A subtask-based image sentiment label distribution learning method

Abstract

This invention provides a subtask-based image sentiment label distribution learning method to address the dependence of existing auxiliary task-based label distribution learning methods on expert knowledge. The method includes image sentiment data collection, structured feature preprocessing, construction of a subtask label distribution set, and a learning method based on shallow / deep models considering the subtasks. Compared to existing technologies, this invention subdivides the label distribution into multiple subtask label distributions by learning a subtask label space mask matrix, generating pseudo-supervisory information containing new knowledge. By constructing the subtask label distributions and treating them as auxiliary tasks for joint learning, this invention effectively improves the accuracy of image sentiment label distribution prediction. This invention can be widely applied in fields such as mental health monitoring, sentiment computing, and human-computer interaction, providing a reliable technical solution for complex sentiment recognition.

Description

Technical Field

[0001] This invention relates to the field of image sentiment prediction, belonging to the category of computer vision, and particularly to the learning of labeled distributions of image sentiment. Background Technology

[0002] With the rapid development of social media and human-computer interaction technologies, image emotion recognition technology has been widely applied in fields such as mental health monitoring, emotion computing, and human-computer interaction. Traditional image emotion recognition methods mainly rely on coarse feature extraction, including edge detection, gray-level co-occurrence matrix, or local binary pattern features, and depend on rudimentary single-label annotation. With the rapid development of neural network methods, automated feature extraction has been widely used in emotion recognition tasks; however, single-label methods often face challenges when dealing with complex emotional states.

[0003] In recent years, label distribution learning, as an emerging machine learning paradigm, has been gradually applied in sentiment recognition. Unlike traditional single-label classification, label distribution learning assigns label relevance to sample examples, becoming an effective solution to the problem of label ambiguity. Specifically, in image sentiment prediction tasks, label distribution learning can handle the complex distribution of sentiment labels across multiple emotions, providing more abundant supervised information to improve the information richness and accuracy of image sentiment prediction results.

[0004] Despite this, labeled distribution learning still faces the dilemma of insufficient data in real-world applications, as acquiring high-quality labeled distributions is both time-consuming and labor-intensive. To address the data shortage problem in real-world applications, auxiliary task learning offers a solution. Specifically, it introduces domain-specific expert knowledge and prior knowledge, as well as additional data from similar domains, as auxiliary tasks to guide the learning of the main task. Typically, when both the main and auxiliary tasks are learned together, the model can learn more discriminative and robust representations, thus improving performance. However, auxiliary tasks may constrain the model's generalization performance; due to its dependence on domain-specific expert knowledge and prior knowledge, the model can only solve specific problems within a single domain. It is understood that some existing multi-label learning methods partition the label space to construct sub-tasks, which are only dependent on the main task itself. Furthermore, transferring this construction method to labeled distribution learning tasks can generate pseudo-supervised information containing new knowledge, not only compensating for the data scarcity problem of existing technologies but also guiding the learning of the main task.

[0005] Therefore, this invention regards the sub-tasks obtained by dividing the label space as auxiliary tasks, explores a label distribution learning method based on sub-tasks, and applies it to the label distribution prediction of image sentiment. Summary of the Invention

[0006] Purpose of the invention: To alleviate the problem of existing auxiliary task-based label distribution learning methods' dependence on expert knowledge, this invention provides a subtask-based image sentiment label distribution learning method. By constructing an efficient subtask-based label distribution learning framework, it generates pseudo-supervised information containing new knowledge, making up for the problem of data scarcity in existing technologies, guiding the model to learn more discriminative representations, thereby improving the accuracy and robustness of sentiment prediction.

[0007] Technical solution: To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0008] A subtask-based image sentiment label distribution learning method includes the following steps:

[0009] Collect a labeled distribution dataset containing various image emotions, and construct the emotion label distribution using visual images as the feature part;

[0010] The feature portion of the labeled distribution dataset is preprocessed, including image cropping and data normalization, to generate a structured feature representation;

[0011] Learn the subtask label space mask matrix to subdivide the label distribution into multiple subtask label distributions to generate pseudo-supervisory information containing new knowledge;

[0012] The label distribution estimator is used to learn the label distribution of subtasks, and the structured features and the set of all subtask label distributions are concatenated as features to train a model corresponding to the label space of the main task.

