CN108229552B - A model processing method, device and storage medium - Google Patents

Abstract

The present invention provides a model processing method, comprising: extracting a first object feature from a source library and extracting a second object feature from a target library, wherein the first object feature and the second object feature are respectively The sample sets in the respective databases; the first object feature and the second object feature are respectively mapped to the same feature space, and the mapped first object feature and the mapped second object feature are obtained correspondingly; according to the mapping After the first object feature and the second object feature after the mapping, determine the transformation matrix when the difference between the source library and the target library satisfies the preset minimum difference condition; The mapped first object feature is converted to obtain a third object feature; the third object feature and the corresponding label are applied to a model for classifying samples in the target library. The present invention also provides a model processing device and a storage medium.

Description

A model processing method, device and storage medium

technical field

The present invention relates to electric digital data processing technology, in particular to a model processing method, device and storage medium.

Background technique

With the advent of the era of big data, people can obtain a large amount of data more easily. In addition, due to the continuous development of the field of machine learning, how to make computers have the ability to draw inferences from others, and how to make large amounts of data play a better role, these problems have changed. In order to solve these problems, transfer learning has been proposed and received more and more attention.

There is an important assumption in conventional machine learning that the samples of the source library and the samples of the target library must have the same distribution or come from the same feature space, but in real life, this assumption is difficult to realize; specifically , for a classification problem, if the samples of the source library and the samples of the target library do not have the same distribution (that is, it can be roughly understood that the source does not belong to the same library), it can be understood that the source library and the target library do not have the same distribution Feature space: In the field of image recognition, a model trained from an image library can get better recognition results when applied to this image library, but it is often unsatisfactory when applied to other image libraries or in real environments.

To sum up, since the source library and the target library in the prior art do not have the same feature space, the model trained from the source library cannot obtain good recognition results when applied to the target library.

SUMMARY OF THE INVENTION

In view of this, the embodiments of the present invention are expected to provide a model processing method, apparatus, and storage medium, which can overcome the problem that the accuracy of the model is affected by the difference in the feature space of the source library and the target library.

In order to achieve the above-mentioned purpose, the technical scheme of the embodiment of the present invention is realized as follows:

An embodiment of the present invention provides a model processing method, and the method includes:

Extracting the first object feature from the source library and extracting the second object feature from the target library, respectively, wherein the first object feature and the second object feature are sample sets in the respective databases;

The first object feature and the second object feature are respectively mapped to the same feature space, and the mapped first object feature and the mapped second object feature are obtained correspondingly;

determining, according to the mapped first object feature and the mapped second object feature, a transformation matrix when the difference between the source library and the target library satisfies a preset minimum difference condition;

Converting the mapped first object feature through the transformation matrix to obtain a third object feature;

The third object feature and the corresponding label are applied to the model for classifying the samples in the target library.

An embodiment of the present invention further provides a model processing device, the device includes: an extraction module, a mapping module, a determination module, a conversion module, and an application module; wherein,

The extraction module is used to extract the first object feature from the source library and the second object feature from the target library respectively, wherein the first object feature and the second object feature are samples in the respective databases respectively gather;

The mapping module is used to map the first object feature and the second object feature to the same feature space respectively, and correspondingly obtain the mapped first object feature and the mapped second object feature;

The determining module is configured to determine, according to the mapped first object feature and the mapped second object feature, when the difference between the source library and the target library satisfies a preset minimum difference condition The transformation matrix of ;

The conversion module is configured to convert the mapped first object feature through the transformation matrix to obtain a third object feature;

The application module is configured to apply the third object feature and the corresponding label to a model for classifying samples in the target library.

An embodiment of the present invention further provides a storage medium on which an executable program is stored, and when the executable program is executed by a processor, any one of the foregoing model processing methods is implemented.

An embodiment of the present invention further provides a model processing apparatus, including a memory, a processor, and an executable program stored in the memory and capable of being run by the processor, and the processor executes any one of the foregoing when running the executable program a model processing method.

The model processing method, device, and storage medium provided by the embodiments of the present invention respectively extract a first object feature from a source library and a second object feature from a target library, wherein the first object feature and the second object feature are The object features are respectively the sample sets in the respective databases; the first object feature and the second object feature are respectively mapped to the same feature space, and the mapped first object feature and the mapped second object feature are obtained correspondingly; According to the mapped first object feature and the mapped second object feature, determine the transformation matrix when the difference between the source library and the target library satisfies the preset minimum difference condition; The transformation matrix transforms the mapped first object features to obtain third object features; and applies the third object features and corresponding labels to a model for classifying samples in the target library. In this way, a model suitable for the target library is trained through the source library. Therefore, when the model trained from the source library is applied to the target library, a good recognition result can be obtained, thereby improving the accuracy of the model.

Description of drawings

FIG. 1 is a schematic diagram of an implementation flowchart of a model processing method provided by an embodiment of the present invention;

2 is a schematic flowchart of a specific implementation of a model processing method provided by an embodiment of the present invention;

3 is a schematic diagram of face calibration provided by an embodiment of the present invention;

4 is a schematic diagram of a Gabor space provided by an embodiment of the present invention;

5 is a schematic diagram of the structure of the AlexNet image classification model of deep learning provided by an embodiment of the present invention;

FIG. 6 is a schematic diagram of the composition and structure of a model processing apparatus provided by an embodiment of the present invention;

FIG. 7 is a schematic diagram of a hardware structure of a model processing apparatus provided by an embodiment of the present invention.

Detailed ways

In order to be able to understand the features and technical contents of the embodiments of the present invention in more detail, the implementation of the embodiments of the present invention will be described in detail below with reference to the accompanying drawings. The accompanying drawings are for reference only and are not used to limit the present invention.

