Sample label smoothing method, device and equipment for multi-level classification neural network

By introducing Gaussian distribution probability and position weight into multi-level classification neural networks, the problem of classification results tending to be weak and one-size-fits-all probability distribution in traditional multi-label classification methods is solved, and the classification accuracy is improved.

CN114758188BActive Publication Date: 2024-11-22PING AN TECH (SHENZHEN) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210156418.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-02-21
Publication Date
2024-11-22
Estimated Expiration
2042-02-21

AI Technical Summary

Technical Problem

The traditional multi-label classification method leads to the weak final classification results, and there is a one-size-fits-all probability distribution problem for other labels, resulting in low accuracy of classification results.

Method used

The sample label smoothing method of multi-level classification neural network is adopted. By obtaining the sample initial label, preset initial label smoothing is performed, Gaussian distribution probability is introduced, position weight is configured according to the position distance, and the normalized probability value is smoothed with the sample. The values ​​are normalized and the non-sample smoothing of this class label value is replaced to obtain the target smoothing label.

Benefits of technology

The problem of equal probability trust in progressive gradual relationship classification categories in traditional technology is avoided, and different trust degrees are assigned to other classification categories are achieved, so that the identification probability distribution is more in line with the actual situation, and the accuracy of classification results is improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114758188B_ABST
    Figure CN114758188B_ABST
Patent Text Reader

Abstract

The present invention discloses a sample label smoothing method, device and equipment for a multi-level classification neural network. The method includes: obtaining the sample initial label corresponding to the training sample, and using a preset initial label smoothing method to obtain the sample initial smoothed label, taking the position of the sample class label value in the sample initial label as the symmetry center, obtaining a position weight set according to the distance between the position of each value and the symmetry center, obtaining the sample Gaussian distribution probability based on Gaussian distribution, normalizing the probability values ​​of other positions in the distribution except the origin position to obtain the corresponding normalized probability value, and normalizing with the sample smoothed class label value to obtain the corresponding target probability value, and replacing the non-sample smoothed class label value with the target probability value to obtain the target smoothed label. The problem of one-size-fits-all probability distribution for other labels is solved, and different trust levels are assigned to other classification categories, so that the recognition probability distribution of classification categories is more in line with the actual situation and reasonable.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of intelligent decision-making technology of artificial intelligence, and in particular to a sample label smoothing method, device, equipment and medium of a multi-level classification neural network. Background Art

[0002] At present, the multi-label classification method (average distance weight and value ML-kNN, DVML-kNN) is generally used for traditional technical label classification. The information of neighbors with labels and neighbors without labels is added to the classification, and the influence of neighbor samples on the samples to be tested is comprehensively considered. It calculates the distance of neighbor samples, quantifies the distance and selects appropriate weights to obtain a new classification function, and adds the concept of value when calculating the posterior probability, so that the final classification result is more biased towards the disadvantaged class. However, even if the final classification result is more biased towards the disadvantaged class, there is still a problem of one-size-fits-all probability distribution for other labels, resulting in low accuracy of the final classification result. Summary of the invention

[0003] The embodiments of the present invention provide a sample label smoothing method, device and equipment for a multi-layer classification neural network, aiming to solve the problem that the traditional multi-label classification method makes the final classification result more biased towards the disadvantaged class, and there is also a one-size-fits-all probability distribution problem for other labels, resulting in the problem that the final classification result is not accurate.

[0004] In a first aspect, an embodiment of the present invention provides a sample label smoothing method for a multi-level classification neural network, comprising:

[0005] Obtaining a sample initial label corresponding to the training sample, wherein the sample initial label is a one-dimensional matrix, wherein the one-dimensional matrix includes multiple values, each of which is used to describe the probability that the training sample is identified as a classification category corresponding to the value, and the multiple values ​​include a sample class label value and a non-sample class label value, wherein the sample class label value is used to describe the probability that the training sample is identified as a classification category with the highest probability, and the non-sample class label value is used to describe the probability that the training sample is identified as a classification category other than the classification category with the highest probability;

[0006] The sample initial label is smoothed by using a preset initial label smoothing method to obtain the sample initial smoothed label, wherein the sample class label value is correspondingly smoothed to the sample class smoothed label value;

