A method for constructing a domain adaptation model based on information selection

Abstract

This invention proposes a method for constructing a domain adaptation model based on information selection. This method addresses the problems of source model misprediction and interference of easily confused instances with target model training in existing methods by using semantic information selection based on the degree of association between each category and instance-level information selection based on the information content of instance classification prediction. Specifically, it performs efficient and accurate knowledge distillation through information selection to complete the target model training task when the source model parameters are not visible. This includes semantic information selection based on the degree of association between each category and instance-level information selection based on the information content of instance classification prediction. At the same time, it uses the confidence index of classification prediction to screen out easily confused instances for instance-level information selection to solve the performance degradation problem caused by this type of instance.

Description

Technical Field

[0001] This invention belongs to the domain adaptation branch of computer vision, specifically relating to a method for constructing a domain adaptation model based on information selection. Background Technology

[0002] With the rapid development of internet technology, all kinds of information permeate people's daily learning, work, and life, making the 21st century an era of data explosion. Supported by massive amounts of data, deep neural networks have achieved remarkable success in various fields such as classification, retrieval, and object detection. However, in real-life scenarios, the performance of deep learning often suffers due to the lack of sufficient labeled data. To address this issue, unsupervised domain adaptation (UDA) was proposed, aiming to process unlabeled target domain data using knowledge learned from labeled source domain data. Through methods such as knowledge transfer and domain alignment, unsupervised domain adaptation tasks have found wide applications in image classification, object detection, and semantic segmentation.

[0003] However, in real-life and work scenarios, people are paying increasing attention to data security. Therefore, in many situations, labeled source domain data is invisible during the training process of the target model. This means the target model can only obtain information from the source model to process the target domain data. This type of unsupervised domain adaptation task is called source-domain adaptation and has attracted increasing attention. Current research has seen many scholars attempting to filter information from the target domain data by introducing structured learning or constructing confidence scores, thus achieving knowledge transfer in scenarios where source domain data is invisible, and some progress has been made.

[0004] However, these methods still require complex training of the source model, and the training of the target domain model is a continuation of the training on the source model, making it impossible to achieve the goal of reducing the parameters of the target domain model, which is undesirable for users with limited resources. Furthermore, for security reasons, current models often only provide input and output application programming interfaces (APIs), making it difficult to learn based on the parameters of the source model. In this scenario, both the source domain data and the parameters of the source model are unknown. Therefore, research not only needs to address the distributional differences between the source and target domain data but also needs to consider how to obtain sufficiently accurate information from the source model using only the API.

[0005] In existing research, some scholars have introduced knowledge distillation in scenarios where the parameters of the source model are not visible. This involves constraining the output of the target model to be close to the output of the source model to achieve knowledge transfer. However, due to the distribution differences between the target and source domain data, the source model trained on the source domain data often makes incorrect predictions about the target domain data. Furthermore, because many categories of individual features are not obvious in the collected images, or because some images contain multiple individuals, the model struggles to provide a confident classification prediction for such image instances. Undoubtedly, both incorrect predictions from the source model and easily confused image instances significantly impact the training performance of the target model, leading to a decline in the final target model's performance.

[0006] In summary, current domain adaptation models in scenarios where source model parameters are not visible face the dual challenges of incorrect predictions from the source model and easily confused instances. When performing knowledge distillation, it is necessary to filter out erroneous semantic information and easily confused instance information. Summary of the Invention

[0007] In view of this, the present invention provides a method for constructing a domain adaptation model based on information selection to handle domain adaptation tasks when the source model parameters are not visible. By selecting semantic information based on the degree of association between each category and selecting instance-level information based on the information content of instance classification prediction, the present invention solves the problems of source model misprediction and interference of easily confused instances on target model training in existing methods. Specifically, efficient and accurate knowledge distillation is performed through information selection to complete the target model training task when the source model parameters are not visible, including semantic information selection based on the degree of association between each category and instance-level information selection based on the information content of instance classification prediction.