FIG. 1 is a model processing method provided by an embodiment of the present invention; as shown in FIG. 1 , the implementation process of the model processing method in the embodiment of the present invention may include the following steps:

Step 101: Extract the first object feature from the source library and the second object feature from the target library, respectively, wherein the first object feature and the second object feature are sample sets in the respective databases.

In some embodiments, the first object feature is extracted from the source library and the second object feature is extracted from the target library, wherein different types of object features can be extracted from the source library, and the extracted different types of object features can be combined. Then, it is used as the first object feature of the source library; different types of object features can be extracted from the target library, and the extracted object features of different types can be combined to serve as the second object feature of the target library.

Then, dimensionality reduction is performed on the first object feature extracted from the source library and the second object feature extracted from the target library; the specific process of dimensionality reduction may include: the first object feature extracted from the source library, and The second object features extracted from the target library are respectively normalized; the specific object features with preset dimensions are respectively extracted from the first object features and the second object features, which correspond to the to-be-mapped objects in the source library and the target library. Object characteristics. Wherein, the preset dimension is smaller than the corresponding dimension before extraction; respectively extracting the specific object feature of the preset dimension from the first object feature and the second object feature may include: from the feature vector corresponding to the first object feature, and the third In the feature vectors corresponding to the two object features, the object features corresponding to the largest d-dimensional feature values are respectively selected as the respective specific object features.

Step 102: Map the first object feature and the second object feature to the same feature space respectively, and obtain the mapped first object feature and the mapped second object feature correspondingly.

In some embodiments, in order to avoid the loss of information caused by directly mapping the source library and the target library to the same feature space, the first object feature and the second object feature are respectively multiplied by a transformation matrix and projected to their corresponding Therefore, through spatial mapping, the mapped first object feature and the mapped second object feature located in the same dimension space are obtained.

Step 103 : According to the mapped first object feature and the mapped second object feature, determine a transformation matrix when the difference between the source library and the target library satisfies a preset minimum difference condition.

In some embodiments, a difference function representing the distance between the source library and the target library may be constructed based on the factored first object feature after mapping and the second object feature after mapping; in a way of finding the minimum value of the difference function, Determine the transformation matrix when the difference between the source library and the target library satisfies the preset minimum difference condition. Here, the product of the mapped first object feature and the transformation matrix, and the difference between the mapped second object feature and the mapped second object feature may be used as a factor; a function for calculating a norm based on the factor is constructed.

Step 104: Convert the mapped first object feature through the transformation matrix to obtain a third object feature.

In some embodiments, the transformation matrix may convert the subspace coordinate system where the mapped first object feature is located into the mapped second object by aligning the mapped first object feature with the mapped second object feature The subspace coordinate system where the feature is located, and the third object feature is obtained.

Step 105: Apply the third object feature and the corresponding label to a model for classifying samples in the target library.

In some embodiments, model training can be performed on the third object feature and the corresponding label, specifically: taking the third object feature and the corresponding label as a new sample; solving the update value component of the model parameter relatively new sample; The samples in the target library extract the object features, and calculate the probability values of the samples in the target library with different labels based on the model; select the labels that meet the probability conditions as the labels of the samples in the target library.

In some embodiments, the samples in the target library may be classified as neighbors based on a similarity function with the third object feature as a factor, so as to obtain the labels of the samples in the target library.

In some embodiments, when the applied models include at least two types, the output results for the samples in the target library are respectively determined based on each model; the output results of each model are compared, and the samples in the target library are determined according to the comparison results. Tag of.

For example, when the output results of each model are the same, select the output result of any one model as the label of the samples in the target library; when the output results of each model are different and do not include a specific label, select the output result of any one model. As the label of the sample in the target library; when the output results of each model are different and include a specific label, select the specific label as the label of the sample in the target library.

The specific implementation process of the model processing method according to the embodiment of the present invention will be further described in detail below.

FIG. 2 shows a schematic diagram of a specific implementation flow of the model processing method according to the embodiment of the present invention; as shown in FIG. 2 , the model processing method according to the embodiment of the present invention includes the following steps:

Step 201: Extract the first object feature from the source library and extract the second object feature from the target library, respectively, wherein the first object feature and the second object feature are sample sets in the respective databases.

In the embodiment of the present invention, taking the extraction of the facial expression feature S from the human face pictures in the source library and the extraction of the facial expression feature T from the human face pictures in the target library as examples, here, it is possible to extract the facial expression features from the source library. The facial expression feature S is understood as the first object feature, and the facial expression feature T extracted from the target library is understood as the second object feature.

For example, different types of facial expression features can be extracted from the source library, and the extracted facial expression features of different types can be combined as the facial expression feature S of the source library; different types of facial expression features can be extracted from the target library. The facial expression feature, after combining the extracted facial expression features of different types, is used as the facial expression feature T of the target library.

For the source library and the target library, the following two methods can be used to extract different types of facial expression features.

1) Gabor feature g can be extracted:

For example, you can use the face alignment Intraface processing tool for face calibration. Intraface can calibrate 49 points of people's eyebrows, eyes, nose and mouth, and 49 points can be very good for facial features. , a specific example can be seen in Figure 3; when calibrating a face picture, other processing tools can also be used, and the number of calibration points used by different processing tools is different, and the embodiment of the present invention is not limited to using 49 points Intraface processing tool.

Then, feature sampling can be performed on the calibrated points to obtain facial expression features; for example, after calibration, Gabor features can be extracted from the calibrated part, which can be used to distinguish people such as anger, disgust, fear, happiness, sadness and surprise. The basic emotion expressed by the face; among them, the Gabor feature is a feature that can be used to describe the texture information of the image. The frequency and direction of the Gabor filter are similar to the human visual system, and it is especially suitable for texture representation and discrimination.