[0007] Taking the position of the sample class label value in the sample initial label as the symmetry center, assigning position weights from small to large to the positions of the values ​​according to the order from near to far of the position of the values, and obtaining a position weight set including all the position weights;

[0008] Based on the Gaussian distribution, taking the symmetry center as the center, obtaining the distribution weight corresponding to each position weight in the position weight set, and obtaining the sample Gaussian distribution probability;

[0009] Normalizing the probability values ​​of other positions in the sample Gaussian probability distribution except the origin position to obtain corresponding normalized probability values;

[0010] All the normalized probability values ​​are taken as a whole and normalized with the sample smoothed class label value to obtain the weight of the whole. Based on the weight, the target probability value corresponding to each normalized probability value is obtained, and the target probability value replaces the non-sample smoothed class label value at the corresponding position in the sample initial smooth label to obtain the target smooth label of the training sample.

[0011] In a second aspect, an embodiment of the present invention provides a sample label smoothing device for a multi-level classification neural network, comprising:

[0012] A sample initial label acquisition unit is used to acquire the sample initial label corresponding to the training sample, wherein the sample initial label is a one-dimensional matrix, and the one-dimensional matrix includes multiple values, each of which is used to describe the probability that the training sample is identified as the classification category corresponding to the value, and the multiple values ​​include sample class label values ​​and non-sample class label values, wherein the sample class label values ​​are used to describe the probability that the training sample is identified as the classification category with the highest probability, and the non-sample class label values ​​are used to describe the probability that the training sample is identified as other classification categories except the classification category with the highest probability;

[0013] A label smoothing unit is used to smooth the sample initial label by using a preset initial label smoothing method to obtain the sample initial smoothed label, wherein the sample class label value is correspondingly smoothed to the sample smoothed class label value;

[0014] A position weight set acquisition unit is used to configure position weights from small to large for the positions of the values ​​according to the order from near to far of the positions of the values ​​in the symmetry center, taking the position of the sample class label value in the sample initial label as the symmetry center, and obtaining a position weight set including all the position weights;

[0015] A Gaussian distribution probability acquisition unit, configured to acquire, based on Gaussian distribution and with the symmetry center as the center, the distribution weight corresponding to each position weight in the position weight set, and obtain the sample Gaussian distribution probability;

[0016] A probability value normalization unit, used to normalize the probability values ​​of other positions in the sample Gaussian probability distribution except the origin position to obtain corresponding normalized probability values;

[0017] The target smooth label acquisition unit is used to normalize all the normalized probability values ​​as a whole with the sample smoothed class label value to obtain the weight of the whole, and based on the weight, obtain the target probability value corresponding to each normalized probability value, and replace the non-sample smoothed class label value at the corresponding position in the sample initial smooth label with the target probability value to obtain the target smooth label of the training sample.

[0018] In a third aspect, an embodiment of the present invention further provides a computer device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the sample label smoothing method of a multi-layer classification neural network described in the first aspect is implemented.

[0019] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor executes the sample label smoothing method of a multi-layer classification neural network described in the first aspect above.

[0020] The embodiment of the present invention provides a sample label smoothing method, device and equipment for a multi-level classification neural network. For classification problems with levels, especially for multi-level classification problems with progressive relationships, by introducing Gaussian probability distribution in the initial smoothed labels of the samples, the label weights assigned to non-sample class labels are inversely proportional to the distance from the sample class labels, so that the non-sample class labels are assigned different weights according to the distance from the sample class labels, and the non-sample class labels that are closer to the sample class labels are assigned larger weights, and the non-sample class labels that are farther from the sample class labels are assigned smaller weights. This can avoid the problem that in the traditional technology label smoothing process, for classification categories with progressive relationships, the probability of setting other positions to alpha / (K-1) is equivalent to trusting several other classification categories with equal probability, solves the problem of one-size-fits-all probability distribution for other labels, and achieves different trusts for other classification categories, so that the recognition probability distribution of classification categories is more in line with the actual situation and more reasonable. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying any creative work.

