A zero-shot classification method based on dynamic feature extraction and attribute correction

Abstract

The present invention discloses a zero-shot classification method based on dynamic feature extraction and attribute correction, comprising the following steps: obtaining visual samples and semantic features; constructing a zero-shot learning network based on dynamic feature extraction and attribute correction; transmitting the visual samples and semantic features to the zero-shot learning network based on dynamic feature extraction and attribute correction, obtaining and calculating a loss value based on the visual sample features and the corrected semantic features, backpropagating the loss value, and repeating this step until training is completed; verifying the trained zero-shot learning network based on dynamic feature extraction and attribute correction; if the accuracy rate is higher than a predetermined value, proceeding to the next step; otherwise, returning to the previous step; and classifying a data set using the trained zero-shot learning network based on dynamic feature extraction and attribute correction. The present invention adopts different feature extraction methods according to the different properties of different attributes, proposes the concept of attribute correction, and enhances the network's representational capabilities.

Description

Technical Field

[0001] The present invention relates to the field of zero-shot recognition, and specifically includes a zero-shot classification method based on dynamic feature extraction and attribute correction. Background Art

[0002] In traditional deep learning classification algorithm research, the samples in the training set include all label distribution information in the dataset. In this case, the model can learn all the knowledge about the sample distribution by learning from the training set, and the model's learning accuracy can be verified by testing the model's prediction accuracy on the test set. In this case, having the same labels in the training and test sets is crucial for validating the model's performance. However, in some specialized application scenarios, training samples for certain categories may be difficult to obtain or label. Without these labels, pre-trained models cannot predict for these categories. This practical challenge significantly limits the application of deep learning models. Therefore, to address the problem of predicting for novel categories, the zero-shot learning task has been proposed. This task attempts to enable a model to accurately identify samples from categories not seen in the training set, even if it already recognizes the categories already in the training set. This task, which allows a model to learn knowledge about unseen categories without ever having seen any samples, significantly expands the application of deep learning and holds great research value.

[0003] In order to study zero-shot learning, researchers have proposed and designed several datasets, each of which contains a large number of visual samples X. Assume that all categories of all visual samples are in It is the visible class that the model will see during training, the number is N s , the visual samples belonging to the visible class can be recorded as X s , Class is an unseen class used by the test set to detect the zero-sample learning performance of the model, and the number is N u , the visual samples belonging to the unseen class can be recorded as X u It is worth noting that there is no overlap between the visible and invisible classes and they cover all the classes in the dataset, i.e. In order to enable the model to learn without samples, researchers introduced the concept of semantic features in the dataset. Each class has a corresponding semantic feature. Assume that the semantic features of all classes in the dataset can be represented as A, where represents the semantic features of all visible classes, represents the semantic features of the unseen class, K represents the dimension of the semantic feature vector, where each dimension can be represented as a specific attribute, so each semantic feature can be represented by a combination of K attributes. When the zero-shot learning model is trained, the model can see the visual samples X of the seen class. s , and the semantics A of all classes including the unseen classes. Zero-shot learning aims to use the semantic feature A as a bridge, allowing the model to learn the relationship between the corresponding visual samples through the relationship between the semantics of the seen and unseen classes, thereby making accurate predictions on the visual samples of the unseen classes in the test set.

[0004] Currently, there are three mainstream technologies for zero-shot learning:

[0005] Existing technology 1: A learning algorithm based on cross-modal mapping, which maps visual samples originally distributed in the visual space and semantic features distributed in the semantic space into the same space, aligns the distribution of visual samples with the semantic features as the center point, and maps visual samples of unseen classes into this space for classification during the test phase.

[0006] The disadvantage of this technology is that the quality of feature extraction of visual samples cannot be guaranteed. This method uses the global features and semantic features of visual samples for alignment but ignores the extraction and understanding of local features of the samples. This will cause some redundant features of the visual samples to affect the training of the model and ultimately reduce the performance of the algorithm.

[0007] Existing technology 2: Generative methods, which directly address the core problem of zero-shot learning: the lack of samples of unseen classes. By generating a large number of samples of unseen classes based on semantics, the zero-shot learning task is ultimately transformed into a standard supervised learning task.

