as follows: the computer equipment judges whether the data type label of the first sample data is the same as the data type label of the second sample data, and if the data type label of the first sample data is determined to be the same as the data type label of the second sample data, a first difference value between the similarity measurement characteristic of the first sample data and the second sample data and the first parameter is obtained; in particular, the computer device may calculate a first difference between the similarity metric feature and the first parameter between the first sample data and the second sample data using a contrast loss function. And then carrying out linear operation on the first difference value to obtain a contrast loss value between the first sample data and the second sample data. The first parameter can be set according to requirements, the first parameter can be understood as a distance threshold, and for sample data of the same data category, similarity measurement features between the first sample data and the second sample data are gathered in the first parameter as much as possible. The computer device may calculate a first difference between the similarity measurement characteristic between the first sample data and the second sample data and the first parameter, and then perform linear operation on the first difference by using a linear rectification function to obtain a contrast loss value between the first sample data and the second sample data.

If the data type label of the first sample data is determined to be different from the data type label of the second sample data, acquiring a second difference value between the similarity measurement characteristic between the first sample data and the second parameter; in particular, the computer device may calculate a first difference between the similarity metric feature between the first sample data and the second parameter using a contrast loss function. And then carrying out linear operation on the second difference value to obtain a contrast loss value between the first sample data and the second sample data. The second parameter can be set according to requirements, the second parameter can be understood as a distance threshold, and for sample data of different data categories, the similarity measurement feature is higher than the second parameter as much as possible. The computer device may calculate a second difference between the similarity measurement feature between the first sample data and the second parameter, and then perform linear operation on the second difference by using a linear rectification function to obtain a contrast loss value between the first sample data and the second sample data.

The formula for the contrast loss function for the first sample data and the second sample data is as follows:

wherein, contextureloss is a value of a loss of contrast between the first sample data and the second sample data, output1 and output2 respectively represent the first sample data and the second sample data, and softMargin1 represents the first parameter; softMargin2 represents the second argument, relu () represents a linear rectification function, label1 represents a data class label of the first sample data, and label2 represents a data class label of the second sample data.

It should be noted that, the above is only exemplified by the sample data set including the first sample data and the second sample data, and at this time, the value of the final contrast loss function is the contrast loss value calculated by the first sample data and the second sample data; when the sample data set comprises a plurality of sample data, the contrast loss value of every two sample data can be calculated according to the contrast loss values of the first sample data and the second sample data, and finally, all the obtained contrast loss values are subjected to preset operation processing, so that the value of the contrast loss function can be obtained.

In one embodiment, when training the clustering network by using a plurality of sample data, the plurality of sample data needs to be preprocessed, so that the clustering network can extract feature vectors with the same size and dimension when extracting features. In the embodiment of the present application, there is a certain difference in the preprocessing for different sample data. For example, when the sample data is a sample image, the pre-processing of the plurality of sample images by the computer device may include: carrying out image standardization on a plurality of sample images, namely adjusting the sizes of the sample images into a uniform size; and denoising the plurality of sample images. For another example, when the sample data is a sample text, the pre-processing of the plurality of sample texts by the computer device may include: carrying out text standardization on each sample text, namely adjusting the size of each sample text to be uniform; invalid characters in each sample text are removed, wherein the invalid characters can be set according to experience values or business requirements, and characters such as numbers, punctuation marks and the like can be set as the invalid characters.

In an embodiment, training data included in each of the N cluster sets may be repeated, and in order to make the sampled training data in the sampled training data set different when step S103 is executed subsequently, deduplication processing needs to be performed before the N cluster sets are obtained. The computer device performs clustering processing on a plurality of training data included in an original training data set and a plurality of training data included in an incremental training data set to obtain N clustering sets, and the specific implementation mode of the method is as follows: clustering a plurality of training data included in an original training data set and a plurality of training data included in an incremental training data set to obtain N initial clustering sets; and then, carrying out de-duplication processing on each initial cluster set in the N initial cluster sets to obtain N cluster sets. For example, the initial cluster set a in the N initial cluster sets includes training data 1, training data 2, and training data 3, and the computer device performs deduplication processing on the initial cluster set a to obtain a cluster set a, where the cluster set a includes training data 1, training data 2, and training data 3.

S103, carrying out sampling compression processing on the training data included in the N clustering sets to obtain a sampling training data set.