[0022] Figure 1 A schematic flow chart of a sample label smoothing method for a multi-level classification neural network provided by an embodiment of the present invention;

[0023] Figure 2 A schematic diagram of a sub-process of a sample label smoothing method for a multi-level classification neural network provided by an embodiment of the present invention;

[0024] Figure 3 A schematic block diagram of a sample label smoothing device for a multi-level classification neural network provided by an embodiment of the present invention;

[0025] Figure 4 A schematic block diagram of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0026] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] It should be understood that when used in this specification and the appended claims, the terms "include" and "comprises" indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0028] It should also be understood that the terms used in this specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0029] It should be further understood that the term "and / or" used in the present description and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0030] See also Figure 1 , Figure 1A schematic flow chart of a sample label smoothing method for a multi-level classification neural network provided by an embodiment of the present invention; Figure 1 As shown, the method includes steps S11 to S16.

[0031] S11. Obtain the sample initial label corresponding to the training sample, wherein the sample initial label is a one-dimensional matrix, and the one-dimensional matrix includes multiple values, each of which is used to describe the probability of the training sample being identified as the classification category corresponding to the value, and the multiple values ​​include sample class label values ​​and non-sample class label values. The sample class label values ​​are used to describe the probability of the training sample being identified as the classification category with the highest probability, and the non-sample class label values ​​are used to describe the probability of the training sample being identified as other classification categories except the classification category with the highest probability.

[0032] Specifically, classification is to identify the target as a corresponding category, and the classification neural network is a neural network based on deep learning to classify the target so as to identify the target as the corresponding category as accurately as possible. Classification can be divided into binary classification and multi-classification. Binary classification means that the target is either of this category or another category. Multi-classification means that the target belongs to one of three or more categories. Multi-classification includes multi-level classification, which can also be called hierarchical multi-classification. The hierarchical multi-classification includes progressive multi-level classification, which can also be called multi-level classification, which is a category classification in which there is a progressive relationship between multiple categories of hierarchical classification. For example, taking the six categories of image blur (that is, the six categories of image clarity) as an example, the image can be divided into 6 categories according to the degree of blur (that is, the image clarity can be divided into 6 levels), which can be divided into: high-definition, clear, relatively clear, relatively blurred, blurred and very blurred. The six categories of image blur are progressive multi-level classification.

[0033] When training a multi-level classification neural network model, training samples are usually used to train the multi-level classification neural network model. Generally, training samples are prepared in advance and labeled to obtain the labeled categories to which the training samples are labeled. Then, the training samples are identified using the multi-level classification neural network model to identify the training samples as predicted categories. Then, the multi-level classification neural network model performs deep learning based on the training samples, labeled categories, and predicted categories to achieve training of the multi-level classification neural network model. The goal of training the multi-level classification neural network model is to make the predicted categories and the labeled categories consistent, that is, the multi-level classification neural network model is considered to have correctly identified the training samples. Otherwise, the multi-level classification neural network model is considered to have incorrectly identified the training samples.