[0008] The main drawback of this technique, similar to existing techniques, is that it uses global features as the characteristic representation of visual samples for model training, while ignoring the importance of local features. To achieve high-quality generation of unseen samples, we often want the model to have good generation capabilities for specific semantically relevant attributes, while the irrelevant background is less important. However, this generation method based on global features often fails to take this into account, and therefore the generation quality cannot be guaranteed.

[0009] Existing technology three: A method based on the attention mechanism, which decomposes semantics into different attributes, extracts features from visual images based on attributes, and then aligns the extracted attribute features with the semantics as the feature expression of the image. Since semantics is composed of different attributes and attributes are universal between different classes, the features extracted based on the attributes can well predict visual samples of unseen classes.

[0010] While this technical approach considers the importance of local features for the first time, it still suffers from two major shortcomings. The first is a lack of targeted discussion of attribute categories. Semantic attributes can generally be divided into two types. The first is low-level texture-based attributes, which often describe the color or shape characteristics of specific parts of the visual sample's subject and are easily extracted by models. The other type of attributes are high-level, abstract attributes that require understanding of the relevant context, such as an animal's herbivorous nature, and cannot be captured through low-level textures. Existing solutions use a single method for extracting low-level texture attributes to extract features for all attributes, lacking consideration for high-level, abstract attributes. Another drawback is that existing technologies often target a fixed attribute for prediction. However, semantic features can vary depending on the angle and lighting of different visual samples. Therefore, using a fixed attribute value to describe all visual samples in a class ignores the characteristic variations of the attribute across different visual samples, ultimately resulting in poor feature extraction. Summary of the Invention

[0011] In response to the above-mentioned deficiencies in the prior art, the present invention provides a zero-sample classification method based on dynamic feature extraction and attribute correction, which solves the problem that the prior art lacks consideration of high-level abstract attributes and ignores the characteristic changes of attributes in different visual samples.

[0012] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: a zero-shot classification method based on dynamic feature extraction and attribute correction, comprising the following steps:

[0013] S1, obtain visual sample x and semantic feature α;

[0014] S2. Construct a zero-shot learning network based on dynamic feature extraction and attribute correction;

[0015] S3. Transmit the visual samples and semantic features to a zero-shot learning network based on dynamic feature extraction and attribute correction, obtain and calculate the loss function based on the visual sample features and the corrected semantic features, calculate the loss value based on the loss function, and return the gradient of the loss value; repeat this step until the training is completed;

[0016] S4, verifying the trained zero-shot learning network based on dynamic feature extraction and attribute correction; if the accuracy is higher than a predetermined value, proceeding to step S5; otherwise, proceeding to step S3;

[0017] S5. Use the trained zero-shot learning network based on dynamic feature extraction and attribute correction to classify the dataset.

[0018] Furthermore, the zero-shot learning network based on dynamic feature extraction and attribute correction includes a feature extraction backbone network, an attribute localization network, an attribute correction network, a scale control unit, and a loss value calculation module;

[0019] The first output end of the feature extraction backbone network is connected to the first input end of the attribute correction network; the second output end of the feature extraction backbone network is connected to the first input end of the attribute positioning network; the third output end of the feature extraction backbone network is connected to the input end of the scale control unit; the first output end of the scale control unit is connected to the second input end of the attribute correction network; the second output end of the scale control unit is connected to the second input end of the attribute positioning network; the output end of the attribute positioning network and the output end of the attribute correction network are connected to the loss value calculation module.

[0020] Furthermore, the specific implementation of step S3 is as follows:

[0021] S3-1. Use the attribute localization network to locate the feature attributes of the visual sample and extract local features and global features;

[0022] S3-2, extracting local features and global features required for attribute correction through the attribute correction network;

[0023] S3-3, fusing the local features and global features extracted by the attribute localization network and the attribute correction network through the scale control unit to obtain attribute correction values and visual sample features;

[0024] S3-4, correcting the semantic feature according to the attribute correction value to obtain a corrected semantic feature;

[0025] S3-5. Calculate the loss value based on the distance between the visual sample features and the corrected semantic features; return the loss value and update the parameters of the zero-shot learning network based on dynamic feature extraction and attribute correction.