[0013] In other embodiments, after the subtask label distribution set is constructed, a deeper model can also be used for learning: the structured feature representation is mapped to obtain a subspace representation, the subspace representation and the predicted subtask label distribution set are linked, and the label distribution learner is trained based on the linking result.

[0014] Furthermore, the above methods can be applied to various variant tasks of label distribution learning, including classification-oriented label distribution learning, incomplete label distribution learning, and label augmentation.

[0015] Furthermore, the learning subtask label space mask matrix, which subdivides the label distribution into multiple subtask label distributions to generate pseudo-supervisory information containing new knowledge, includes:

[0016] Simultaneously considering the information diversity and local ignorance of the subtask label space, a subtask label space mask matrix is ​​constructed and learned.

[0017] A partition is created based on the subtask label space mask matrix, and the original label distribution matrix is ​​sliced ​​based on the partition;

[0018] The descriptive power of the subtask label distribution is normalized to ensure that the subtask label distribution has a probabilistic simplex constraint.

[0019] Furthermore, the subtask label space mask matrix M∈{0,1} is obtained by solving the following problem. T×L Where T is the expected number of subtasks, and L is the number of sentiment tags; the m-th row of matrix M t , represents the mask corresponding to the label space in the possible t-th subtask, and the matrix elements 1 and 0 represent whether the label space of the subtask contains the label or not, respectively;

[0020]

[0021] Where N is the sample size, λ is the equilibrium hyperparameter, and d i Let represent the label distribution corresponding to sample i in the dataset, T represents the transpose operation, and ||·|| represents the modulo operation of the vector.

[0022] In some embodiments, when predicting the sentiment label distribution of a new image sample, the sample to be predicted is concatenated with all the predicted subtask label distributions to obtain the concatenated data to be predicted, and the predicted main task label distribution is obtained using an estimator.

[0023] In some embodiments, when predicting the sentiment label distribution of a new image sample, the sample to be predicted is input into a subspace representation operator to obtain a subspace representation, the subspace representation is input into a subtask label distribution prediction operator to obtain a subtask label distribution, the subspace representation and all predicted subtask label distributions are connected, and the predicted main task label distribution is obtained using a label distribution learner.

[0024] The types of learners involved include linear transformation matrices, nonlinear models, and neural networks based on encoder structures.

[0025] Furthermore, the predicted results and true values ​​of the label distribution can be reduced in dimensionality using the implicit Dirichlet allocation algorithm and mapped to the color space by the sigmoid mathematical operation. The structured features or subspace representations can be reduced in dimensionality using the t-distribution random nearest neighbor embedding algorithm, so that different label distributions can present different colors in the two-dimensional space obtained by the algorithm.

[0026] A computer program product includes a computer program / instructions that, when executed by a processor, implement the steps of the above-described subtask-based image sentiment label distribution learning method.

[0027] Beneficial effects: This invention uses sub-tasks as auxiliary tasks, reducing the dependence of existing methods on expert knowledge and thus enhancing the generalization ability of the model; the pseudo-supervisory information containing new knowledge generated by this invention, namely the sub-task label distribution, can expand existing data, thereby enhancing the model's adaptability under data-scarce conditions; this invention guides representation learning with sub-tasks, making the model less prone to overfitting; in addition, this invention has "plug-and-play" capability, is more independent than existing technologies, and can be extended to derivative tasks of label distribution learning. Attached Figure Description

[0028] Figure 1 This is a flowchart of a method according to an embodiment of the present invention.

[0029] Figure 2 This is a schematic diagram of the learning method based on a shallow model that considers sub-tasks in an embodiment of the present invention.

[0030] Figure 3 This is a schematic diagram of the deep model-based learning method considering sub-tasks in an embodiment of the present invention. Detailed Implementation

[0031] To better understand the working principle and application of the present invention, and to further elaborate on the present invention, the implementation details of the present invention will be described in detail below with reference to specific drawings and embodiments.

[0032] The abbreviations and key terms used in this embodiment are defined as follows:

[0033] (1) Label Distribution Learning (LDL).

[0034] (2) Local Binary Pattern (LBP).

[0035] (3) Stochastic Gradient Descent (SGD).

[0036] (4) Latent Dirichlet Allocation (LDA).

[0037] (5) t-distributed stochastic neighbor embedding (t-SNE).

[0038] (6) Kullback-Leibler Divergence (KLD).