[0034] When training a multi-level classification neural network model, the multi-level classification neural network model first classifies the training samples to obtain the probability that the multi-level classification neural network model identifies the training samples as each classification category in the multi-level classification. The category corresponding to the maximum probability is the category to which the training sample belongs by default. The classification category to which the training sample belongs can be called the class category of the training sample, and other classification categories can be called the non-class category of the training sample, thereby obtaining the sample initial label corresponding to the training sample. Since this is a multi-level classification problem with multiple classification categories, the sample initial label is a one-dimensional matrix, which contains multiple values, each of which is used to describe the probability that the training sample is identified as the classification category corresponding to the value (that is, the predicted probability). Among all the predicted probabilities, the maximum value is used to describe the probability corresponding to the classification category to which the training sample should most belong, that is, the probability corresponding to the classification category to which the training sample is most likely to be identified. The classification category to which the training sample is most likely to be identified can be described by the sample class category, and the maximum value in the predicted probability can be described by the sample class label value, and the sample class category is the classification category to which the training sample is predicted to be most likely to belong. Since the classification problem is a probabilistic prediction problem, it can be based on a multi-level classification neural network model by labeling the training samples, especially numerically labeling the training samples, such as labeling the training samples by 10, to describe the probability of the training samples being identified as each classification category in the multi-level classification. For example, one-hot label is used as the label, and one-hot label is the label obtained by One-hot encoding the sample. For example, for the above-mentioned six-category image blurriness classification, assuming that the blurriness of a training sample image is relatively blurred, the one-hot label 01 label corresponding to the training sample image is [0, 0, 0, 1, 0, 0], that is, the sample initial label of the training sample image is [0, 0, 0, 1, 0, 0], which is used to describe that the probability of the training sample image being identified as the sample class category is 1, and the classification category of the sample class label value with a probability of 1 corresponds to the classification category of "relative blur", and the non-sample class label values ​​of other non-sample class categories in the multi-level classification are [0, 0, 0, 0, 0] respectively, which is used to describe that the probability of the training sample image being identified as other non-class categories is 0.Therefore, training samples, the sample class categories annotated with the training samples, and the multi-level classifications to which the sample class categories belong are obtained, and the training samples are labeled according to the sample class categories and the multiple non-sample class categories contained in the multi-level classifications, so as to obtain an initial label matrix in the form of a one-dimensional matrix corresponding to the training samples. The initial label matrix includes sample class label values ​​describing the sample class categories and non-sample class label values ​​describing the non-sample class categories. Each value in the initial label matrix is ​​used to describe the probability that the training sample is identified as the classification category corresponding to the value. Since the probability of a training sample being identified as a classification category is the highest, under normal circumstances, the sample class label value is generally the maximum value corresponding to the maximum probability described in the initial label matrix.

[0035] S12. Smoothing the sample initial labels by using a preset initial label smoothing method to obtain sample initial smoothed labels, wherein the sample class label value is correspondingly smoothed to the sample class smoothed label value.

[0036] Among them, label smoothing is a modification of the loss function, which adjusts the training target of the neural network from "1" to "1-label smoothing adjustment".

[0037] Specifically, the classification problem is essentially a probability prediction problem. The multi-level classification neural network model is trained with the goal of minimizing the gap between the predicted probability and the true probability of the training samples when the multi-level classification neural network model classifies the training samples, thereby obtaining the most optimized probability distribution, so that the multi-level classification neural network model learns in the direction of maximizing the difference between the correct label and the wrong label. At the same time, in order to prevent the overfitting of the multi-level classification labels, label smoothing is usually adopted, such as adding noise through the soft label, thereby reducing the weight of the category of the real sample label when calculating the loss function, to solve the problem of overfitting. For example, for the sample initial label [0, 0, 0, 1, 0, 0] of the above-mentioned six-category image blur, when the soft label is used to achieve label smoothing, the original 1 position becomes 1-alpha, and the other positions become alpha / (K-1), where K is the number of categories of the six-category image blur, that is, K is 6. If alpha is 0.1, the sample initial label is converted from [0, 0, 0, 1, 0, 0] to the sample initial smoothed label [0.02, 0.02, 0.02, 0.9, 0.02, 0.02, where the sample smoothed class label value is 0.9, and the smoothed label values ​​of other classification categories in the preset multi-level classification are [0.02, 0.02, 0.02, 0.02, 0.02], which are non-sample smoothed class label values. Therefore, after obtaining the sample initial label corresponding to the training sample, the sample initial label is smoothed by the preset initial label smoothing method to obtain the sample initial smoothed label, where the sample class label value is correspondingly smoothed to the sample smoothed class label value, and the smoothed label values ​​other than the sample smoothed class label value are non-sample smoothed class label values.

[0038] S13. Taking the position of the sample class label value in the sample initial label as the symmetry center, assign position weights from small to large to the positions of the values ​​according to the order from near to far of the position of each value from the symmetry center, and obtain a position weight set containing all the position weights.