The Gabor features represent the matrices B _s and B _t of the face image through convolution with the Gabor filter to obtain the Gabor features S _g and T _g respectively;

B _s *g=S _g ; B _t *g=T _g ;

The filter is calculated as follows:

where λ represents the wavelength, θ represents the rotation angle, ψ represents the phase offset, σ represents the standard deviation of the Gaussian function, γ represents the spatial ratio, and x and y represent the pixel points; it can be seen from the above formula that the Gabor filter can be divided into a real part R and the imaginary number I, after trigonometric transformation, we can get:

Gabor滤波器的实数部分可以看作是各个方向的边缘检测算子，使用这个特性，可以得到Gabor空间，具体可以参见图4。in

The real part of the Gabor filter can be regarded as the edge detection operator in all directions. Using this feature, the Gabor space can be obtained. For details, see Figure 4.

2) Deep learning features can be extracted for the entire face. Convolutional neural networks can be regarded as a special type of feedforward network. These models are designed to imitate the behavior of the visual cortex. Convolutional neural networks CNN are used in visual recognition tasks. Performing well, the CNN has special layers called convolutional and pooling layers that allow the network to encode certain image attributes, through which the CNN features can be extracted: S _c and T _c .

Among them, the deep learning image classification model AlexNet consists of 11 layers of CNN with the following architecture, as shown in Figure 5 for details.

The Gabor feature S _g extracted from the source library and the convolutional neural network CNN feature S _c are combined in series to form a new feature as the object feature of the source library, that is, the facial expression feature S; from the target library The extracted Gabor feature T _g and the convolutional neural network CNN feature T _c are combined in series to form a new feature as the object feature of the target library, that is, the facial expression feature T, namely:

The facial expression features of the face pictures in the source library are: S=[S _g , S _c ];

The facial expression features of the face pictures in the target library are: T=[T _g ,T _c ];

Among them, the number of lines (the number of pictures) of S _g and S _c are the same, and similarly, the number of lines of T _g and T _c are also the same.

Traditional features can better describe the details of the face, while deep learning features can better describe the overall features of the face. Therefore, through the combination of these two features, the facial expression can be more effectively described.

Step 202: Dimensionality reduction is performed on the first object feature extracted from the source library and the second object feature extracted from the target library.

In this embodiment of the present invention, the first object feature extracted from the source library and the second object feature extracted from the target library may be respectively normalized.

For example, taking the facial expression feature S extracted from the source library and the facial expression feature T extracted from the target library for dimensionality reduction as an example; the facial expression feature S extracted from the source library, and The facial expression features T extracted from the target library are respectively normalized. Since the dimensions of S and T are the same, both S and T exist in a given D-dimensional space. In order to obtain a more robust To get a representation of the properties, and to be able to get the difference between the two image libraries, you can convert both S and T into a D-dimensional normalized vector (eg zero mean and unit standard deviation).

In this embodiment of the present invention, specific object features with preset dimensions may be extracted from the first object feature and the second object feature, respectively, which correspond to the object features to be mapped in the source library and the target library. Wherein, the preset dimension is smaller than the corresponding dimension before extraction; respectively extracting the specific object feature of the preset dimension from the first object feature and the second object feature may include: from the feature vector corresponding to the first object feature, and the third In the feature vectors corresponding to the two object features, the object features corresponding to the largest d-dimensional feature values are respectively selected as the respective specific object features.

For example, Kernel Principal Component Analysis (KPCA) can be used to select the largest d-dimensional eigenvalues for the eigenvectors of the source library and the target library, and a Gaussian kernel is used as the kernel.

Step 203: Map the first object feature and the second object feature to the same feature space respectively, and obtain the mapped first object feature and the mapped second object feature correspondingly.

In the embodiment of the present invention, in order to avoid the loss of information caused by directly mapping the source library and the target library to the same feature space, the first object feature and the second object feature are respectively multiplied by a transformation matrix and projected to their corresponding Therefore, through spatial mapping, the mapped first object feature and the mapped second object feature located in the same dimension space are obtained.

其中I_d是d的单位矩阵；为了避免直接将源库和目标库映射到同一特征空间所造成的信息损失，可以将人脸表情特征S和人脸表情特征T采用分别乘以一个转换矩阵A_s,A_t∈R^1×D的方式，投影到各自对应的子空间，从而，通过空间映射，得到位于同一维度空间的映射后的人脸表情特征X_s和映射后的人脸表情特征X_t。For example, the feature vector is used as the basis for the source library and the target library, which are denoted as X _s and X _t (X _s , X _t ∈ R ^D×d ), respectively. Note that X _s and X _t are orthogonal, so

where I _d is the identity matrix of d; in order to avoid the loss of information caused by directly mapping the source library and the target library to the same feature space, the facial expression feature S and the facial expression feature T can be multiplied by a transformation matrix A respectively. _s , A _t ∈ R ^1×D , project to their corresponding subspaces, thus, through spatial mapping, the mapped facial expression feature X _s and the mapped facial expression feature X in the same dimensional space are obtained _t .

Step 204: Determine a transformation matrix when the difference between the source library and the target library satisfies a preset minimum difference condition according to the mapped first object feature and the mapped second object feature.

In this embodiment of the present invention, a difference function representing the distance between the source library and the target library may be constructed based on the first object feature after mapping and the second object feature after mapping as factors; in a way of finding the minimum value of the difference function, Determine the transformation matrix when the difference between the source library and the target library satisfies the preset minimum difference condition. Here, the product of the mapped first object feature and the transformation matrix, and the difference between the mapped second object feature and the mapped second object feature may be used as a factor; a function for calculating a norm based on the factor is constructed.