[0039] This invention discloses a subtask-based image sentiment label distribution learning method, abbreviated as S-LDL. The overall technical solution is as follows: Figure 1 As shown, this includes image sentiment data collection, structured feature preprocessing, construction of subtask label distribution sets, and learning methods based on shallow models that consider subtasks / learning methods based on deep models that consider subtasks. These will be introduced one by one below.

[0040] Step S1. Image sentiment data collection: Collect a labeled distribution dataset containing various image sentiments, and construct the sentiment label distribution using visual images as feature parts.

[0041] Specifically, this step involves collecting labeled distribution datasets containing various types of data that can evoke human emotions, using visual images as the feature components, and defining the original image feature matrix. Where N is the number of samples, and for ease of data processing, it is assumed that the width and height of the image are both P, and C is the number of channels in the image. A sentiment label distribution is constructed based on mass outsourcing. The sentiment labels corresponding to this distribution include happiness, sadness, anger, surprise, disgust, fear, and neutrality. The sentiment label distribution of the entire sample is defined as matrix D. i Let represent the label distribution corresponding to sample number i in the dataset. Let x be the descriptive degree of the example x with respect to the label y. The label distribution satisfies the constraints of the probability simplex, including the constraint that the sum is one and the non-negativity constraint, i.e., d i ∈Δ L-1 Where L is the number of sentiment markers, and the probability simplex Δ has an L-1 dimension. L-1 Defined in the following mathematical form:

[0042]

[0043] The dataset can be sourced from social media, online surveys, or sentiment databases. Each sample in the dataset undergoes a mass outsourcing process involving a certain number of people, followed by quantitative evaluation based on the sentiment markers. The transformation from the mass outsourcing results to the facial sentiment marker distribution matrix takes various forms, typically a normalization function.

[0044] Step S2. Structured feature processing: Preprocess the feature parts of the labeled distribution dataset, including image cropping and data normalization, to generate structured feature representations.

[0045] Specifically, this step preprocesses the feature portion of the dataset, including locating and cropping specific objects that evoke emotions, and uses the LBP algorithm to extract Q-dimensional feature vectors from the original images. A normalization function is then used to process the obtained feature vectors to obtain a structured feature matrix.

[0046] Step S3. Construct the subtask label distribution set.

[0047] The method learns a subtask label space mask matrix, subdivides the label distribution into multiple subtask label distributions, and generates pseudo-supervisory information containing new knowledge, ultimately promoting the method to learn more discriminative and robust representations. Specifically, it includes the following steps.

[0048] Step S301: Learn the subtask label space: Define M∈{0,1} T×L Let M be the subtask mask matrix, where T is the expected number of subtasks. The t-th row of matrix M, i.e., m... t , represents the mask corresponding to the label space in the possible t-th subtask, and the matrix elements 1 and 0 represent whether the label is contained in the label space of the subtask or not.

[0049] To ensure the knowledge generated in the sub-task label space is rich and reliable, and to avoid homogeneous pseudo-supervision information, it is necessary to reasonably control the ignorance of the sub-label distribution space. To address this challenge, this step uses the label distribution matrix to assign weights to the mask and considers the pairwise similarity between rows of the mask matrix. Simultaneously, the exponential trick is applied to transform the solution of M into a minimization problem, as shown in the following mathematical form:

[0050]

[0051] Where λ is the balancing hyperparameter. In the problem objective of the above equation, the first term avoids the ignorance of the sub-label distribution space, and the second term encourages the diversity of the sub-label distribution space. When optimizing the solution, it is important to ensure that each element M of matrix M is M0. tj The following constraints must be satisfied:

[0052] M tj ∈{0,1}; t=1,…,T; j=1,…,L

[0053] The problem described above is slightly more complex than common integer programming problems. For convenience, this step uses the SGD algorithm for solution. To ensure that the matrix constraints are satisfied, the result of this step will undergo final mathematical processing via the sigmoid function, with a threshold set. The default value is 0.95. If the value is higher than this threshold, the matrix elements are set to 1; otherwise, the matrix elements are set to 0.