In one embodiment, the computer device may obtain the total number of the sampled training data to be sampled, then determine the number of training data to be collected from each cluster set based on the total number of the sampled training data to be sampled and the number of training data included in each cluster set, perform sampling compression processing on each cluster set according to the number of training data to be collected in each cluster set, obtain the sampled training data from each cluster set, and add the sampled training data obtained from each cluster set to the sampled training data set. For example, N ═ 2, 2 cluster sets are cluster set 1 and cluster set 2, respectively; the number of training data included in the cluster set 1 is 30, and the number of training data included in the cluster set 2 is 10; the total quantity of sampling training data needing to be sampled, which is acquired by computer equipment, is 10, which means that 10 training data are acquired in total and are used as sampling training data in a sampling training data set; the computer equipment determines that the number of training data needing to be collected from the cluster set 1 is 8 and the number of training data needing to be collected from the cluster set 2 is 2 according to the total number of the sampled training data needing to be sampled, the number of training data included in the cluster set 1 and the number of training data included in the cluster set 2; then the computer device performs sampling compression processing on the training data included in the cluster set 1, acquires 8 sampling training data from the cluster set 1, performs sampling compression processing on the training data included in the cluster set 2, acquires 2 sampling training data from the cluster set 2, and adds the 8 sampling training data and the 2 sampling training data to the sampling training data set.

In another embodiment, a sampling ratio for sampling the training data included in the N cluster sets may be set, and the computer device may perform sampling compression processing on the training data included in the N cluster sets according to the total number of the sampling training data that needs to be sampled and the sampling ratio, to obtain a sampling training data set. For example, N ═ 3, i.e., there are 3 cluster sets, cluster set 1, cluster set 2, and cluster set 3, respectively; the total amount of the sampled training data needing to be sampled is 10; the computer device then obtains a sampling ratio for the training data included in the 3 sets of clusters, assuming a sampling ratio of 1: 2: 2; then, the computer equipment performs sampling compression processing on the cluster set 1 according to the sampling proportion and the total number 10 of the sampling training data needing to be sampled, and acquires 2 sampling training data from the cluster set 1; the computer equipment performs sampling compression processing on the cluster set 2 according to the sampling proportion and the total number 10 of the sampling training data to be sampled, and acquires 4 sampling training data from the cluster set 2; the computer equipment performs sampling compression processing on the cluster set 3 according to the sampling proportion and the total number 10 of the sampling training data to be sampled, and acquires 4 sampling training data from the cluster set 3; the sampled training data obtained from these 3 sets is then added to the sampled training data set.

S104, training the pre-trained data processing model by using a sampling training data set to obtain a target data processing model; the pre-trained data processing model is trained based on an original training data set.

From the perspective of the training data distribution, the original training data set and the incremental training data set are in two different but similar distributions, wherein the two similar distributions can be understood as: the original training data set and the incremental training data set are independent data sets, but the original training data set and the incremental training data set obey probability distribution; the difference between the original training data set and the incremental training data set is mainly influenced by the randomness within the respective training data sets. Therefore, the data processing model that can be pre-trained in the present application can include a correction bias module for correcting the offset caused by the training data in the incremental training data set. The architecture of the pre-trained data processing model, which may include a correction bias module and other data processing modules, may be as shown in fig. 2 c. The correction bias module may comprise a network of simple relu functions (an activation function), e.g. the correction bias module uses a network of two to three layers of relu functions; other data processing modules may include feature extraction layers, hidden layers, fully connected layers, and the like. The embodiment of the present application does not limit the correction bias module and other data processing modules.

In the specific implementation, the computer device firstly trains the data processing model by using an original training data set to obtain a pre-trained data processing model, and then alternately trains a correction bias module and other data processing modules included in the pre-trained data processing model by using a sampling training data set to obtain a target data processing model.

In the embodiment of the application, after acquiring an original training data set and an incremental training data set, a computer device may perform clustering processing on a plurality of training data included in the original training data set and a plurality of training data included in the incremental training data set to obtain N cluster sets; then, carrying out sampling compression processing on training data included in the N clustering sets to obtain a sampling training data set; and training the pre-trained data processing model by utilizing a sampling training data set to obtain a target data processing model. By carrying out sampling compression processing on the training data included by the N clustering sets, the obtained sampling data set comprises the training data in the original training data set and the training data in the incremental training data set, the problem of information redundancy caused by the training data is solved, model training is carried out by cooperation of old and new data, and the accuracy of the trained model is improved.