[0026] Furthermore, the specific implementation of step S3-1 is as follows:

[0027] S3-1-1. Obtain the visual sample feature map of visual sample x through the feature extraction backbone network Among them, C represents the number of channels of the feature map, that is, the dimension of the feature of each pixel; H represents the height of the feature map; W represents the width of the feature map; Represents the shape of the data;

[0028] S3-1-2, according to the formula:

[0029]

[0030] Get the local feature u of the visual sample L ; Where i represents the height of the feature map; j represents the width of the feature map; is the attribute graph, which indicates the distribution of attributes on the feature graph, K indicates the number of attributes, w indicates the attention weight, and v indicates the specific distribution value of the attribute; the softmax function indicates that the pixel values of the feature graph on each channel are normalized to between 0 and 1; φ v and φ w Represents two convolutional layers with a convolution kernel size of 1×1;

[0031] S3-1-3, according to the formula:

[0032]

[0033] Get the global feature u of the visual sample G ;in; i' represents the height of the feature map; j' represents the width of the feature map.

[0034] Furthermore, the specific implementation of step S3-2 is as follows:

[0035] According to the formula:

[0036]

[0037] Get the local feature t of each attribute L and the global feature t for each attribute G ; Among them, φ r The convolution kernel size for calculating the attribute correction value is 1×1; max c’,d’ represents the global maximum pooling; c' represents the height of the feature map; d' represents the width of the feature map;

[0038] Furthermore, the specific implementation of step S3-3 is as follows:

[0039] S3-3-1. According to the formula:

[0040]

[0041] Get the probability g of whether the attribute is a local attribute or a global attribute; where φ s represents a convolution layer with a convolution kernel of 1×1; c represents the height of the feature map; d represents the width of the feature map;

[0042] S3-3-2, according to the formula:

[0043]

[0044] Get attribute modifier and visual sample features ψ(x).

[0045] Furthermore, the specific implementation of step S3-4 is as follows:

[0046] According to the formula:

[0047]

[0048] Get the corrected semantic feature π m (α); where Normalize means normalizing the vector length to 1; Represents the value corresponding to the nth dimension in the semantic feature of the mth category, n = 1, 2, ..., K; express The value corresponding to the nth dimension of .

[0049] Furthermore, the specific implementation of step S3-5 is as follows:

[0050] S3-5-1. According to the formula:

[0051]

[0052] Get the classification loss and distance loss Among them, N B represents the batch size of visual samples sampled in each round of learning; exp represents the natural exponential; cos represents the cosine similarity, τ represents the temperature coefficient, α y Representation and sample x p Semantic features corresponding to the same class; represents the square of L2 norm; Is a visible class; α q is the semantic feature of the qth class in the visible class;

[0053] S3-5-2, according to the formula:

[0054]

[0055] Get the gap between the predicted value and the true value of the zero-shot learning network based on dynamic feature extraction and attribute correction That is the final loss function.

[0056] S3-5-3. Calculate the loss value according to the loss function and perform gradient backpropagation to update the zero-shot learning network parameters based on dynamic feature extraction and attribute correction.

[0057] The beneficial effects of the present invention are as follows: the present invention classifies semantic attributes and designs a comprehensive attribute feature extraction method. For those attributes based on low-level textures, local feature extraction based on the attention mechanism is still retained. For those high-level abstract attributes based on content understanding, the global features of the visual sample are used as their feature expressions. The local features and global features are fused, and the attribute proportions of the two features are adjusted with the gating unit as the weight, so as to finally realize the extraction of visual sample features. The concept of attribute correction is proposed, and an attribute correction module is designed to modify the attribute value so that the attribute value is closer to the true expression of the visual sample. The features extracted from the visual sample are aligned with the corrected attributes to enhance the representation ability of the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 Flowchart of the present invention;

[0059] Figure 2 This is a network structure diagram of the present invention;

[0060] Figure 3 Visualization of the attention mechanism for the attribute localization module. DETAILED DESCRIPTION

[0061] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0062] like Figure 1 As shown, a zero-shot classification method based on dynamic feature extraction and attribute correction includes the following steps:

[0063] S1, obtain visual sample x and semantic feature α;

[0064] S2. Construct a zero-shot learning network based on dynamic feature extraction and attribute correction;

[0065] S3. Transmit the visual samples and semantic features to a zero-shot learning network based on dynamic feature extraction and attribute correction, obtain and calculate the loss function based on the visual sample features and the corrected semantic features, calculate the loss value based on the loss function, and return the gradient of the loss value; repeat this step until the training is completed;

[0066] S4, verifying the trained zero-shot learning network based on dynamic feature extraction and attribute correction; if the accuracy is higher than a predetermined value, proceeding to step S5; otherwise, proceeding to step S3;

[0067] S5. Use the trained zero-shot learning network based on dynamic feature extraction and attribute correction to classify the dataset.