[0039] Specifically, taking the position of the sample class label value in the sample initial label as the symmetry center, according to the order of the positions of the non-sample class label values ​​and the sample class label values ​​from the symmetry center from near to far, the positions of the values ​​are configured with position weights from small to large. The position weight is also the distance weight. The closer the distance, the smaller the position weight, and the farther the distance, the larger the position weight. A position weight set containing the position weights of each position is obtained, so that the classification category closer to the sample class label value has a smaller position weight. For example, for the above-mentioned six categories of image blur, since the blur of the training sample image is relatively blurry, the position weight corresponding to the relative blur can be 0, and the classification categories "relatively clear" and "blurred" based on the distance "relative blur" should have smaller position weights than "clear" and "very blurry", so that the probability distribution of subsequent image blur recognition is more in line with the actual situation and more reasonable.

[0040] Furthermore, if Figure 2 As shown, step S13 takes the position of the sample class label value in the sample initial label as the symmetry center, and configures the position of each value from small to large in the order from the position of each value to the symmetry center, and obtains a position weight set containing all the position weights, including:

[0041] S130, converting the initial labels of the samples into a one-dimensional coordinate system to obtain a one-dimensional coordinate system;

[0042] S131, taking the symmetry center as the origin of the one-dimensional coordinate system, taking the origin coordinates of the origin as the preset reference value, and increasing the absolute values ​​of the coordinates corresponding to the positions of the non-sample class label values ​​in sequence from near to far from the origin coordinates, to obtain the coordinates corresponding to the positions of the non-sample class label values;

[0043] S132: Use the coordinates as position weights, and obtain sample label coordinates corresponding to the sample initial labels.

[0044] Specifically, the sample initial label is converted into a one-dimensional coordinate system, and the position of the sample class label value in the sample initial label is used as the origin of the one-dimensional coordinate system. The origin coordinates of the origin are used as preset reference values. According to the order from near to far of the position of each non-sample class label value from the origin coordinates, the absolute values ​​of the coordinates corresponding to the position of each non-sample class label value are increased in sequence to obtain the coordinates corresponding to the position of the non-sample class label value, and the sample label coordinates corresponding to the sample initial label are obtained.

[0045] Furthermore, in step S131, according to the order from near to far of the position of each non-sample class label value to the origin coordinate, the absolute value of the position corresponding to the coordinate of each non-sample class label value is increased in sequence, including:

[0046] Taking the preset step length as the coordinate unit, according to the order from near to far of the position of each non-sample class label value from the origin coordinate, the absolute value of the coordinate corresponding to the position of each non-sample class label value is increased in sequence.

[0047] Specifically, since the sample initial label is a one-dimensional matrix, the one-dimensional matrix is ​​converted into a one-dimensional coordinate system, that is, the sample initial label is represented by a one-dimensional coordinate system, and the position of the sample class label value in the sample initial label is used as the origin of the one-dimensional coordinate system, that is, the position of the sample class label value in the sample initial label is used as the symmetry center, and the origin coordinates of the origin are used as the preset reference value, and then the preset step size is used as the coordinate unit, according to the order from near to far of the position of each non-sample class label value in the sample initial label from the origin coordinates, the absolute values ​​of the corresponding coordinates of the positions of each non-sample class label value in the sample initial label are incremented in sequence to obtain the coordinates corresponding to the positions of the non-sample class label values, and the sample label coordinates corresponding to the sample initial label are obtained. For example, for the sample initial labels [0, 0, 0, 1, 0, 0] of the six categories of image blurriness mentioned above, the position of the sample class label value in the sample initial label is used as the origin of the one-dimensional coordinate system, and the origin coordinates of the origin are used as the reference value. If the origin coordinates are set to 0, 1 is used as the preset step size, and if one category is moved to the right, the corresponding coordinates increase by 1, and if one category is moved to the left, the corresponding coordinates decrease by 1, the coordinates corresponding to the position of the non-sample class label value can be obtained, and the sample label coordinates corresponding to the sample initial label [0, 0, 0, 1, 0, 0] are [-3, -2, -1, 0, 1, 2].

[0048] S14. Based on the Gaussian distribution, taking the symmetry center as the center, obtain the distribution weight corresponding to each position weight in the position weight set to obtain the sample Gaussian distribution probability.