Based on the model training scheme provided above, please refer to fig. 3, fig. 3 is a schematic flowchart of a method for training an image processing model according to an embodiment of the present application, where the method for training an image processing model is executable by a computer device, the computer device may be a terminal device or a server, and the terminal device may be a smart phone, a tablet computer, a laptop computer, a desktop computer, a vehicle-mounted device, or the like; the server may be an independent physical server, a server cluster or a distributed system formed by a plurality of physical servers, a cloud server providing basic cloud computing services such as cloud service, a cloud database, cloud computing, a cloud function, cloud storage, Network service, cloud communication, middleware service, domain name service, security service, Content Delivery Network (CDN), big data and an artificial intelligence platform, and the like. The image processing model training method may include the following steps S301 to S306:

s301, acquiring an original training data set and an incremental training data set.

S302, clustering a plurality of training data included in the original training data set and a plurality of training data included in the incremental training data set to obtain N clustering sets; each cluster set comprises one or more training data, and the one or more training data of each cluster set belong to the same data category.

The specific implementation manner of the steps S301 to S302 can refer to the steps S101 to S102 in fig. 1, which is not described herein again.

S303, determining the target number of the training data to be sampled in each cluster set based on the sampling parameters and the number of the training data included in each cluster set in the N cluster sets.

The sampling parameter is a variable parameter, and the sampling parameter can be a parameter for controlling the sampling degree, the larger the value of the sampling parameter is, the less training data is obtained by sampling, and the smaller the value of the sampling parameter is, the more training data is obtained by sampling.

In one embodiment, based on the result of clustering the training data included in the original training data set and the training data included in the incremental training data set, the similarity distribution of the training data is obtained (i.e., N clustering sets are obtained), and after multiple times of experimental verification, the class center points of the same data class can be obtained by calculating the feature average values of the training data corresponding to the same data class. And by analyzing the distribution condition of the distance histogram between each training data and the central point of the category, the distance distribution between each training data and the central point of the category can be found to be in accordance with logarithmic Gaussian distribution. Therefore, the number of training data to be sampled for each cluster set is calculated by taking the density sampling function as a logarithmic function into consideration in the embodiment of the present application. In this case, the computer device may determine, by using a logarithmic function, the number of training data included in each of the N cluster sets, and then, for a target cluster set in the N cluster sets, the computer device may determine, according to the sampling parameter and the target number of training data included in the target cluster set, the target number of training data to be sampled in the target cluster set. The expression of the logarithmic function is as follows:

s＝log_α(D+1)

wherein, S is the number of training data needing random sampling in the target cluster set in the N cluster sets, D is the target number of training data in the target cluster set, and alpha is a sampling parameter.

It should be noted that, the implementation process of determining the target number of the training data to be sampled in any one of the N cluster sets may refer to the above-mentioned target number of the training data included in the target cluster set and the sampling parameter to determine the target number of the training data to be sampled in the target cluster set. The sampling parameters used for determining the target number of training data to be sampled in each cluster set according to the sampling parameters and the number of training data included in each cluster set may be the same or different, the sampling parameters may be determined according to the number of training data included in each cluster set, or the sampling parameters may be set according to requirements, and the embodiments of the present application are not limited.

S304, performing compression sampling processing on each cluster set according to the target number of the training data to be sampled in each cluster set, and acquiring sampling training data from each cluster set.

In a specific implementation, the computer device performs compression sampling processing on each cluster set according to the target number of training data to be sampled in each cluster set, and randomly acquires the corresponding target number of sampled training data from each cluster set. For example, let the number of cluster sets be 2, cluster set 1 and cluster set 2, respectively; the target number of the training data needing to be sampled in the cluster set 1 is 2, and the target number of the training data needing to be sampled in the cluster set 2 is 3; the computer equipment can perform compression sampling processing on the cluster set 1 according to the target number 2 of the training data to be sampled in the cluster set 1, and randomly acquire 2 sampling training data from the cluster set 1; and performing compression sampling processing on the cluster set 2 according to the target number 3 of the training data to be sampled in the cluster set 2, and randomly acquiring 3 sampling training data from the cluster set 2.

In one embodiment, the computer device may further divide each cluster set into a plurality of cluster subsets, and then sample the training data in each cluster subset. In a specific implementation, the N cluster sets include a target cluster set, the target cluster set is any one of the N cluster sets, the training data included in each cluster set is compressed and sampled according to target data of training data to be sampled in each cluster set, and a specific implementation manner of acquiring sampled training data from each cluster set may be: dividing the target cluster set into a plurality of cluster subsets according to a division rule; calculating the quantity ratio of the training data in the plurality of clustering subsets; performing compression sampling processing on each clustering subset based on the number ratio and the target number of the training data to be sampled in the target clustering set, and acquiring sampling training data from each clustering subset; and determining the sampled training data acquired from each cluster subset as the sampled training data acquired from the target cluster set.