[0068] The specific implementation of step S3 is as follows:

[0069] S3-1. Use the attribute localization network to locate the feature attributes of visual samples and extract local features and global features;

[0070] S3-2, extracting local features and global features required for attribute correction through the attribute correction network;

[0071] S3-3, fusing the local features and global features extracted by the attribute localization network and the attribute correction network through the scale control unit to obtain attribute correction values and visual sample features;

[0072] S3-4, correcting the semantic feature according to the attribute correction value to obtain a corrected semantic feature;

[0073] S3-5. Calculate the loss value based on the distance between the visual sample features and the corrected semantic features; return the loss value and update the parameters of the zero-shot learning network based on dynamic feature extraction and attribute correction.

[0074] The specific implementation of step S3-1 is as follows:

[0075] S3-1-1. Obtain the visual sample feature map of visual sample x through the feature extraction backbone network Among them, C represents the number of channels of the feature map, that is, the dimension of the feature of each pixel; H represents the height of the feature map; W represents the width of the feature map; Represents the shape of the data;

[0076] S3-1-2, according to the formula:

[0077]

[0078] Get the local feature u of the visual sample L ; Where i represents the height of the feature map; j represents the width of the feature map; is the attribute graph, which indicates the distribution of attributes on the feature graph, K indicates the number of attributes, w indicates the attention weight, and v indicates the specific distribution value of the attribute; the softmax function indicates that the pixel values of the feature graph on each channel are normalized to between 0 and 1; φ v and φ w Represents two convolutional layers with a convolution kernel size of 1×1;

[0079] S3-1-3, according to the formula:

[0080]

[0081] Get the global feature u of the visual sample G ;in; i' represents the height of the feature map; j' represents the width of the feature map.

[0082] The specific implementation of step S3-2 is as follows:

[0083] According to the formula:

[0084]

[0085] Get the local feature t of each attribute L and the global feature t of each attribute G ; Among them, φ r The convolution kernel size for calculating the attribute correction value is 1×1; max c’,d’ represents the global maximum pooling; c' represents the height of the feature map; d' represents the width of the feature map;

[0086] The specific implementation of step S3-3 is as follows:

[0087] S3-3-1. According to the formula:

[0088]

[0089] Get the probability g of whether the attribute is a local attribute or a global attribute; where φ s represents a convolution layer with a convolution kernel of 1×1; c represents the height of the feature map; d represents the width of the feature map;

[0090] S3-3-2, according to the formula:

[0091]

[0092] Get attribute modifier and visual sample features ψ(x).

[0093] The specific implementation of step S3-4 is as follows:

[0094] According to the formula:

[0095]

[0096] Get the corrected semantic feature π m (α); where Normalize means normalizing the vector length to 1; Represents the value corresponding to the nth dimension in the semantic feature of the mth category, n = 1, 2, ..., K; express The value corresponding to the nth dimension of .

[0097] The specific implementation of step S3-5 is as follows:

[0098] S3-5-1. According to the formula:

[0099]

[0100] Get the classification loss and distance loss Among them, N B represents the batch size of visual samples sampled in each round of learning; exp represents the natural exponential; cos represents the cosine similarity, τ represents the temperature coefficient, α y Representation and sample x p Semantic features corresponding to the same class; represents the square of L2 norm; Is a visible class; α q is the semantic feature of the qth class in the visible class;

[0101] S3-5-2, according to the formula:

[0102]

[0103] Get the gap between the predicted value and the true value of the zero-shot learning network based on dynamic feature extraction and attribute correction That is the final loss function.

[0104] S3-5-3. Calculate the loss value according to the loss function and perform gradient backpropagation to update the zero-shot learning network parameters based on dynamic feature extraction and attribute correction.