[0049] Specifically, Gaussian distribution is a weighted distribution mode. The closer to the center position, the larger the value, and the farther away from the center position, the smaller the value. Therefore, if described in a one-dimensional coordinate system, based on Gaussian distribution, with the origin coordinate as the center, obtain the distribution weights corresponding to each coordinate in the sample label coordinate, and obtain the sample Gaussian distribution probability. Then, based on the position relationship of each non-sample class label value from the origin coordinate from near to far, introduce Gaussian probability distribution in the sample initial label so that non-sample class labels have different weights. The closer the non-sample class label is to the sample class label, the larger the weight is. The following Gaussian distribution probability density function can be adopted:

[0050] (1)

[0051] Wherein, in formula (1), x is each coordinate in the sample label coordinates, f(x) is used to describe the probability value of identifying the sample as the category corresponding to the x coordinate, c is the coordinate of the sample's own label (that is, the origin coordinate, which is also the preset reference value) or the mean of the absolute values ​​of each coordinate in the sample label coordinates, e is a constant parameter, and the variance σ is a constant, which can be taken as an empirical value of 1.5, and the probability value of each coordinate in the sample label coordinates can be obtained. For example, for the above image blur six-classification sample initial label [0, 0, 0, 1, 0, 0], the corresponding sample label coordinates are [-3, -2, -1, 0, 1, 2], the variance is 1.5, c is the coordinate of the sample class label and is set to 0. Substituting the coordinates of each position of the sample label coordinate [-3, -2, -1, 0, 1, 2] into formula (1), the probability values ​​of each position category are: 0.0360, 0.1093, 0.2130, 0.2660, 0.2130, 0.1093, and the sample Gaussian distribution probability is [0.0360, 0.1093, 0.2130, 0.2660, 0.2130, 0.1093].

[0052] S15. Normalize the probability values ​​of other positions in the sample Gaussian probability distribution except the origin position to obtain corresponding normalized probability values.

[0053] Specifically, the probability values ​​of other positions except the origin position in the sample Gaussian probability distribution are normalized to obtain corresponding normalized probability values. For example, for the sample Gaussian distribution probability [0.0360, 0.1093, 0.2130, 0.2660, 0.2130, 0.1093], the five probability values ​​of other positions except 0.2660 are added and summed to obtain 0.6806, and the five probability values ​​of other positions are normalized to obtain normalized probability values ​​of 0.06, 0.16, 0.31, 0.31, and 0.16.

[0054] S16. Take all the normalized probability values ​​as a whole, normalize them with the sample smoothed class label value, and obtain the weight of the whole. Based on the weight, obtain the target probability value corresponding to each normalized probability value, and replace the non-sample smoothed class label value at the corresponding position in the sample initial smooth label with the target probability value to obtain the target smooth label of the training sample.

[0055] Specifically, all the normalized probability values ​​are taken as a whole and normalized with the sample smoothed class label value, that is, 1-the sample smoothed class label value is taken as the sum of all the normalized probability values, and the sum is evenly divided into each of the normalized probability values, and the calculated result is taken as the target probability value corresponding to each of the normalized probability values, thereby obtaining the target probability value corresponding to each of the normalized probability values, and replacing the non-sample smoothed class label value at the corresponding position in the sample initial smooth label with the target probability value to obtain the target smoothed label of the training sample, so that the target smoothed label has both the normalized target and the different weights of other classification categories according to the distance from the sample smoothed class label value, thereby avoiding the one-size-fits-all approach to other labels and adding more reasonable weights based on equal trust.

[0056] Furthermore, the step S16 of obtaining the target probability value corresponding to each normalized probability value based on the weight includes:

[0057] All the normalized probability values ​​divide the weight equally to obtain a sub-weight of each normalized probability value accounting for the weight, and the sub-weight is used as the target probability value corresponding to the normalized probability value.