As an embodiment, the dividing rule may be to divide the target cluster set based on the amount of training data in the target cluster set; the computer device may divide the training data in the target cluster set averagely or according to a preset number ratio to obtain a plurality of cluster subsets. As another embodiment, the dividing rule may be that the target cluster set is divided based on the distribution condition of training data in the target cluster set in the feature space; the computer device can determine the spatial distribution distance (or Euclidean distance) of every two training data in the target cluster set, then divide the two training data corresponding to the spatial distribution distance satisfying the first preset value in the target cluster set into one cluster subset, divide the two training data corresponding to the spatial distribution distance satisfying the second preset value into another cluster subset, and so on, and finally obtain a plurality of cluster subsets of the target cluster set. The first preset value, the second preset value and the third preset value can be set according to requirements.

For example, the number of the training data to be sampled in the target cluster set is 11; the target cluster set comprises 20 training data; the division rule is divided according to the spatial distribution condition of the training data; the computer equipment can divide training included in the target clustering set into a clustering subset 1, a clustering subset 2 and a clustering subset 3 according to the spatial distribution condition of the training data, wherein the spatial distribution distance between every two training data in the clustering subset 1 is smaller than a first preset value; the spatial distribution distance between every two training data in the clustering subset 2 is smaller than a second preset value; the spatial distribution distance between every two training data in the clustering subset 3 is smaller than three preset values; wherein the training number in the cluster subset 1 is 10; the number of training in the clustering subset 2 is 6; the number of training in the clustering subset 2 is 6; then the computer device can calculate the number ratio of the training data in the 3 clustering subsets, wherein the number ratio is 5:3:3, then the computer device can perform sampling compression processing on each clustering subset according to the target number and the number ratio of the training data to be sampled slowly in the target clustering set, obtain 5 sampling training data from the clustering subset 1, obtain 3 sampling training data from the clustering subset 2, and obtain 3 sampling training data from the clustering subset 3. Then, the sampled training data obtained from the 3 clustering subsets are determined as the sampled training data obtained from the target clustering set.

And S305, adding the sampled training data acquired from each cluster set to a sampled training data set.

In particular implementations, a computer device may add a target number of sampled training data acquired in each cluster set to a set of sampled training data. The sampled training data set may include training data in each cluster set.

S306, training the pre-trained data processing model by using a sampling training data set to obtain a target data processing model; the pre-trained data processing model is trained based on an original training data set.

The sampling training data set comprises a first type of sampling training data subset and a second type of sampling training data subset, training data in the first type of sampling training data subset is from an incremental training data set, and training data in the second type of sampling training data subset is from an original training data set.

In a specific implementation, the computer device may perform an alternating training of the pre-trained data processing model using the sampled training data set and a subset of the second type of sampled training data in the sampled training data set to obtain the target data processing model.

In an embodiment, as can be seen from the foregoing description, the pre-trained data processing model includes a bias correction module and other data processing modules, and the computer device performs alternating training on the pre-trained data processing model by using the sampling training data set and the second type of sampling training data subset in the sampling training data set, so as to obtain a target data processing model, which may be implemented as shown in fig. 4: firstly, keeping the model parameters of other data processing modules unchanged, and training a correction offset module by adopting a sampling training data set; and when the training of the offset correcting module is finished, keeping the model parameters in the offset correcting module unchanged, and training other data processing modules by adopting a second type of sampling training data subset in the sampling training data set to obtain a target data processing model. Wherein, the completion of the training of the correction bias module means that the training of the correction bias module converges. The alternative training mode is mainly based on the idea that: the correction bias module is used for correcting the offset brought by the training data in the incremental training data set. When the correction bias module converges, the offset of the training data in the incremental training data set may be considered to have been corrected to some extent. At this time, the data with reduced offset participates in training, and the effect can be better improved.

In an embodiment of the application, a computer device may obtain an original training dataset and an incremental training dataset; clustering a plurality of training data included in an original training data set and a plurality of training data included in an incremental training data set to obtain N cluster sets, and determining the target number of the training data to be sampled in each cluster set based on the sampling parameters and the number of the training data included in each cluster set in the N cluster sets; respectively carrying out compression sampling processing on each cluster set according to the target number of training data to be sampled in each cluster set, and acquiring sampling training data from each cluster set; and adding the sampling training data acquired from each cluster set into a sampling training data set, and training the pre-trained data processing model by using the sampling training data set to obtain a target data processing model. The target quantity of the training data to be sampled in each cluster set can be controlled through the sampling parameters, so that the sampled collected training data set can be well controlled, the problem of information redundancy caused by the training data is solved, and the model training efficiency and the accuracy of the trained model are improved.