[0105] like Figure 2 As shown in the figure, the zero-shot learning network based on dynamic feature extraction and attribute correction includes a feature extraction backbone network, an attribute localization network, an attribute correction network, a scale control unit, and a loss value calculation module;

[0106] The first output end of the feature extraction backbone network is connected to the first input end of the attribute correction network; the second output end of the feature extraction backbone network is connected to the first input end of the attribute positioning network; the third output end of the feature extraction backbone network is connected to the input end of the scale control unit; the first output end of the scale control unit is connected to the second input end of the attribute correction network; the second output end of the scale control unit is connected to the second input end of the attribute positioning network; the output end of the attribute positioning network and the output end of the attribute correction network are connected to the loss value calculation module.

[0107] like Figure 3As shown in the figure, SUN represents the scene understanding dataset, and CUB represents the fine-grained bird classification dataset. The model accurately locates local features, such as different parts of a bird's body and still water and fences in complex scenes. Furthermore, for complex attributes requiring contextual understanding, such as the open fields in the SUN dataset, the model assigns a high attention weight to the entire image, which aligns with the definition of an open field.

[0108] In one embodiment of the present invention, the softmax function normalizes the pixel values of the feature map on each channel to between 0 and 1 to represent the attention weight, where pixels with higher values represent higher importance. Global maximum pooling can be considered as a special form of attention mechanism, that is, only one pixel has a weight of 1, and the weights of other pixels are 0. The zero-shot learning network based on dynamic feature extraction and attribute correction uses global average pooling to combine the features of each attribute on the graph to obtain a judgment score for whether the attribute belongs to a global attribute or a local attribute, and finally uses the sigmoid function to normalize the score to between 0 and 1. Classification loss The cosine similarity between the visual sample features extracted by the attribute localization module and the semantic features corrected by the attribute correction module is brought closer, which is the alignment of the overall semantic level; distance loss It directly requires that the sample features and the corrected semantics are the same in each dimension. This requires that the extracted sample features are the same as the corrected semantics in each dimension. This is attribute-level alignment.

[0109] In terms of quantitative analysis, the present invention achieves higher test set prediction accuracy compared with the prior art, as shown in Table 1.

[0110] Table 1

[0111]

[0112]

[0113] Zero-shot learning tasks are quantified using three metrics: the prediction accuracy S of the zero-shot learning network based on dynamic feature extraction and attribute correction on the seen class, the prediction accuracy U of the unseen class, and the harmonic mean H of these two accuracy rates. Generally, a higher harmonic mean indicates better overall algorithm performance. As shown in Table 1, the present invention achieves a significant improvement in the harmonic mean accuracy compared to existing technologies, demonstrating its superiority.

[0114] In qualitative analysis, through the visualization analysis of the attention mechanism, it can be shown that the present invention has achieved good results in the key task of extracting attribute features.

[0115] The present invention classifies semantic attributes and designs a comprehensive attribute feature extraction method. For those attributes based on low-level textures, local feature extraction based on the attention mechanism is still retained. For those high-level abstract attributes based on content understanding, the global features of the visual sample are used as their feature expressions. The local features and global features are fused, and the attribute proportions of the two features are adjusted with the gating unit as the weight, so as to finally realize the extraction of visual sample features. The concept of attribute correction is proposed, and an attribute correction module is designed to modify the attribute value so that the attribute value is closer to the true expression of the visual sample. The features extracted from the visual sample are aligned with the corrected attributes to enhance the representation ability of the network.

Claims

1. A zero-shot classification method based on dynamic feature extraction and attribute correction, characterized in that: The following steps are involved: S1. Obtain visual samples x and semantic feature α; S2. Construct a zero-shot learning network based on dynamic feature extraction and attribute correction; S3, transmitting the visual sample and semantic features to a zero-shot learning network based on dynamic feature extraction and attribute correction, obtaining and calculating a loss function based on the visual sample features and the corrected semantic features, calculating a loss value based on the loss function, and performing gradient backpropagation on the loss value; Repeat this step until the training is completed; The zero-shot learning network based on dynamic feature extraction and attribute correction includes a feature extraction backbone network, an attribute localization network, an attribute correction network, a scale control unit, and a loss value calculation module; S3-1. Use the attribute localization network to locate the feature attributes of visual samples and extract local features and global features; S3-2, extracting local features and global features required for attribute correction through the attribute correction network; S3-3, fusing the local features and global features extracted by the attribute localization network and the attribute correction network through the scale control unit to obtain attribute correction values and visual sample features; S3-4, correcting the semantic feature according to the attribute correction value to obtain a corrected semantic feature; S4, verifying the trained zero-shot learning network based on dynamic feature extraction and attribute correction; if the accuracy is higher than a predetermined value, proceeding to step S5; Otherwise, proceed to step S3; S5. Use the trained zero-shot learning network based on dynamic feature extraction and attribute correction to classify the dataset.