[0058] Specifically, all the normalized probability values ​​are taken as a whole and normalized with the sample smoothed class label value to obtain the weights occupied by all the normalized probability values, and then the weights are evenly distributed to each of the normalized probability values ​​to obtain the sub-weight of each normalized probability value, and the sub-weight is used as the target probability value corresponding to the normalized probability value, and then the target probability value replaces the non-sample smoothed class label value at the corresponding position in the sample initial smooth label to obtain the target smooth label of the training sample. For example, for the above normalized probability values ​​of 0.06, 0.16, 0.31, 0.31, 0.16, it is necessary to consider normalizing them again with the sample smoothed class label value 0.9, that is, the above normalized probability values ​​0.06, 0.16, 0.31, 0.31, 0.16 need to be normalized to 0.1, that is, the probability values ​​of other positions except the origin position in the Gaussian distribution probability distribution of the sample are normalized twice, so as to obtain [0.006, 0.016, 0.031, 0.031, 0.016] corresponding to the above normalized probability values ​​0.06, 0.16, 0.31, 0.31, 0.16 normalized to 0.1, and then [0.006, 0.016, 0.031, 0.031, 0.016] replaces the corresponding non-sample smoothed class label values ​​at each position except 0.9 in the sample initial smoothed labels [0.02, 0.02, 0.02, 0.9, 0.02, 0.02] to obtain the target smoothed labels of the training samples [0.006, 0.016, 0.031, 0.9, 0.031, 0.016].

[0059] In an embodiment of the present application, the sample initial label corresponding to the training sample is obtained, and the sample initial label is smoothed by a preset initial label smoothing method to obtain the sample initial smoothed label, the position of the sample class label value in the sample initial label is taken as the symmetry center, and the position of each value is configured with a position weight from small to large according to the order from near to far of the position of each value to the symmetry center, to obtain a position weight set containing all the position weights, based on a Gaussian distribution, with the symmetry center as the center, to obtain the distribution weight corresponding to each position weight in the position weight set, to obtain the sample Gaussian distribution probability, to normalize the probability values ​​of other positions except the origin position in the sample Gaussian probability distribution, to obtain the corresponding normalized probability value, to normalize all the normalized probability values ​​as a whole with the sample smoothed class label value, to obtain each The target probability value corresponding to the normalized probability value is obtained, and the target probability value is used to replace the non-sample smoothed class label value at the corresponding position in the sample initial smoothed label to obtain the target smoothed label of the training sample, thereby optimizing the label of the training sample. For hierarchical classification problems, especially for multi-level classification problems with progressive relationships, Gaussian probability distribution is introduced into the sample initial smoothed label so that the label weight assigned to the non-sample class label is inversely proportional to the distance from the sample class label, so that the non-sample class label is assigned different weights according to the distance from the sample class label. The non-sample class label closer to the sample class label is assigned a larger weight, and the non-sample class label farther from the sample class label is assigned a smaller weight. This can avoid the problem that in the traditional technology label smoothing process, for the classification category with progressive relationships, the probability of setting other positions to alpha / (K-1) is equivalent to trusting other classification categories with equal probability, which solves the problem of one-size-fits-all probability distribution for other labels, realizes the assignment of different trusts to other classification categories, and makes the recognition probability distribution of classification categories more in line with the actual situation and reasonable.

[0060] The embodiment of the present invention also provides a sample label smoothing device for a multi-level classification neural network, which is used to execute any embodiment of the sample label smoothing method for a multi-level classification neural network. Figure 3 , Figure 3 A schematic block diagram of a sample label smoothing device for a multi-level classification neural network provided by an embodiment of the present invention.

[0061] like Figure 3As shown, the sample label smoothing device 100 of the multi-level classification neural network includes a sample initial label acquisition unit 110, a label smoothing unit 120, a position weight set acquisition unit 130, a Gaussian distribution probability acquisition unit 140, a probability value normalization unit 150 and a target smooth label acquisition unit 160.

[0062] The sample initial label acquisition unit 110 converts the sample initial label into a one-dimensional coordinate system to obtain a one-dimensional coordinate system;

[0063] The label smoothing unit 120 uses the symmetry center as the origin of the one-dimensional coordinate system, uses the origin coordinates of the origin as a preset reference value, and increases the absolute values ​​of the coordinates corresponding to the positions of the non-sample class label values ​​in sequence according to the order from near to far of the positions of the non-sample class label values ​​from the origin coordinates, to obtain the coordinates corresponding to the positions of the non-sample class label values;

[0064] The position weight set acquisition unit 130 is used to take the position of the sample class label value in the sample initial label as the symmetry center, and configure position weights from small to large for the positions of each value according to the order from near to far of the position of each value from the symmetry center, so as to obtain a position weight set containing all the position weights.