The image processing model training method provided by the embodiment of the application can be applied to model training of image processing models, image classification models and the like. When the image processing model training method is applied to model training of an image processing model, the whole model training process is as follows: acquiring an initial original training image set and an initial incremental training image set, preprocessing a plurality of training images included in the initial original training image set and a plurality of training images included in the initial incremental training image set by computer equipment to obtain the original training image set and the incremental training image set, and then calling a clustering network to perform feature extraction on the plurality of training images included in the original training image set and the plurality of training images included in the incremental training image set to obtain a feature vector corresponding to each training image; then, clustering each training image according to the feature vector corresponding to each training image by adopting a clustering algorithm to obtain a plurality of similarity distribution areas (namely N clustering sets) in a feature space; the training images of the same image category are gathered in the same similar distribution area as much as possible through a clustering network and a clustering algorithm.

Then, the computer device calculates the target number of training images needing to be acquired in each similarity distribution area by adopting a logarithmic function, then performs compression sampling processing on the training images included in each similarity distribution area according to the target number of the training images needing to be acquired in each similarity distribution area, acquires the training images of the target number from the training images included in each similarity distribution area, and adds the training images of the target number acquired from the training images included in each similarity distribution area into a sampling training image set; the sampling training image set comprises a first type of sampling training image subset and a second type of sampling training image subset, wherein the sampling training images in the first type of sampling training image subset are from an incremental training image set; the sampled training images in the second type of sampled training image subset are from the original training image set; and then the computer equipment firstly trains the image processing model by using the original training image set to obtain a pre-trained image processing model. Then, the relu functions of the last three layers in the pre-trained image processing model are used as a shallow network (namely the mentioned bias correcting module) for correcting bias, then model parameters of other image processing modules except the shallow network in the pre-trained image processing model are fixed, a sampling training image set is adopted to train the shallow network, when the shallow network is trained, the model parameters of the shallow network can be fixed, and a second type of sampling training image subset in the sampling training image set is adopted to train other image processing modules except the shallow network in the pre-trained image processing model, so that the target image processing model is obtained.

After the target image processing model is obtained, experiments are performed on some business data in order to further verify the effectiveness of the image processing model training method. The result shows that the image processing model training method can basically approach the result obtained by co-training on the basis of the number of the compressed training images, and exceeds the method for fine-tuning the original training image set by using the incremental training image set. And at low false acceptance rate, the target image processing model accuracy rate is higher. The accuracy of the target image processing model in low error rate is improved, and a paradigm capable of continuously learning the incremental training image set is provided.

Please refer to fig. 5, which is a schematic structural diagram of an image processing model training apparatus according to an embodiment of the present disclosure. As shown in fig. 5, the image processing model training apparatus may be applied to the computer device in the corresponding embodiment of fig. 1 or fig. 3; specifically, the image processing model training device may be a computer program (including program code) running on a computer device, for example, the image processing model training device is an application software; the image processing model training device can be used for executing corresponding steps in the method provided by the embodiment of the application.

An obtaining unit 501, configured to obtain an original training data set and an incremental training data set;

a processing unit 502, configured to perform clustering on a plurality of training data included in the original training data set and a plurality of training data included in the incremental training data set to obtain N cluster sets; each cluster set comprises one or more training data, and the one or more training data of each cluster set belong to the same data category;

the processing unit 502 is further configured to perform sampling compression processing on the training data included in the N cluster sets to obtain a sampled training data set;

the processing unit 502 is further configured to train a pre-trained data processing model by using the sampling training data set to obtain a target data processing model; the pre-trained data processing model is trained based on the original training data set.

In an embodiment, when the processing unit 502 performs clustering on a plurality of training data included in the original training data set and a plurality of training data included in the incremental training data set to obtain N cluster sets, it may specifically be configured to:

calling a clustering network to perform feature extraction on a plurality of training data included in the original training data set and a plurality of training data included in the incremental training data set to obtain a feature vector corresponding to each training data;

and clustering each training data based on the characteristic vector corresponding to each training data to obtain N cluster sets.

In another embodiment, the clustering network is trained based on a sample data set, where the sample data set includes a plurality of sample data and a data category label of each sample data, and the processing unit 502 is further configured to:

calling a clustering network to perform feature extraction on each sample data to obtain a feature vector of each sample data;

performing similarity calculation on the feature vectors of every two sample data to obtain similarity measurement features between every two sample data;

and determining the value of a contrast loss function according to the similarity measurement characteristic between every two sample data and the data category label of each sample data, and optimizing the network parameters of the clustering network according to the direction of reducing the value of the contrast loss function.