For example, based on the above-mentioned mapped facial expression features X _s and X _t as factors, a difference function representing the distance between the facial expression feature X _s and the facial expression feature X _t can be constructed; Bregman Matrix Divergence (Bregman Matrix Divergence) Bregman, determines the corresponding transformation matrix M.

To achieve this goal, the spatial alignment method can be used to align the reference vectors by using a transformation matrix M from X _s to X _t , which is learned by Bregman:

M ^* = argmin _M (F(M)) (5)

是Frobenius规范；由于X_s和X_t是由第d个特征矢量生成的，事实上它们趋于内在正则化，由于Frobenius范数对于正交操作是不变的，因此，可以将公式(4)重写成如下形式：in,

is the Frobenius norm; since X _s and X _t are generated by the d-th eigenvector, in fact they tend to be intrinsically regularized, and since the Frobenius norm is invariant to orthogonal operations, Equation (4) Rewrite as follows:

计算得到，如果源库和目标库是相同的，则X_s＝X_t，且M^*为单位矩阵。It can be seen from formula (6) that the optimal M ^* can be given by

It is calculated that if the source library and the target library are the same, then X _s =X _t , and M ^* is the identity matrix.

Step 205: Transform the mapped first object feature through a transformation matrix to obtain a third object feature.

In the embodiment of the present invention, the transformation matrix may convert the subspace coordinate system where the mapped first object feature is located into the mapped second object by aligning the mapped first object feature with the mapped second object feature The subspace coordinate system where the feature is located, and the third object feature is obtained.

For example, the above-mentioned transformation matrix M converts the source image space coordinate system to the target image space coordinate system by aligning the facial expression feature X _s after mapping in the source library and the facial expression feature X _t after mapping in the target library. system, by multiplying the mapped facial expression feature X _s by the transformation matrix M, the converted facial expression feature is obtained; if the mapped facial expression feature X _s in the source library and the mapped facial expression feature in the target library If the facial expression feature X _t is orthogonal, it can be ignored; when the mapped facial expression feature X _s in the source library is well aligned with the mapped facial expression feature X _t in the target library, it is given high weight.

Step 206: Apply the third object feature and the corresponding label to the model for classifying the samples in the target library.

In the embodiment of the present invention, model training can be performed on the third object feature and the corresponding label, specifically: the third object feature and the corresponding label are used as new samples; the update value components of the relatively new samples of model parameters are solved; The samples in the target library extract the object features, and calculate the probability values of the samples in the target library with different labels based on the model; select the labels that meet the probability conditions as the labels of the samples in the target library.

For example, a support vector SVM model can be selected, or other types of models such as neural networks can be selected for training: take the converted facial expression features and corresponding labels as new samples, and put them into the SVM model for training; The update value component of the samples with relatively new model parameters; extract facial expression features from the samples in the target database, and calculate the probability values of the samples in the target database with different labels based on the model; select the labels that meet the probability conditions as the labels of the samples in the target database.

In the embodiment of the present invention, the samples in the target library may be classified into neighbors based on the similarity function with the third object feature as a factor, so as to obtain the labels of the samples in the target library.

For example, taking the K-nearest neighbor criterion model applied to classify the samples in the target database as an example, in order to compare the A _s corresponding to the facial expression feature S in the source database with the A _t corresponding to the facial expression feature S in the target database. For comparison, a metric similarity function Sim(A _s , _{At )} is needed. Projecting _{As and At into their respective subspaces X s and X t and} applying the optimal transformation matrix M ^* , the nearest neighbor classification is aided by this similarity function, where the similarity function can be defined as follows:

是编码向量的不同分量在其原始空间中的相对贡献。in,

is the relative contribution of the different components of the encoded vector in its original space.

By transforming the source image with the optimal M ^* obtained by learning, the similarity function Sim can be directly used as the measurement tool of the nearest neighbor classifier KNN. The KNN here does not need model training, and can be directly classified; assuming that the source library There are 1-6 expression categories in , and there are 10 expression features in each category. Then, after substituting any sample in the target library into the Sim formula, expression categories 1-6 get a value respectively, and take the largest value (ie The label corresponding to the most similar) is used as the expression label of the sample; wherein, the value of each expression classification is determined by the average value obtained by substituting the 10 expression features it contains.

In the embodiment of the present invention, when the applied model includes two types, here, it is assumed that the applied model is the K-nearest neighbor criterion model and the SVM model, and the output results for the samples in the target library are respectively determined based on these two models; The output results of the two models are compared, and the labels of the samples in the target library are determined according to the comparison results.

For example, when the output results of the above two models are the same, select the output result of any one model as the expression label in the target library; when the output results of the above two models are different and do not include specific expression labels, select any expression label. The output result of a model is used as the expression label in the target library. Here, the specific expression label can include fear, sadness and disgust; when the output results of the above two models are different and include a specific label, the specific label is selected as the target library. Emoji tab, here, you can choose sadness, disgust and fear in order of priority.

This judgment method is adopted because the emoji tags happy and surprised are relatively strong positive emotions, which are easy to distinguish in the trained model; while the intuitively similar emoji tags such as fear, sadness, and disgust have only obtained relatively general emotions. Recognition rate, when recognizing expressions such as fear, sadness, and disgust, it is easy to be recognized as positive emotions such as happiness. Therefore, the probability of error is high, which can alleviate the tendency of the classification model. Through the above migration method, the accuracy of the test results is much higher than the accuracy of the model trained using the source library directly used in the target library for testing.