[0054] To perform subsequent operations, the sub-tag space set is defined as follows: and initialized to Furthermore, the set of subtask label distribution matrices is defined as follows: and initialized to

[0055] Step S302: Slice the original label distribution matrix: Create a partition based on the subtask label space mask matrix, and slice the original label distribution matrix based on the partition. Specifically, traverse each row of matrix M, denoted by the traversal index t, and define... For a possible t-th sub-label space, if there exists a sentiment label index j such that M tj =1, then y j Add to collection

[0056] This step also requires checking for invalid subtask masks. (That is, the subtask is completely identical to the main task) or If (i.e., the subtask label distribution cannot be constructed), skip the remaining operations on the mask for that row.

[0057] Step S303 normalizes the descriptors of the subtask label distribution to ensure that the subtask label distribution has a probabilistic simplex constraint. To avoid mathematical errors in subsequent calculations, a clip operation needs to be performed on each descriptor of the original label distribution matrix. The clip(x,a,b) function restricts the input x to the range [a,b]. Define D (t) Label the distribution matrix for the possible t-th subtask, specifically, that is, all subtasks that satisfy... clip(d) ·j ,ε,1), where d ·j Let be the j-th column of matrix D.

[0058] After the clip operation, the descriptive properties of the subtask label distribution are normalized, and the matrix D is traversed. (t) For each row, let the traversal index be i. Then the normalization steps can be expressed in the following mathematical form:

[0059]

[0060] Among them, normalization operation The definition can be in the following mathematical form:

[0061]

[0062] in,[·] j Let j be the j-th element of the vector being operated on.

[0063] At the end of each row traversal of matrix M, the subtask label spaces and subtask label distribution matrices that were not successfully preserved due to the aforementioned illegal checks are summarized and merged, i.e., execution is performed. as well as

[0064] Finally, the module returns a set of subtask tag spaces. and the set of subtask label distribution matrices The time complexity of the entire module is O(n).

[0065] The construction of the subtask label distribution set in this step does not rely on additional data from domains similar to the dataset, expert knowledge from any domain, or prior knowledge.

[0066] Step S4. The joint learning of the main task label distribution and the sub-task label distribution can be achieved using a learning method based on a shallow model or a learning method based on a deep model.

[0067] The following describes a shallow model-based learning method that considers subtasks. In short, it utilizes a shallow label distribution estimator and simple connection operations to quickly model and jointly learn the label distributions of the main task and subtasks. The overall process is as follows: Figure 2 As shown, specifically, it can be divided into the following steps:

[0068] Step S401: Learning the Subtask Label Distribution Estimator: Selecting a shallow label distribution estimator, totaling... There are several estimators, which can be the same type of labeled distribution learning algorithm or different types, such as SA-IIS, LDSVR, AA-kNN, etc. Initialize the model parameters of each estimator.

[0069] Learn the subtask label distribution matrix corresponding to different subtask label spaces, and then... Traverse the distribution of all subtask labels in the dataset, denoted by the traversal index t, and construct a new dataset {X,D}. (t)}, and train the t-th estimator on this dataset, denoted as f. (t) For a new sample x′ to be predicted, the predicted subtask label distribution is obtained using the estimator, i.e.:

[0070] d (t)′ ←f (t) (x′)

[0071] By the end of the traversal, the sample to be predicted has a corresponding predicted subtask label distribution for all subtasks.

[0072] Step S402 involves the structured feature X and the distribution set D of all subtask labels. (t) (in, Connect the links to obtain the linked data Z.

[0073] Step S403: Learn the merged label distribution estimator: Use the connected data Z as features to construct a new dataset {Z,D}, and train the model f corresponding to the main task label space.

[0074] Step S404: Predict the label distribution of unknown samples: Predict the sentiment label distribution of new image samples, and connect the sample to be predicted x′ with all predicted subtask label distributions d. (t)′ (in, ), obtain the linked data z′ to be predicted, and use the estimator to obtain the distribution of the main task label for prediction, that is:

[0075] d′←f(z′)

[0076] Finally, the module returns the predicted main task label distribution d′ as the module output.