In another embodiment, when determining the value of the contrast loss function according to the similarity metric feature between every two sample data and the data category label of each sample data, the processing unit 502 may specifically be configured to:

obtaining a contrast loss value between every two sample data according to the similarity measurement characteristics between every two sample data and the data category label of every sample data;

and carrying out preset operation processing on the obtained contrast loss value to obtain a value of a contrast loss function.

In another embodiment, the sample data set includes a first sample data and a second sample data, and when the processing unit 502 obtains the contrast loss value between every two sample data according to the similarity metric characteristic between every two sample data and the data category label of every sample data, it may specifically be configured to:

if the data type label of the first sample data is the same as the data type label of the second sample data, acquiring a first difference value between the similarity measurement characteristic of the first sample data and the second sample data and a first parameter; performing linear operation on the first difference to obtain a contrast loss value between the first sample data and the second sample data;

if the data type label of the first sample data is different from the data type label of the second sample data, acquiring a second difference value between the similarity measurement characteristic of the first sample data and the second sample data and a second parameter; and performing linear operation on the second difference to obtain a contrast loss value between the first sample data and the second sample data.

In another embodiment, when the processing unit 502 performs sampling compression processing on the training data included in the N cluster sets to obtain a sampled training data set, the processing unit may specifically be configured to:

determining the target number of training data to be sampled in each cluster set based on the sampling parameters and the number of training data included in each cluster set in the N cluster sets;

respectively performing compression sampling processing on each cluster set according to the target number of the training data to be sampled in each cluster set, and acquiring sampling training data from each cluster set;

adding the sampled training data obtained from each cluster set to a sampled training data set.

In another embodiment, the N cluster sets include a target cluster set, where the target cluster set is any one of the N cluster sets, and the processing unit 502 performs compression sampling processing on training data included in each cluster set according to a target number of training data to be sampled in each cluster set, where when sampling training data is obtained from each cluster set, the processing unit may be specifically configured to:

dividing the target cluster set into a plurality of cluster subsets according to a division rule;

calculating the number ratio of the training data in the plurality of clustering subsets;

performing compression sampling processing on each clustering subset based on the number ratio and the target number of the training data to be sampled in the target clustering set, and acquiring sampling training data from each clustering subset;

and determining the obtained sampling training data in each clustering subset as the obtained sampling training data in the target clustering set.

In yet another embodiment, the sampled training data set includes a first type of sampled training data subset and a second type of sampled training data subset, the training data in the first type of sampled training data subset is from the incremental training data set, and the training data in the second type of sampled training data subset is from the original training data set;

when the processing unit 502 trains the pre-trained data processing model by using the sampling training data set to obtain the target data processing model, it may specifically be configured to:

and alternately training the pre-trained data processing model by adopting the sampling training data set and a second type of sampling training data subset in the sampling training data set to obtain a target data processing model.

In another embodiment, the pre-trained data processing model includes a bias correction module and other data processing modules, and when the processing unit 502 performs alternate training on the pre-trained data processing model by using the sampling training data set and a second type of sampling training data subset in the sampling training data set, to obtain the target data processing model, the processing unit may be specifically configured to:

keeping the model parameters of the other data processing modules unchanged, and training the correction bias module by adopting the sampling training data set;

and when the training of the offset correcting module is finished, keeping the model parameters in the offset correcting module unchanged, and training the other data processing modules by adopting a second type of sampling training data subset in the sampling training data set to obtain a target data processing model.

According to an embodiment of the present application, the steps involved in the image processing model training method shown in fig. 1 and 3 may be performed by the units in the image processing model training apparatus shown in fig. 5. For example, step S101 shown in fig. 1 may be performed by the obtaining unit 501 in the image processing model training apparatus shown in fig. 5, and steps S202 to S204 may be performed by the processing unit 502 in the image processing model training apparatus shown in fig. 5; as another example, step S301 in the image processing model training method shown in fig. 3 may be performed by the obtaining unit 501 in the image processing model training apparatus shown in fig. 5, and steps S302 to S306 may be performed by the processing unit 502 in the image processing model training apparatus shown in fig. 5.

According to another embodiment of the present application, the units in the image processing model training apparatus shown in fig. 5 may be respectively or entirely combined into one or several other units to form the image processing model training apparatus, or some unit(s) thereof may be further split into multiple units with smaller functions to form the image processing model training apparatus, which may achieve the same operation without affecting the achievement of the technical effect of the embodiment of the present application. The units are divided based on logic functions, and in practical application, the functions of one unit can be realized by a plurality of units, or the functions of a plurality of units can be realized by one unit. In other embodiments of the present application, the training apparatus based on the image processing model may also include other units, and in practical applications, these functions may also be implemented by the assistance of other units, and may be implemented by cooperation of a plurality of units.