2. A zero-shot classification method based on dynamic feature extraction and attribute correction according to claim 1, characterized in that: The first output end of the feature extraction backbone network is connected to the first input end of the attribute correction network; the second output end of the feature extraction backbone network is connected to the first input end of the attribute positioning network; the third output end of the feature extraction backbone network is connected to the input end of the scale control unit; the first output end of the scale control unit is connected to the second input end of the attribute correction network; the second output end of the scale control unit is connected to the second input end of the attribute positioning network; the output end of the attribute positioning network and the output end of the attribute correction network are connected to the loss value calculation module.

3. The zero-shot classification method based on dynamic feature extraction and attribute correction according to claim 1, characterized in that: The specific implementation of step S3-1 is as follows: S3-1-1. Obtain visual samples through feature extraction backbone network x Visual sample feature map ;in, C Indicates the number of channels of the feature map, that is, the dimension of the features of each pixel; H Indicates the height of the feature map; W Indicates the width of the feature map; Represents the shape of the data; S3-1-2, according to the formula: Get local features of visual samples u L ;in, i Indicates the height of the feature map; j Indicates the width of the feature map; is the attribute graph, which represents the distribution of attributes on the feature graph. K Indicates the number of attributes, w represents the attention weight, v Indicates the specific distribution value of the attribute; The function represents normalizing the pixel values of the feature map on each channel to between 0 and 1; and Represents two convolutional layers with a convolution kernel size of 1×1; ; S3-1-3, according to the formula: Get the global features of visual samples u G ;in; ; Indicates the height of the feature map; Indicates the width of the feature map.

4. The zero-shot classification method based on dynamic feature extraction and attribute correction according to claim 3, characterized in that: The specific implementation of step S3-2 is as follows: According to the formula: Get the local features of each attribute t L and the global characteristics of each attribute t G ;in, The convolution kernel size for calculating the attribute correction value is 1×1; max c’,d’ represents global maximum pooling; Indicates the height of the feature map; Indicates the width of the feature map; ; .

5. The zero-shot classification method based on dynamic feature extraction and attribute correction according to claim 4, characterized in that: The specific implementation of step S3-3 is as follows: S3-3-1. According to the formula: Get the probability of whether the attribute is a local attribute or a global attribute g ;in, Indicates a convolution layer with a convolution kernel of 1×1; c Indicates the height of the feature map; d Indicates the width of the feature map; S3-3-2, according to the formula: Get attribute modifier and visual sample features .

6. The zero-shot classification method based on dynamic feature extraction and attribute correction according to claim 5, characterized in that: The specific implementation of step S3-4 is as follows: According to the formula: Get the corrected semantic features π m (α) ;in, Normalize Indicates normalizing the vector length to 1; Indicates the m The semantic features of the class n The value corresponding to the dimension, ; express No. n The value corresponding to the dimension.

7. The zero-shot classification method based on dynamic feature extraction and attribute correction according to claim 6, characterized in that: The specific implementation of step S3-5 is as follows: S3-5-1. According to the formula: Get the classification loss and distance loss ;in, N B represents the batch size of visual samples sampled in each round of learning; exp represents the natural exponential; cos represents the cosine similarity, τ represents the temperature coefficient, α y Representation and samples x p Semantic features corresponding to the same class; represents the square of L2 norm; Is a visible class; α q Is the first in the visible class q Semantic features of each class; S3-5-2, according to the formula: Get the gap between the predicted value and the true value of the zero-shot learning network based on dynamic feature extraction and attribute correction , which is the final loss function; S3-5-3. Calculate the loss value according to the loss function and perform gradient backpropagation to update the zero-shot learning network parameters based on dynamic feature extraction and attribute correction.

Images