[0077] The following describes a deep model-based learning method that considers sub-tasks, specifically using deep learning techniques to jointly learn the main task label distribution and the sub-task label distribution. The overall process is as follows: Figure 3 As shown, specifically, this module includes the following steps:

[0078] Step S411 Mapping Subspace Representation: Mapping the structured feature representation to obtain a subspace representation, defining... The subspace representation mapping operator, whose parameters are guided by the joint learning of the subtask and the main task, makes the representation more discriminative and robust. Specifically, this step can be expressed in the following mathematical form:

[0079]

[0080] Define ψ(·) as the subtask label distribution prediction operator, which can obtain the label distribution of each subtask based on the input representation, i.e., in the following mathematical form:

[0081]

[0082] The subtask label distribution is learned by minimizing the subtask learning loss, which is weighted by the sum of descriptiveness grouped by label, thus giving more attention to more reliable label spaces during optimization. Specifically, the subtask learning loss is defined as follows:

[0083]

[0084] in, These represent the actual values ​​of the main task descriptiveness, the actual values ​​of the subtask descriptiveness, and the predicted values ​​of the subtask descriptiveness, respectively.

[0085] Step S412 involves subspace representation R and the predicted subtask label distribution set. Make a connection, specifically, Z = (R, ψ(R)).

[0086] Step S413 trains the label distribution learner based on the connection results. The label distribution learner is defined as ω(·). This learner can be any existing label distribution learning model (provided that the model can be theoretically expressed as a network structure).

[0087] Step S414: Predict the label distribution of unknown samples: Predict the sentiment label distribution of new image samples, i.e.

[0088] Finally, the module returns the predicted main task label distribution matrix. As module output.

[0089] The final predicted values ​​and the true results of the label distribution can be reduced in dimensionality using the LDA algorithm and mapped to the color space using the sigmoid mathematical operation. The enhanced structured features obtained by the structured feature enhancement module can be reduced in dimensionality using the t-SNE algorithm, so that different label distributions can present different colors in the two-dimensional space obtained by the t-SNE algorithm.

[0090] This invention can be implemented using the deep learning open-source tool TensorFlow 2.0. The LDL loss function is defined as follows:

[0091]

[0092] The specific implementation of l(·,·) is KLD, that is, the following mathematical form:

[0093]

[0094] Note that the final implementation of this invention can be applied to various variant tasks of labeled distribution learning, such as classification-oriented labeled distribution learning, incomplete labeled distribution learning, and label augmentation. The following details how to modify the invention to adapt to these various variant tasks.

[0095] For label distribution learning tasks oriented towards classification, the construction of subtasks still depends on the original label distribution matrix, while the main task loss is replaced by a task-specific loss. Its specific definition is beyond the scope of this specification and will not be elaborated upon here. The degenerate label distribution matrix is ​​defined as follows:

[0096]

[0097] in,[·] ij The element with index j corresponding to the i-th row of the matrix being operated on.

[0098] For incomplete label distribution learning tasks, subtask construction depends on the label distribution matrix with limited observable range. in The operation is defined as follows:

[0099]

[0100] in,[·] ij The definition is the same as above. The main task loss is replaced with a task-specific loss for learning the incomplete labeled distribution. Specific definitions are beyond the scope of this manual and will not be elaborated here.

[0101] For the label augmentation task, the subtask construction depends on the logical label matrix L, and the underlying theme is similar to the above schemes, which will not be elaborated here.

[0102] All learners involved in the steps of this invention can be represented as any learning model, including any linear transformation matrix, nonlinear model, and neural network based on encoder-decoder structure. In this example, a neural network based on encoder-decoder structure is used to implement the deep model-based learning method of this invention, and based on this, image sentiment label distribution learning is performed.

[0103] Using TensorFlow 2.0, we load the sentiment label distribution dataset from the network and perform a label distribution learning task, treating the LDL loss function as the main task loss. Define the parameters of the entire model as Θ, and train the model using the following loss:

[0104]

[0105] Where α is the balancing hyperparameter of the subtask.

[0106] This concludes the training phase. Through the training method described above, this invention can effectively improve the accuracy of sentiment label distribution prediction in various data distribution environments, providing a more accurate and reliable sentiment analysis tool for related applications.

[0107] This invention also discloses a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the aforementioned subtask-based image sentiment tag distribution learning methods. The program / instruction code for implementing the methods of this invention can be written in any combination of one or more programming languages. This program / instruction code can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program / instruction code causes the steps of the methods of this invention to be implemented. The program / instruction code can be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server. All aspects not detailed in this invention are well-known to those skilled in the art.