[0008] To address the above problems, the technical method of this invention employs a method for constructing a domain adaptation model based on information selection, comprising: preprocessing the target domain data, including image cropping and image flipping; constructing a domain adaptation model based on information selection, which consists of a source model and a target model. The source model is a supervised training model on the source domain data, whose parameters and model structure are not visible; input and output can only be completed through an interface, and the output is the source classification features. The target model consists of a feature extraction network and a classifier, and the final output is the target classification features. The source classification feature vector of the target domain data is extracted through the source model. Clustering of the source classification feature vector yields cluster centers for each category and classification predictions for the target domain data. A semantic similarity matrix is ​​constructed using the cluster centers. Confidence scores for each prediction are constructed using the classification predictions. The target classification feature vector of the target domain data is extracted through the target model. The target classification feature vector is saved until the end of the next training round. The source classification feature vector is reconstructed using the semantic similarity matrix, and the reconstructed source classification feature vector is named the source classification signal. The semantic distillation loss is calculated based on the source classification signal and the target classification feature vector. The self-distillation loss is calculated based on the target classification feature vector saved in the previous training round and the target classification feature vector in this training round. The semantic distillation loss and the self-distillation loss are then scaled at the instance level using the confidence score. The parameters of the target model are adjusted and updated based on the scaled losses to obtain the trained domain adaptation model.

[0009] The feature extraction network of the target model is constructed as a residual neural network ResNet-50. A randomly initialized 256-dimensional fully connected layer is added to the end of the ResNet-50 network, resulting in a final output vector with a dimension of 256. The classification network of the target model is a randomly initialized fully connected layer. The output of the feature extraction network is used as the input to the classifier, and the final output of the target model is the target classification feature with the dimension of the number of categories.

[0010] Further, source classification feature vectors of the target domain data are extracted using the source model. Clustering of these source classification feature vectors yields cluster centers for each category and classification predictions for the target domain data. A semantic similarity matrix is ​​constructed using the cluster centers, and confidence scores for each prediction are built using the classification predictions. Finally, target classification feature vectors of the target domain data are extracted using the target model. These target classification feature vectors will be saved until the end of the next training round, specifically including:

[0011] The preprocessed target domain data is input into the source model, and the output of the source model is called the source classification feature vector of the target domain data. The Gaussian Mixture Model (GMM) clustering algorithm is used to cluster the extracted source classification feature vectors to obtain the classification prediction for each input image instance. and the cluster centers for each category in, For classifying and predicting probabilities, f S For the source model, u i Let C represent the cluster center for each category in the training set, and N represent the number of categories in the training set. T Indicates the number of data in the target domain. This represents a data instance in the target domain, and GMM(*) indicates clustering using a Gaussian mixture model. Since cluster centers are obtained by clustering the source classification feature vectors of the target domain data, each cluster center can be considered a representative of its corresponding category. Therefore, the similarity between cluster centers can represent the similarity between categories. A semantic similarity matrix is ​​constructed using the cluster centers by calculating the cosine similarity between them. The i-th row and j-th column of the matrix represents the cosine similarity between the i-th and j-th cluster centers, signifying the semantic similarity between the i-th and j-th categories. The construction formula is as follows:

[0012]

[0013] Among them ||·|| 2 Let M represent the sum of squares, and M be the semantic similarity matrix. In this method, all data values ​​in the semantic similarity matrix are between 0 and 1, and therefore can be used as weights to scale the source classification feature vectors of the target domain data.