According to another embodiment of the present application, the image processing model training apparatus as shown in fig. 5 may be constructed by running a computer program (including program codes) capable of executing the steps involved in the respective methods shown in fig. 1 and 3 on a general-purpose computing device such as a computer including a processing element such as a Central Processing Unit (CPU), a random access storage medium (RAM), a read-only storage medium (ROM), and a storage element, and the image processing model training method according to an embodiment of the present application may be implemented. The computer program may be embodied on a computer-readable storage medium, for example, and loaded into and executed by the computer apparatus described above via the computer-readable storage medium.

Further, please refer to fig. 6, which is a schematic structural diagram of a computer device according to an embodiment of the present application. The computer device may specifically be the computer device in the embodiment corresponding to fig. 2 or fig. 5. As shown in fig. 6, the computer apparatus may include: a processor 601, an input device 602, an output device 603, and a computer storage medium 604. The processor 601, input device 602, output device 603, and computer storage medium 604 described above are connected by a bus 605.

The computer storage medium 604 described above may be stored in a memory of a computer device, the computer storage medium 604 being used for storing computer programs, and the processor 601 being used for executing the computer programs stored by the computer storage medium 604. The processor 601 (or CPU) is a computing core and a control core of a computer device, and is adapted to implement one or more computer programs, and specifically adapted to load and execute: acquiring an original training data set and an incremental training data set; clustering a plurality of training data included in the original training data set and a plurality of training data included in the incremental training data set to obtain N cluster sets; each cluster set comprises one or more training data, and the one or more training data of each cluster set belong to the same data category; sampling and compressing the training data included in the N cluster sets to obtain a sampling training data set; training the pre-trained data processing model by using the sampling training data set to obtain a target data processing model; the pre-trained data processing model is trained based on the original training data set.

An embodiment of the present application further provides a computer storage medium (Memory), which is a Memory device in a computer device and is used to store programs and data. It is understood that the computer storage medium herein may include both built-in storage media of the computer device and, of course, extended storage media supported by the computer device. Computer storage media provide storage space that stores an operating system for a computer device. Also stored in this memory space are one or more computer programs adapted to be loaded and executed by the processor 601. The computer storage medium may be a high-speed RAM memory, or may be a non-volatile memory (non-volatile memory), such as at least one disk memory; and optionally at least one computer storage medium located remotely from the processor.

In one embodiment, the computer storage medium may be loaded with one or more computer programs and executed by processor 601 to implement the corresponding steps of the image processing model training methods described above in fig. 1 and 3. In particular implementations, one or more computer programs in a computer storage medium are loaded by processor 601 and perform the steps of: acquiring an original training data set and an incremental training data set;

clustering a plurality of training data included in the original training data set and a plurality of training data included in the incremental training data set to obtain N cluster sets; each cluster set comprises one or more training data, and the one or more training data of each cluster set belong to the same data category; sampling and compressing the training data included in the N cluster sets to obtain a sampling training data set; training the pre-trained data processing model by using the sampling training data set to obtain a target data processing model; the pre-trained data processing model is trained based on the original training data set.

In one embodiment, when clustering the training data included in the original training data set and the training data included in the incremental training data set to obtain N cluster sets, one or more computer programs in the computer storage medium are loaded by the processor 601 and specifically perform the following steps:

In one embodiment, the clustering network is trained on a sample data set comprising a plurality of sample data and a data class label for each sample data, one or more computer programs in the computer storage medium being loaded by the processor 601 and performing the steps of:

In one embodiment, when determining the value of the contrast loss function according to the similarity metric characteristic between every two sample data and the data category label of each sample data, one or more computer programs in the computer storage medium are loaded by the processor 601 and specifically perform the following steps:

In an embodiment, the sample data set includes a first sample data and a second sample data, and when a contrast loss value between every two sample data is obtained according to the similarity metric characteristic between every two sample data and the data category label of each sample data, one or more computer programs in the computer storage medium are loaded by the processor 601 and specifically perform the following steps:

In one embodiment, when the training data included in the N cluster sets is subjected to sampling compression processing to obtain a sampled training data set, one or more computer programs in the computer storage medium are loaded by the processor 601 and specifically perform the following steps:

In an embodiment, the N cluster sets include a target cluster set, where the target cluster set is any one of the N cluster sets, and when the training data included in each cluster set is respectively compressed and sampled according to a target number of training data to be sampled in each cluster set, and the sampled training data is obtained from each cluster set, one or more computer programs in the computer storage medium are loaded by the processor 601 and specifically perform the following steps:

In one embodiment, the sampled training data set includes a first type of sampled training data subset and a second type of sampled training data subset, the training data in the first type of sampled training data subset being from the incremental training data set, the training data in the second type of sampled training data subset being from the original training data set; when the pre-trained data processing model is trained using the sample training dataset to obtain a target data processing model, one or more computer programs in the computer storage medium are loaded by the processor 601 and specifically perform the following steps:

In one embodiment, the pre-trained data processing model includes a bias correction module and other data processing modules, and when the pre-trained data processing model is alternatively trained by using the sampling training data set and a second type of sampling training data subset in the sampling training data set to obtain the target data processing model, one or more computer programs in the computer storage medium are loaded by the processor 601 and specifically execute the following steps:

Embodiments of the present application further provide a computer program product, where the computer program product includes a computer program, where the computer program is stored in a computer-readable storage medium, and when executed by a processor of a computer device, the computer program performs: acquiring an original training data set and an incremental training data set; clustering a plurality of training data included in the original training data set and a plurality of training data included in the incremental training data set to obtain N cluster sets; each cluster set comprises one or more training data, and the one or more training data of each cluster set belong to the same data category; sampling and compressing the training data included in the N cluster sets to obtain a sampling training data set; training the pre-trained data processing model by using the sampling training data set to obtain a target data processing model; the pre-trained data processing model is trained based on the original training data set.

It will be understood by those skilled in the art that all or part of the processes of the methods of the embodiments described above can be implemented by a computer program, which can be stored in a computer-readable storage medium, and when executed, can include the processes of the embodiments of the methods described above. The storage medium may be a magnetic disk, an optical disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), or the like.

While the invention has been described with reference to a preferred embodiment, it will be understood by those skilled in the art that various changes in form and detail may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.

Claims

1. An image processing model training method, comprising:

acquiring an original training data set and an incremental training data set;

2. The method of claim 1, wherein clustering the plurality of training data included in the original training data set and the plurality of training data included in the incremental training data set to obtain N cluster sets comprises:

3. The method of claim 2, wherein the clustering network is trained based on a sample data set comprising a plurality of sample data and a data class label for each sample data, the method further comprising:

4. The method of claim 3, wherein determining the value of the contrast loss function according to the similarity metric characteristic between each two sample data and the data class label of each sample data comprises:

5. The method of claim 4, wherein the sample data set comprises a first sample data and a second sample data, and the deriving the value of the loss of contrast between each two sample data according to the similarity metric between each two sample data and the data class label of each sample data comprises:

6. The method of claim 1, wherein the performing sample compression processing on the training data included in the N cluster sets to obtain a sampled training data set comprises:

7. The method according to claim 6, wherein the N cluster sets include a target cluster set, the target cluster set is any one of the N cluster sets, and the obtaining of the sampled training data from each cluster set by performing the compressed sampling processing on the training data included in each cluster set according to the target number of the training data to be sampled in each cluster set comprises:

calculating the quantity ratio of the training data in each cluster subset;

performing compression sampling processing on each clustering subset based on the quantity ratio of each clustering subset and the target quantity of training data to be sampled in the target clustering set, and acquiring sampling training data from each clustering subset;

8. The method of claim 1, wherein the sampled training data set comprises a first subset of sampled training data from the incremental training data set and a second subset of sampled training data from the original training data set;

the training of the pre-trained data processing model by using the sampling training data set to obtain a target data processing model comprises the following steps:

9. The method of claim 8, wherein the pre-trained data processing model comprises a bias correction module and other data processing modules, and the alternately training the pre-trained data processing model with the subset of the sampled training data set and the second type of sampled training data in the sampled training data set to obtain the target data processing model comprises:

10. An image processing model training apparatus, comprising:

11. A computer device, characterized in that,

a processor adapted to implement one or more computer programs;

a computer storage medium storing one or more computer programs adapted to be loaded by the processor and to perform the image processing model training method of any one of claims 1-9.

12. A computer storage medium, characterized in that the computer storage medium stores a computer program which, when being executed by a processor, is adapted to load and carry out the image processing model training method according to any one of claims 1 to 9.

13. A computer product or computer program, characterized in that the computer program product comprises a computer program which, when being executed by a processor, is adapted to load and carry out the image processing model training method according to any of the claims 1-9.