[0108] The above are embodiments of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. Various variations and modifications can be made to the present invention without departing from its spirit and scope, and all such variations and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A subtask-based image sentiment label distribution learning method, characterized in that, Includes the following steps: Collect a labeled distribution dataset containing various image emotions, and construct the emotion label distribution using visual images as the feature part; The feature portion of the labeled distribution dataset is preprocessed, including image cropping and data normalization, to generate a structured feature representation; The method involves learning a subtask label space mask matrix to subdivide the label distribution into multiple subtask label distributions, thereby generating pseudo-supervised information containing new knowledge. This includes: simultaneously considering the information diversity and local ignorance of the subtask label space; constructing and learning a subtask label space mask matrix; creating partitions based on the subtask label space mask matrix; slicing the original label distribution matrix based on these partitions; and normalizing the descriptive degree of the subtask label distribution to ensure that the subtask label distribution has probabilistic simplex constraints. The label distribution estimator is used to learn the label distribution of subtasks, and the structured features and the set of all subtask label distributions are concatenated as features to train a model corresponding to the label space of the main task. The subtask label space mask matrix is ​​obtained by solving the following problem. ,in, This represents the expected number of subtasks. The number of sentiment tags; matrix The OK , representing the possible first The mask corresponding to the label space in the subtask, where matrix elements 1 and 0 represent whether the label space of the subtask contains or does not contain the label, respectively; ； in, For the sample size, To balance hyperparameters, Indicates that the data is numbered The label distribution corresponding to the samples, This indicates the transpose operation. Represents the modulo operation of vectors.

2. A subtask-based image sentiment label distribution learning method, characterized in that, Includes the following steps: Collect a labeled distribution dataset containing various image emotions, and construct the emotion label distribution using visual images as the feature part; The feature portion of the labeled distribution dataset is preprocessed, including image cropping and data normalization, to generate a structured feature representation; The method involves learning a subtask label space mask matrix to subdivide the label distribution into multiple subtask label distributions, thereby generating pseudo-supervised information containing new knowledge. This includes: simultaneously considering the information diversity and local ignorance of the subtask label space; constructing and learning a subtask label space mask matrix; creating partitions based on the subtask label space mask matrix; slicing the original label distribution matrix based on these partitions; and normalizing the descriptive degree of the subtask label distribution to ensure that the subtask label distribution has probabilistic simplex constraints. The structured feature representation is mapped to obtain a subspace representation. The subspace representation and the predicted subtask label distribution set are linked together, and the label distribution learner is trained based on the linking result. The subtask label space mask matrix is ​​obtained by solving the following problem. ,in, This represents the expected number of subtasks. The number of sentiment tags; matrix The OK , representing the possible first The mask corresponding to the label space in the subtask, where matrix elements 1 and 0 represent whether the label space of the subtask contains or does not contain the label, respectively; ； in, For the sample size, To balance hyperparameters, Indicates that the data is numbered The label distribution corresponding to the samples, This indicates the transpose operation. Represents the modulo operation of vectors.

3. The image sentiment marker distribution learning method based on subtasks according to claim 1 or 2, characterized in that, The method is applied to various variants of label distribution learning tasks, including classification-oriented label distribution learning, incomplete label distribution learning, and label augmentation.

4. The image sentiment label distribution learning method based on subtasks according to claim 1, characterized in that, When predicting the sentiment label distribution for new image samples, the sample to be predicted is concatenated with all the predicted subtask label distributions to obtain the concatenated data to be predicted. The estimator is then used to obtain the predicted main task label distribution.

5. The image sentiment marker distribution learning method based on subtasks according to claim 2, characterized in that, When predicting the sentiment label distribution of a new image sample, the sample to be predicted is input into the subspace representation operator to obtain the subspace representation, the subspace representation is input into the subtask label distribution prediction operator to obtain the subtask label distribution, the subspace representation and all predicted subtask label distributions are connected, and the predicted main task label distribution is obtained using the label distribution learner.

6. The image sentiment marker distribution learning method based on subtasks according to claim 2, characterized in that, The types of learners involved include linear transformation matrices, nonlinear models, and neural networks based on encoder structures.

7. The subtask-based image sentiment tag distribution learning method according to claim 1 or 2, characterized in that, The predicted results and true values ​​of the labeled distribution can be reduced in dimensionality using the implicit Dirichlet allocation algorithm and mapped to the color space by the sigmoid mathematical operation. The structured features or subspace representations can be reduced in dimensionality using the t-distribution random nearest neighbor embedding algorithm, so that different labeled distributions can present different colors in the two-dimensional space obtained by the algorithm.

8. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1-7.

Images