[0014] The confidence score for each instance is calculated through the classification prediction. The confidence score should be proportional to the degree of certainty of the model's prediction result. For instances that are not easily confused, the classification prediction obtained after clustering should yield a result with high certainty, i.e., a high value in one dimension and low values ​​in other dimensions. For instances that are easily confused, the classification prediction should have many dimensions with similar values. Based on this characteristic, this invention provides a method for calculating the confidence score:

[0015]

[0016] Where x T For target domain data, f T (x T Let H(t) be the target classification feature vector of the target domain data, σ(·) represent the softmax function, and H(t) = Σ-t i log(t iC represents the number of classes in the training dataset. Instances with high confidence scores have significant features, clear compositions, and similar shapes, thus facilitating model distillation training. Instances with low confidence scores have more or less interfering information and special shapes, which can interfere with the training results. Therefore, the confidence score, as a weight, can effectively filter all instances by their strength level, greatly reducing the influence of easily confused instances on the final training results.

[0017] This invention inputs target domain data into the feature extraction network of the target model, then inputs the output of the feature extraction network into the classifier of the target model, and finally outputs the target classification feature vector of the target data.

[0018] The target classification features will be saved in the next round of training and participate in the training of the target model in the next round.

[0019] Furthermore, the source classification feature vector is reconstructed using the semantic similarity matrix, and the reconstructed source classification feature vector is named the source classification signal. The reconstruction method involves retrieving the similarity between the category of each instance and other categories from the semantic similarity matrix, and scaling each dimension of the instance's source classification feature vector based on all similarities between the instance's category and other categories. The scaling method is a dot product of the similarity and the corresponding dimension of the source classification feature vector. The scaled source classification feature vector is named the source classification signal.

[0020] The source classification signal is calculated as follows:

[0021]

[0022] in This represents the source classification feature vector of the target model; * indicates that the values ​​in the matrices are multiplied together. Represents the source classification signal, and represents matrix multiplication. The source classification prediction probability obtained by clustering the target model using a Gaussian mixture model represents the target model's value. This represents the semantic similarity matrix.

[0023] Furthermore, semantic distillation loss and self-distillation loss are calculated based on the source classification signal, the target classification feature vector, and the target classification feature vector saved in the previous training round. These two losses are then scaled at the instance level using the confidence score. The parameters of the target model are adjusted and updated based on the scaled losses. Semantic distillation loss is calculated based on the source classification signal and the target classification feature vector. Semantic distillation loss is used to replace traditional classification loss for preliminary supervised training of the target model, endowing it with semantic information. Unlike traditional distillation loss, semantic distillation loss does not directly constrain the target classification feature vector and the source classification feature vector to conform to the same distribution characteristics. Instead, it constrains the target classification feature vector to align with the distribution characteristics of the source classification signal obtained after semantically filtering the source classification feature vector. Therefore, it avoids interference from incorrect predictions in the source classification feature vector and improves the efficiency of knowledge distillation. The semantic distillation loss is calculated as follows:

[0024]

[0025] Where x T For target domain data, f T (x T Let σ(·) be the target classification feature vector of the target domain data, and let σ(·) represent the softmax function. CE(·,·) represents the source classification signal of the target domain data, and CE(·,·) represents the cross-entropy loss.

[0026] The self-distillation loss is calculated based on the target classification feature vector saved in the previous training round and the target classification feature vector in this training round.

[0027] Furthermore, semantic distillation loss is calculated based on the source classification signal and the target classification feature vector. The semantic distillation loss is obtained by constraining the target classification feature vector to be close to the source classification signal, where the source classification signal does not participate in gradient updates. The self-distillation loss is obtained by constraining the target classification feature vector in this round to be similar to the target classification feature vector saved in the previous round. Self-distillation loss is not calculated for the first round of training. As training progresses, since the source model does not participate in gradient updates, its performance remains constant. In the initial stage, the source model's high performance can be aided by semantic distillation in training the target model. However, as the target model is continuously optimized, its performance may surpass that of the source model. In this case, self-distillation loss is needed to reduce the impact of insufficient source model performance on training. The formula for calculating self-distillation loss is as follows:

[0028]

[0029] Where x T For target domain data, Let σ(·) be the target classification feature vector of the target domain data in the k-th training round of the target model, and let σ(·) represent the softmax function. Let be the target classification feature vector of the target domain data in the (k-1)th round of training, and CE(·,·) represent the cross-entropy loss.