To implement the above method, an embodiment of the present invention further provides a model processing device. As shown in FIG. 6 , the device includes an extraction module 601, a mapping module 602, a determination module 603, a conversion module 604, and an application module 605; wherein,

The extraction module 601 is used to extract the first object feature from the source library and the second object feature from the target library respectively, wherein the first object feature and the second object feature are respectively in the respective databases. sample collection.

The extraction module 601 is specifically configured to: extract different types of object features from the source library, and combine the extracted different types of object features as the first object features of the source library;

Different types of object features are extracted from the target library, and the extracted different types of object features are combined to serve as the second object features of the target library.

The mapping module 602 is configured to map the first object feature and the second object feature to the same feature space respectively, and correspondingly obtain the mapped first object feature and the mapped second object feature.

The determining module 603 is configured to determine, according to the mapped first object feature and the mapped second object feature, that the difference between the source library and the target library satisfies a preset minimum difference condition transformation matrix when .

The determining module 603 is specifically configured to: construct a difference representing the distance between the source library and the target library based on the mapped first object feature and the mapped second object feature as factors function;

A transformation matrix when the difference between the source library and the target library satisfies a preset minimum difference condition is determined in the manner of finding the minimum value of the difference function.

The determining module 603 is specifically configured to: use the product of the mapped first object feature and the transformation matrix, and the difference between the mapped second object feature and the mapped second object feature as a factor; construct a calculation based on the factor norm function.

The transformation module 604 is configured to transform the mapped first object feature through the transformation matrix to obtain a third object feature.

The application module 605 is configured to apply the third object feature and the corresponding label to a model for classifying samples in the target library.

The application module 605 is specifically configured to: take the third object feature and the corresponding label as a new sample; solve the update value component of the model parameter relative to the new sample; extract the object feature from the sample in the target library , and calculate the probability value that the samples in the target library have different labels based on the model; select the labels that meet the probability conditions as the labels of the samples in the target library.

The application module 605 is specifically configured to: perform neighbor classification on the samples in the target library based on the similarity function with the third object feature as a factor, and obtain the labels of the samples in the target library.

The apparatus further includes a dimension reduction module 606, configured to: perform dimension reduction on the first object feature extracted from the source library and the second object feature extracted from the target library.

The dimensionality reduction module 606 is specifically configured to: respectively perform normalization processing on the first object feature extracted from the source library and the second object feature extracted from the target library;

The specific object features of preset dimensions are respectively extracted, and the preset dimensions are smaller than the corresponding dimensions before extraction.

The apparatus further includes a comparison module 607, configured to: when the applied models include at least two types, determine the output results for the samples in the target library respectively based on each model;

The output results of the various models are compared, and the labels of the samples in the target library are determined according to the comparison results.

The comparing module 607 is specifically configured to: when the output results of the models are the same, select the output result of any model as the label of the sample in the target library;

When the output results of the models are different and do not include a specific label, select the output result of any model as the label of the sample in the target library;

When the output results of the models are different and include a specific label, the specific label is selected as the label of the sample in the target library.

In practical applications, the extraction module 601 , the mapping module 602 , the determination module 603 , the conversion module 604 , the application module 605 , the dimension reduction module 606 and the comparison module 607 can all be implemented by a central processing unit (CPU, Central Processing Processing) located on a computer device. Unit), Microprocessor (MPU, Micro Processor Unit), Digital Signal Processor (DSP, Digital Signal Processor) or Field Programmable Gate Array (FPGA, Field Programmable GateArray) etc.

It should be noted that: when the model processing apparatus provided in the above embodiment performs model processing, only the division of the above-mentioned program modules is used as an example for illustration. In practical applications, the above-mentioned processing can be allocated to different program modules as required. That is, the internal structure of the device is divided into different program modules to complete all or part of the processing described above. In addition, the model processing apparatus and the model processing method embodiments provided by the above embodiments belong to the same concept, and the specific implementation process thereof is detailed in the method embodiments, which will not be repeated here.

In order to implement the above model processing method, an embodiment of the present invention further provides a hardware structure of a model processing apparatus. The model processing apparatus for implementing the embodiments of the present invention will now be described with reference to the accompanying drawings. The model processing apparatus may be implemented by a terminal device, such as a computer device such as a smart phone, a tablet computer, and a palmtop computer. The hardware structure of the model processing apparatus provided by the embodiments of the present invention will be further described below. It can be understood that FIG. 7 only shows an exemplary structure of the model processing apparatus but not the entire structure. full structure.

Referring to FIG. 7, FIG. 7 is a schematic diagram of the hardware structure of a model processing apparatus provided by an embodiment of the present invention, which can be applied to the aforementioned terminal equipment running an application program in practical applications. The model processing apparatus 700 shown in FIG. 7 includes: at least one A processor 701 , a memory 702 , a user interface 703 and at least one network interface 704 . The various components in the model processing apparatus 700 are coupled together through a bus system 705 . It will be understood that the bus system 705 is used to implement the connection communication between these components. In addition to the data bus, the bus system 705 also includes a power bus, a control bus and a status signal bus. However, for clarity of illustration, the various buses are labeled as bus system 705 in FIG. 7 .

The user interface 703 may include a display, a keyboard, a mouse, a trackball, a click wheel, keys, buttons, a touch pad or a touch screen, and the like.

It will be appreciated that memory 702 may be either volatile memory or non-volatile memory, and may include both volatile and non-volatile memory.

The memory 702 in the embodiment of the present invention is used for storing various types of data to support the operation of the model processing apparatus 700 . Examples of these data include: any computer program for operating on the model processing apparatus 700 , such as an executable program 7021 and an operating system 7022 , and the program implementing the model processing method of the embodiment of the present invention may be included in the executable program 7021 .