[0030] The two losses are then scaled at the instance level using the confidence scores. The final loss is as follows:

[0031]

[0032] Where λ is a hyperparameter used to balance semantic distillation loss and self-distillation loss.

[0033] Furthermore, the trained domain adaptation model generates classification results based on the image of the target classification result. The target model extracts the target classification feature vector from the image to be detected, and then clusters the target classification feature vector using the existing clustering algorithm Gaussian Mixture Model (GMM) to obtain the classification prediction result.

[0034] The improvement of this invention lies in that the domain adaptation model and method based on information selection is used to handle the domain adaptation problem when the parameters of the source model are not visible. It obtains cluster centers through data feature clustering, constructs a semantic similarity matrix through the cluster centers, and then performs semantic-level information selection on the semantic information in the source model classification signal through the semantic similarity matrix to solve the performance loss caused by the source model's incorrect classification prediction. At the same time, it uses the confidence index of classification prediction to screen out easily confused instances for instance-level information selection to solve the performance degradation problem caused by this type of instance. Attached Figure Description

[0035] Figure 1 This is a simplified flowchart of the method for constructing a domain adaptation model based on information selection according to the present invention;

[0036] Figure 2 This is a schematic diagram of the overall framework of the present invention;

[0037] Figure 3 This is an explanatory diagram of the effect of the confidence score in this invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings, so that those skilled in the art can better understand this invention. It should be particularly noted that the described embodiments are only some embodiments of this invention, not all embodiments, and are not intended to limit the scope of the claimed invention. All other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this invention.

[0039] Figure 1 This is a simplified flowchart of the method for constructing a domain adaptation model based on information selection proposed in this invention. The first step is data preprocessing. Since the data is collected from various sources, and due to limitations such as collection time and equipment, the formats of the various image instances in the data are not entirely the same. The first step is to unify the image format by converting all images to JPG format. Next, each image is randomly flipped horizontally or vertically with a probability of 0.5. Finally, all images are cropped to a size of 224*224 pixels.

[0040] Furthermore, after data preprocessing, the target domain data is processed separately using pre-designed source and target model networks. The source and target models are then used to extract source and target classification feature vectors from the target domain data. The overall structure of the domain adaptation model construction method based on information selection in this invention is as follows: Figure 2 As shown, it consists of a source model f S A target model f T The system comprises the following components: The source model is a supervised training model on source domain data. Its parameters and model structure are not visible; input and output are only completed through an interface, with the output being source classification features. The target model consists of a feature extraction network and a classifier, ultimately outputting target classification features. The source model extracts source classification feature vectors from the target domain data. Clustering of these source classification feature vectors yields cluster centers for each category and classification predictions for the target domain data. A semantic similarity matrix is ​​constructed using the cluster centers, and confidence scores for each prediction are generated. The target model then extracts target classification feature vectors from the target domain data. These target classification feature vectors are saved until the end of the next training round. The source feature vectors are reconstructed using the semantic similarity matrix, and the reconstructed source classification feature vectors are named the source classification signal. A semantic distillation loss is calculated based on the source classification signal and the target classification feature vectors. A self-distillation loss is calculated based on the target classification feature vectors saved in the previous training round and those in this training round. The two losses are then scaled at the instance level using the confidence scores. The parameters of the target model are adjusted and updated based on the scaled losses.

[0041] Furthermore, an information-selection-based domain adaptation model is constructed, consisting of a source model and a target model. The source model is a supervised model trained on source domain data; its parameters and structure are not visible, and input / output can only be completed through an interface, with the output being the source classification features. The target model consists of a feature extraction network and a classifier, ultimately outputting the target classification features.

[0042] The feature extraction network of the target model is constructed as a ResNet-50 residual neural network, with a randomly initialized 256-dimensional fully connected layer added at the end of the network, resulting in a final output vector of 256 dimensions. The classification network of the target model is a randomly initialized fully connected layer, using the output of the feature extraction network as input, and the final output dimension is the target classification feature of the number of categories.