The model processing method disclosed in the embodiment of the present invention may be applied to the processor 701 or implemented by the processor 701 . The processor 701 may be an integrated circuit chip with signal processing capability. In the implementation process, each step of the above-mentioned model processing method can be completed by an integrated logic circuit of hardware in the processor 701 or an instruction in the form of software. The above-mentioned processor 701 may be a general-purpose processor, a DSP, or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, and the like. The processor 701 may implement or execute each model processing method, steps, and logic block diagram provided in the embodiments of the present invention. A general purpose processor may be a microprocessor or any conventional processor or the like. In combination with the steps of the model processing method provided by the embodiments of the present invention, it can be directly embodied as being executed by a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software module may be located in a storage medium, and the storage medium is located in the memory 702. The processor 701 reads the information in the memory 702, and completes the steps of the aforementioned model processing method in combination with its hardware.

The embodiment of the present invention also provides a hardware structure of a model processing apparatus, the model processing apparatus 700 includes a memory 702, a processor 701, and an executable program 7021 stored in the memory 702 and executable by the processor 701, When the processor 701 runs the executable program 7021, it realizes:

Extracting the first object feature from the source library and extracting the second object feature from the target library respectively, wherein the first object feature and the second object feature are sample sets in the respective databases; The object feature and the second object feature are respectively mapped to the same feature space, and the mapped first object feature and the mapped second object feature are obtained correspondingly; according to the mapped first object feature and the mapped The second object feature is to determine the transformation matrix when the difference between the source library and the target library satisfies the preset minimum difference condition; the first object feature after mapping is converted by the transformation matrix, and the result is obtained The third object feature; the third object feature and the corresponding label are applied to the model for classifying the samples in the target library.

In some embodiments, when the processor 701 runs the executable program 7021, it realizes:

Extract different types of object features from the source library, and combine the extracted different types of object features as the first object features of the source library; extract different types of object features from the target library, and extract After combining the different types of object features, it is used as the second object feature of the target library.

In some embodiments, when the processor 701 runs the executable program 7021, it realizes:

Dimensionality reduction is performed on the first object feature extracted from the source library and the second object feature extracted from the target library.

In some embodiments, when the processor 701 runs the executable program 7021, it realizes:

The first object feature extracted from the source library and the second object feature extracted from the target library are respectively normalized; specific object features of preset dimensions are respectively extracted, and the preset dimension The number is less than the corresponding dimension before extraction.

In some embodiments, when the processor 701 runs the executable program 7021, it realizes:

Based on the first object feature after mapping and the second object feature after mapping as factors, construct a difference function representing the distance between the source library and the target library; to solve the minimum value of the difference function way of determining the transformation matrix when the difference between the source library and the target library satisfies a preset minimum difference condition.

In some embodiments, when the processor 701 runs the executable program 7021, it realizes:

Using the product of the mapped first object feature and the transformation matrix, and the difference between the mapped second object feature and the mapped second object feature as a factor; constructing a function for calculating a norm based on the factor.

In some embodiments, when the processor 701 runs the executable program 7021, it realizes:

Taking the third object feature and the corresponding label as a new sample; solving the update value component of the model parameter relative to the new sample; extracting the object feature from the sample in the target library, and calculating the target library based on the model The probability value of the samples in the target library with different labels; select the labels that meet the probability conditions as the labels of the samples in the target library.

In some embodiments, when the processor 701 runs the executable program 7021, it realizes:

Based on the similarity function with the third object feature as a factor, the samples in the target library are classified as neighbors to obtain the labels of the samples in the target library.

In some embodiments, when the processor 701 runs the executable program 7021, it realizes:

When the applied models include at least two types, the output results for the samples in the target library are determined based on each model; the output results of the various models are compared, and the samples in the target library are determined according to the comparison results. Tag of.

In some embodiments, when the processor 701 runs the executable program 7021, it realizes:

When the output results of the models are the same, select the output results of any one model as the label of the samples in the target library; when the output results of the models are different and do not include specific labels, select the output results of any one model. The output result is used as the label of the sample in the target library; when the output results of the models are different and include a specific label, the specific label is selected as the label of the sample in the target library.

An embodiment of the present invention further provides a storage medium, where the storage medium may be a storage medium such as an optical disc, a flash memory, or a magnetic disk, and may optionally be a non-transitory storage medium. Wherein, an executable program 7021 is stored on the storage medium, and the executable program 7021 is implemented when executed by the processor 701:

Extracting the first object feature from the source library and extracting the second object feature from the target library respectively, wherein the first object feature and the second object feature are sample sets in the respective databases; The object feature and the second object feature are respectively mapped to the same feature space, and the mapped first object feature and the mapped second object feature are obtained correspondingly; according to the mapped first object feature and the mapped The second object feature is to determine the transformation matrix when the difference between the source library and the target library satisfies the preset minimum difference condition; the first object feature after mapping is converted by the transformation matrix, and the result is obtained The third object feature; the third object feature and the corresponding label are applied to the model for classifying the samples in the target library.

In some embodiments, when the executable program 7021 is executed by the processor 701, it realizes:

Extract different types of object features from the source library, and combine the extracted different types of object features as the first object features of the source library; extract different types of object features from the target library, and extract After combining the different types of object features, it is used as the second object feature of the target library.

In some embodiments, when the executable program 7021 is executed by the processor 701, it realizes:

Dimensionality reduction is performed on the first object feature extracted from the source library and the second object feature extracted from the target library.

In some embodiments, when the executable program 7021 is executed by the processor 701, it realizes:

The first object feature extracted from the source library and the second object feature extracted from the target library are respectively normalized; specific object features of preset dimensions are respectively extracted, and the preset dimension The number is less than the corresponding dimension before extraction.