[0043] Furthermore, the source classification feature vector of the target domain data is extracted through the source model. Clustering of the source classification feature vectors yields cluster centers for each category and classification predictions for the target domain data. A semantic similarity matrix is ​​constructed using the cluster centers, and confidence scores for each prediction are built using the classification predictions. Finally, the target classification feature vector of the target domain data is extracted through the target model. This target classification feature vector will be saved until the end of the next training round, including:

[0044] The preprocessed target domain data is input into the source model, and the output of the source model is called the source classification feature of the target domain data. The Gaussian Mixture Model (GMM) clustering algorithm is used to cluster the extracted source classification features to obtain the classification prediction for each instance. and the cluster centers for each category Since cluster centers are obtained by clustering the source classification features of the target domain data, each cluster center can be regarded as a representative of the category corresponding to that cluster center. Therefore, the similarity between the cluster centers can represent the similarity between the categories. A semantic similarity matrix is ​​constructed using the cluster centers by calculating the cosine similarity between the cluster centers. The construction formula is as follows:

[0045]

[0046] Among them ||·|| 2 This represents the sum of squares. In this method, all data values ​​in the semantic similarity matrix are between 0 and 1, and therefore can be used as weights to scale the source classification features of the target domain data.

[0047] The confidence score for each instance is calculated through the classification prediction. The confidence score should be proportional to the degree of certainty of the model's prediction result. For instances that are not easily confused, the classification prediction obtained after clustering should yield a result with high certainty, i.e., a high value in one dimension and low values ​​in other dimensions. For instances that are easily confused, the classification prediction should have many dimensions with similar values. Based on this characteristic, this invention provides a method for calculating the confidence score:

[0048]

[0049] Where x T For target domain data, f T (x T ) represents the target classification feature of the target domain data, σ(·) represents the softmax function, and H(t) = Σ-t i log(t i ), where C represents the number of categories. Figure 3 After training, instances from three categories were randomly selected from the target domain data to find the three with the highest and three with the lowest confidence scores. It can be seen that instances with high confidence scores possess significant features, clear compositions, and similar shapes, thus facilitating model distillation training. Instances with low confidence scores contain varying degrees of interfering information and unique structural features, which can interfere with the training results. Therefore, the confidence score, as a weight, can effectively filter the strength levels of all instances, significantly reducing the influence of easily confused instances on the final training results.

[0050] This invention uses the target classification feature vector extracted from the target domain data through the target model, inputs the target domain data into the feature extraction network of the target model, and then inputs the output of the feature extraction network into the classifier of the target model, finally outputting the target classification features of the target data.

[0051] The target classification features will be saved in the next round of training and participate in the training of the target model in the next round. For example... Figure 2 As shown, in the k-th round of training, the target classification features from the (k-1)-th round are used in the calculation of the loss.

[0052] Furthermore, the source feature vector is reconstructed using the semantic similarity matrix, and the reconstructed source classification feature vector is named the source classification signal. The reconstruction method involves retrieving the similarity between the category of each instance and other categories in the semantic similarity matrix, and then scaling each dimension of the instance's source classification feature based on this similarity. The scaled source classification feature is named the source classification signal.

[0053] The source classification signal is calculated as follows:

[0054]

[0055] in Represents the classification features of the target model; * indicates that the values ​​between matrices are multiplied together; represents matrix multiplication. The source classification prediction probability represents the target model. This represents the semantic similarity matrix.