In some embodiments, when the executable program 7021 is executed by the processor 701, it realizes:

Based on the first object feature after mapping and the second object feature after mapping as factors, construct a difference function representing the distance between the source library and the target library; to solve the minimum value of the difference function way of determining the transformation matrix when the difference between the source library and the target library satisfies a preset minimum difference condition.

In some embodiments, when the executable program 7021 is executed by the processor 701, it realizes:

Using the product of the mapped first object feature and the transformation matrix, and the difference between the mapped second object feature and the mapped second object feature as a factor; constructing a function for calculating a norm based on the factor.

In some embodiments, when the executable program 7021 is executed by the processor 701, it realizes:

Taking the third object feature and the corresponding label as a new sample; solving the update value component of the model parameter relative to the new sample; extracting the object feature from the sample in the target library, and calculating the target library based on the model The probability value of the samples in the target library with different labels; select the labels that meet the probability conditions as the labels of the samples in the target library.

In some embodiments, when the executable program 7021 is executed by the processor 701, it realizes:

Based on the similarity function with the third object feature as a factor, the samples in the target library are classified as neighbors to obtain the labels of the samples in the target library.

In some embodiments, when the executable program 7021 is executed by the processor 701, it realizes:

When the applied models include at least two types, the output results for the samples in the target library are determined based on each model; the output results of the various models are compared, and the samples in the target library are determined according to the comparison results. Tag of.

In some embodiments, when the executable program 7021 is executed by the processor 701, it realizes:

When the output results of the models are the same, select the output results of any one model as the label of the samples in the target library; when the output results of the models are different and do not include specific labels, select the output results of any one model. The output result is used as the label of the sample in the target library; when the output results of the models are different and include a specific label, the specific label is selected as the label of the sample in the target library.

To sum up, in the model processing method, device, and storage medium provided by the embodiments of the present invention, the first object feature is extracted from the source library and the second object feature is extracted from the target library, wherein the first object feature The first object feature and the second object feature are respectively mapped to the same feature space, and the mapped first object feature and the mapped first object feature are obtained correspondingly. second object feature; according to the mapped first object feature and the mapped second object feature, determine the transformation when the difference between the source library and the target library satisfies a preset minimum difference condition Matrix; transform the mapped first object feature through the transformation matrix to obtain a third object feature; apply the third object feature and the corresponding label to a method for classifying samples in the target library Model. In this way, a model suitable for the target library is trained through the source library. Therefore, when the model trained from the source library is applied to the target library, a good recognition result can be obtained, thereby improving the accuracy of the model.

As will be appreciated by those skilled in the art, embodiments of the present invention may be provided as a method, system, or executable program product. Accordingly, the invention may take the form of a hardware embodiment, a software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of an executable program product embodied on one or more computer-usable storage media having computer-usable program code embodied therein, including but not limited to disk storage, optical storage, and the like.

The present invention is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and executable program products according to embodiments of the invention. It will be understood that each flow and/or block in the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by executable program instructions. These executable program instructions may be provided to a general purpose computer, a special purpose computer, an embedded processor or a processor of a reference programmable data processing apparatus to produce a machine such that the instructions executed by the computer or a processor of a reference programmable data processing apparatus produce a Means for implementing the functions specified in a flow or flow of a flowchart and/or a block or blocks of a block diagram.

The executable program instructions may also be stored in a computer-readable memory capable of directing a computer or reference programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory result in an article of manufacture comprising instruction means, the The instruction means implement the functions specified in the flow or flow of the flowcharts and/or the block or blocks of the block diagrams.

These executable program instructions may also be loaded onto a computer or reference programmable data processing apparatus, such that a series of operational steps are performed on the computer or reference programmable apparatus to produce a computer-implemented process for execution on the computer or reference programmable apparatus The instructions provide steps for implementing the functions specified in one or more of the flowcharts and/or one or more blocks of the block diagrams.

The above are only preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention shall be included in the within the protection scope of the present invention.

Claims (22)

1. a model processing method, is characterized in that, described method comprises: Extracting the first object feature from the source library and extracting the second object feature from the target library, respectively, wherein the first object feature and the second object feature are sample sets in the respective databases; The first object feature and the second object feature are respectively mapped to the same feature space, and the mapped first object feature and the mapped second object feature are obtained correspondingly; determining, according to the mapped first object feature and the mapped second object feature, a transformation matrix when the difference between the source library and the target library satisfies a preset minimum difference condition; Converting the mapped first object feature through the transformation matrix to obtain a third object feature; Apply the third object feature and the corresponding label to the model for classifying the samples in the target library; when the applied model includes at least two types, if the output of each model in the at least two models is If the results are different and include a specific tag, the specific tag is selected as the tag of the sample in the target library; wherein, in the case that the specific tag includes multiple tags, based on each tag in the multiple tags The priorities determine the labels of the samples in the target library. 2. The method according to claim 1, wherein the extracting the first object feature from the source library and the extracting the second object feature from the target library respectively comprise: Extract different types of object features from the source library, and combine the extracted different types of object features as the first object features of the source library; Different types of object features are extracted from the target library, and the extracted different types of object features are combined to serve as the second object features of the target library. 3. The method according to claim 1, wherein the method further comprises: Before mapping the first object feature and the second object feature to the same feature space, respectively, Dimensionality reduction is performed on the first object feature extracted from the source library and the second object feature extracted from the target library. 4. The method according to claim 3, wherein the performing dimension reduction on the first object feature extracted from the source library and the second object feature extracted from the target library comprises: normalizing the first object feature extracted from the source library and the second object feature extracted from the target library; The specific object features of preset dimensions are respectively extracted, and the preset dimensions are smaller than the corresponding dimensions before extraction. 5 . The method according to claim 1 , wherein, according to the mapped first object feature and the mapped second object feature, determining the difference between the source library and the target library. 6 . The transformation matrix when the difference of satisfies the preset minimum difference condition, including: Constructing a difference function representing the distance between the source library and the target library based on taking the mapped first object feature and the mapped second object feature as factors; A transformation matrix when the difference between the source library and the target library satisfies a preset minimum difference condition is determined in the manner of finding the minimum value of the difference function. 6. The method according to claim 5, wherein the constructing a difference function representing the distance between the source library and the target library comprises: Using the product of the mapped first object feature and the transformation matrix, and the difference between the mapped second object feature as a factor; Build a function that computes the norm based on the factors. 7. The method according to claim 1, wherein the applying the third object feature and the corresponding label to a model for classifying samples in the target library comprises: Taking the third object feature and the corresponding label as a new sample; solving for the updated value components of the model parameters relative to the new sample; Extract object features from the samples in the target library, and calculate the probability values that the samples in the target library have different labels based on the model; A label that meets the probability condition is selected as the label of the sample in the target library. 8. The method according to claim 1, wherein the applying the third object feature and the corresponding label to a model for classifying samples in the target library comprises: Based on the similarity function with the third object feature as a factor, the samples in the target library are classified as neighbors to obtain the labels of the samples in the target library. 9. The method according to claim 7 or 8, wherein the method further comprises: When the applied model includes at least two types, Determine the output results for the samples in the target library based on each model; The output results of the various models are compared, and the labels of the samples in the target library are determined according to the comparison results. 10. The method according to claim 9, wherein the comparing the output results of the various models, and determining the labels of the samples in the target library according to the comparison results, comprising: When the output results of each model are the same, select the output result of any model as the label of the sample in the target library; When the output results of the models are different and do not include a specific label, select the output result of any model as the label of the sample in the target library; When the output results of the models are different and include a specific label, the specific label is selected as the label of the sample in the target library. 11. A model processing device, characterized in that the device comprises: an extraction module, a mapping module, a determination module, a conversion module and an application module; wherein, The extraction module is used to extract the first object feature from the source library and the second object feature from the target library respectively, wherein the first object feature and the second object feature are samples in the respective databases respectively gather; The mapping module is used to map the first object feature and the second object feature to the same feature space respectively, and correspondingly obtain the mapped first object feature and the mapped second object feature; The determining module is configured to determine, according to the mapped first object feature and the mapped second object feature, when the difference between the source library and the target library satisfies a preset minimum difference condition The transformation matrix of ; The conversion module is configured to convert the mapped first object feature through the transformation matrix to obtain a third object feature; The application module is used to apply the third object feature and the corresponding label to a model for classifying samples in the target library; when the applied model includes at least two types, if at least two The output results of each model in the model are different and include a specific tag, then the specific tag is selected as the tag of the sample in the target library; wherein, in the case that the specific tag includes multiple tags, based on the multiple tags The priority of each of the labels determines the label of the sample in the target library. 12. The apparatus according to claim 11, wherein the extraction module is specifically configured to: Extract different types of object features from the source library, and combine the extracted different types of object features as the first object features of the source library; Different types of object features are extracted from the target library, and the extracted different types of object features are combined to serve as the second object features of the target library. 13. The apparatus according to claim 11, characterized in that, the apparatus further comprises a dimensionality reduction module for: Dimensionality reduction is performed on the first object feature extracted from the source library and the second object feature extracted from the target library. 14. The apparatus according to claim 13, wherein the dimensionality reduction module is specifically used for: normalizing the first object feature extracted from the source library and the second object feature extracted from the target library; The specific object features of preset dimensions are respectively extracted, and the preset dimensions are smaller than the corresponding dimensions before extraction. 15. The apparatus according to claim 11, wherein the determining module is specifically configured to: Constructing a difference function representing the distance between the source library and the target library based on taking the mapped first object feature and the mapped second object feature as factors; A transformation matrix when the difference between the source library and the target library satisfies a preset minimum difference condition is determined in the manner of finding the minimum value of the difference function. 16. The apparatus according to claim 15, wherein the determining module is specifically configured to: Using the product of the mapped first object feature and the transformation matrix, and the difference between the mapped second object feature as a factor; Build a function that computes the norm based on the factors. 17. The device according to claim 11, wherein the application module is specifically used for: Taking the third object feature and the corresponding label as a new sample; solving for the updated value components of the model parameters relative to the new sample; Extract object features from the samples in the target library, and calculate the probability values that the samples in the target library have different labels based on the model; A label that meets the probability condition is selected as the label of the sample in the target library. 18. The device according to claim 11, wherein the application module is specifically used for: Based on the similarity function with the third object feature as a factor, the samples in the target library are classified as neighbors to obtain the labels of the samples in the target library. 19. The device according to claim 17 or 18, wherein the device further comprises a comparison module for: When the applied model includes at least two types, Determine the output results for the samples in the target library based on each model; The output results of the various models are compared, and the labels of the samples in the target library are determined according to the comparison results. 20. The device according to claim 19, wherein the compared modules are specifically used for: When the output results of each model are the same, select the output result of any model as the label of the sample in the target library; When the output results of the models are different and do not include a specific label, select the output result of any model as the label of the sample in the target library; When the output results of the models are different and include a specific label, the specific label is selected as the label of the sample in the target library. 21. A storage medium on which an executable program is stored, characterized in that, when the executable program is executed by a processor, the model processing method according to any one of claims 1 to 10 is implemented. 22. A model processing device, comprising a memory, a processor, and an executable program stored in the memory and executable by the processor, characterized in that, when the processor executes the executable program, the processor executes the process as claimed in the claims The model processing method according to any one of 1 to 10.

Images