[0056] Furthermore, semantic distillation loss and self-distillation loss are calculated based on the source classification signal, the target classification feature vector, and the target classification feature vector saved in the previous training round. The two losses are then scaled at the instance level using the confidence score. The parameters of the target model are adjusted and updated based on the scaled losses. Semantic distillation loss is calculated based on the source classification signal and the target classification feature vector. Semantic distillation loss is used to replace traditional classification loss for preliminary supervised training of the target model, endowing it with semantic information. Unlike traditional distillation loss, semantic distillation loss does not directly constrain the target classification features to be close to the source classification features, but rather constrains the target classification features to be close to the source classification signal obtained after semantically filtering the source classification features. Therefore, it avoids interference from incorrect predictions in the source classification features and improves the efficiency of knowledge distillation. The semantic distillation loss is calculated as follows:

[0057]

[0058] Where x T For target domain data, f T (x T ) represents the target classification feature of the target domain data, and σ(·) represents the softmax function. CE(·,·) represents the source classification signal of the target domain data, and CE(·,·) represents the cross-entropy loss.

[0059] The self-distillation loss is calculated based on the target classification feature vector saved in the previous training round and the target classification feature vector in this training round.

[0060] Furthermore, semantic distillation loss is calculated based on the source classification signal and the target classification feature vector. The semantic distillation loss is obtained by constraining the target classification feature vector to be close to the source classification signal, where the source classification signal does not participate in gradient updates. The self-distillation loss is obtained by constraining the target classification feature vector in this round to be similar to the target classification feature vector saved in the previous round. Self-distillation loss is not calculated for the first round of training. As training progresses, since the source model does not participate in gradient updates, its performance remains constant. In the initial stage, the source model's high performance can be aided by semantic distillation in training the target model. However, as the target model is continuously optimized, its performance may surpass that of the source model. In this case, self-distillation loss is needed to reduce the impact of insufficient source model performance on training. The formula for calculating self-distillation loss is as follows:

[0061]

[0062] Where x T For target domain data, Let σ(·) represent the target classification features of the target domain data in the k-th training round. Let CE(·,·) represent the target classification features of the target domain data in the (k-1)th round of training, and let CE(·,·) represent the cross-entropy loss.

[0063] The two losses are then scaled at the instance level using the confidence scores. The final loss is as follows:

[0064]

[0065] Where λ is a hyperparameter used to balance semantic distillation loss and self-distillation loss.

[0066] Furthermore, the domain adaptation model generates classification results based on the image of the classification result to be detected. The target model extracts target classification features from the image to be detected, and then clusters these features using a Gaussian mixture model to obtain classification prediction results. Experiments were conducted on three datasets widely used in the field of domain adaptation:

[0067] Office is a widely used benchmark for cross-domain object recognition, containing 31 categories across three different domains. Office-Home is also a popular benchmark with 4 domains and 65 classes, making it more challenging than Office. VISDA-C is a large-scale dataset consisting of 12 classes of instances. Its source domain includes 152,000 images generated from rendered 3D models, and the target domain contains 55,000 real object images from Microsoft COCO. The experimental results are presented in Tables 1-3, showing the test results of this invention on the Office dataset, the Office-Home dataset, and the VISDA-C dataset, respectively.

[0068] Table 1: Test results of the information selection-based domain adaptation model and method on the Office dataset.

[0069]

[0070] Table 2: Test results of the information selection-based domain adaptation model and method on the Office-Home dataset.

[0071]

[0072] Table 3: Test results of the information selection-based domain adaptation model and method on the VISDA-C dataset.

[0073]

[0074] It can be seen that the knowledge distillation-based method outperforms other methods, indicating that the concept of knowledge distillation is suitable for black-box UDA tasks. This invention achieves state-of-the-art performance on all three datasets, demonstrating the advancement of knowledge extraction under the information selection strategy of this invention. The performance improvement of this invention increases with the data size; on the VISDA-C dataset, this invention achieves an 8.4% improvement over the second-best DINE, because the instance-level information filtering of this invention effectively filters out cluttered data.

[0075] The above description is merely a specific embodiment of the present invention. Any feature disclosed in this specification may be replaced by other equivalent or similar features unless otherwise specified. All disclosed features, or steps in all methods or processes, may be combined in any way except for mutually exclusive features and / or steps.

Claims

1. A method for constructing a domain adaptation model based on information selection, characterized in that, This method is applied to domain adaptation under the condition that the source model parameters are not visible. The construction method includes: The target domain data is preprocessed, including image cropping and image flipping. A domain adaptation model based on information selection is constructed. This model consists of a source model and a target model. The source model is a supervised training model on source domain data. Its parameters and model structure are not visible. It can only complete input and output through an interface. The output is the source classification feature vector. The target model consists of a feature extraction network and a classifier. The final output is the target classification feature vector. The process involves extracting source classification feature vectors from the target domain data using the source model, obtaining cluster centers for each category and classification predictions for the target domain data by clustering the source classification feature vectors, constructing a semantic similarity matrix using the cluster centers, constructing confidence scores for each prediction using the classification predictions, and extracting target classification feature vectors from the target domain data using the target model. These target classification feature vectors will be saved until the end of the next training round. The process includes inputting preprocessed target domain data into the source model, the output of which is called the source classification feature vector of the target domain data; using a Gaussian Mixture Model (GMM) clustering algorithm to cluster the extracted source classification feature vectors to obtain the classification prediction for each instance and the cluster center for each category. A semantic similarity matrix is ​​constructed using the cluster centers by calculating the cosine similarity between the cluster centers. A confidence score for each instance is calculated using the classification prediction, and the confidence score should be proportional to the degree of certainty of the model's prediction result. The target model extracts the target classification feature vector of the target domain data, inputs the target domain data into the feature extraction network of the target model, and then inputs the output of the feature extraction network into the classifier of the target model, finally outputting the target classification feature vector of the target data. The target classification feature vector will be saved in the next round of training and participate in the training of the target model in the next round; The source classification feature vector is reconstructed using the semantic similarity matrix, and the reconstructed source classification feature vector is named the source classification signal. The semantic distillation loss is calculated based on the source classification signal and the target classification feature vector. The self-distillation loss is calculated based on the target classification feature vector saved in the previous training round and the target classification feature vector in this training round. The two losses are scaled at the instance level using the confidence score, and the parameters of the target model are adjusted and updated based on the scaled loss.

2. The method for constructing a domain adaptation model based on information selection according to claim 1, characterized in that, Image cropping and image flipping of the target domain data include: Flip all images horizontally or vertically with a probability of 0.5, and then crop the flipped images to a size of 224*224.

3. The method for constructing a domain adaptation model based on information selection according to claim 2, characterized in that, The feature extraction network of the target model is constructed as a residual neural network ResNet-50, with a randomly initialized 256-dimensional fully connected layer added at the end of the network, resulting in a final output vector dimension of 256 dimensions. The classifier of the target model is a randomly initialized fully connected layer that uses the output of the feature extraction network as input, resulting in a final output dimension of the target classification features equal to the number of categories.

4. The method for constructing a domain adaptation model based on information selection according to claim 3, characterized in that, The source classification feature vector is reconstructed using the semantic similarity matrix, and the reconstructed source classification feature vector is named the source classification signal. Specifically, this includes... The source classification feature vector is reconstructed using the semantic similarity matrix. The reconstruction method involves retrieving the similarity between the category of each instance and other categories in the semantic similarity matrix, and scaling each dimension of the source classification feature vector of the instance according to this similarity. The scaled source classification feature vector is named the source classification signal.

5. The method for constructing a domain adaptation model based on information selection according to claim 4, characterized in that, The semantic distillation loss is obtained by constraining the target classification feature vector to be close to the source classification signal, wherein the source classification signal does not participate in gradient update; The self-distillation loss is obtained by constraining the target classification feature vector of this round with the target classification feature vector saved in the previous round. If it is the first round of training, the self-distillation loss is not calculated.

6. The method for constructing a domain adaptation model based on information selection according to claim 5, characterized in that, The domain adaptation model generates classification prediction results based on the image of the classification result to be detected; The target model extracts the target classification feature vector of the image to be detected based on the image to be detected, and then clusters the target classification feature vector using a Gaussian mixture model to obtain the classification prediction result.

Images