[0065] Specifically, taking the position of the sample class label value in the sample initial label as the symmetry center, according to the order of the positions of the non-sample class label values ​​and the sample class label values ​​from the symmetry center from near to far, the positions of the values ​​are configured with position weights from small to large. The position weight is also the distance weight. The closer the distance, the smaller the position weight, and the farther the distance, the larger the position weight. A position weight set containing the position weights of each position is obtained, so that the classification category closer to the sample class label value has a smaller position weight. For example, for the above-mentioned six categories of image blur, since the blur of the training sample image is relatively blurry, the position weight corresponding to the relative blur can be 0, and the classification categories "relatively clear" and "blurred" based on the distance "relative blur" should have smaller position weights than "clear" and "very blurry", so that the probability distribution of subsequent image blur recognition is more in line with the actual situation and more reasonable.

[0066] Furthermore, the position weight set acquisition unit 130 is specifically used for:

[0067] Converting the initial labels of the samples into a one-dimensional coordinate system to obtain a one-dimensional coordinate system;

[0068] Taking the symmetry center as the origin of the one-dimensional coordinate system, taking the origin coordinate of the origin as the preset reference value, and according to the order from near to far of the position of each non-sample class label value from the origin coordinate, increasing the absolute value of the coordinate corresponding to the position of each non-sample class label value in sequence, to obtain the coordinate corresponding to the position of the non-sample class label value;

[0069] The coordinates are used as position weights, and the sample label coordinates corresponding to the sample initial labels are obtained.

[0070] Specifically, the sample initial label is converted into a one-dimensional coordinate system, and the position of the sample class label value in the sample initial label is used as the origin of the one-dimensional coordinate system. The origin coordinates of the origin are used as preset reference values. According to the order from near to far of the position of each non-sample class label value from the origin coordinates, the absolute values ​​of the coordinates corresponding to the position of each non-sample class label value are increased in sequence to obtain the coordinates corresponding to the position of the non-sample class label value, and the sample label coordinates corresponding to the sample initial label are obtained.

[0071] Furthermore, the absolute values ​​of the coordinates corresponding to the positions of the non-sample class label values ​​are increased in sequence according to the order from near to far of the positions of the non-sample class label values ​​from the origin coordinates, including:

[0072] Taking the preset step length as the coordinate unit, according to the order from near to far of the position of each non-sample class label value from the origin coordinate, the absolute value of the coordinate corresponding to the position of each non-sample class label value is increased in sequence.

[0073] Specifically, since the sample initial label is a one-dimensional matrix, the one-dimensional matrix is ​​converted into a one-dimensional coordinate system, that is, the sample initial label is represented by a one-dimensional coordinate system, and the position of the sample class label value in the sample initial label is used as the origin of the one-dimensional coordinate system, that is, the position of the sample class label value in the sample initial label is used as the symmetry center, and the origin coordinates of the origin are used as the preset reference value, and then the preset step size is used as the coordinate unit, according to the order from near to far of the position of each non-sample class label value in the sample initial label from the origin coordinates, the absolute values ​​of the corresponding coordinates of the positions of each non-sample class label value in the sample initial label are incremented in sequence to obtain the coordinates corresponding to the positions of the non-sample class label values, and the sample label coordinates corresponding to the sample initial label are obtained. For example, for the sample initial labels [0, 0, 0, 1, 0, 0] of the six categories of image blurriness mentioned above, the position of the sample class label value in the sample initial label is used as the origin of the one-dimensional coordinate system, and the origin coordinates of the origin are used as the reference value. If the origin coordinates are set to 0, 1 is used as the preset step size, and if one category is moved to the right, the corresponding coordinates increase by 1, and if one category is moved to the left, the corresponding coordinates decrease by 1, the coordinates corresponding to the position of the non-sample class label value can be obtained, and the sample label coordinates corresponding to the sample initial label [0, 0, 0, 1, 0, 0] are [-3, -2, -1, 0, 1, 2].

[0074] The Gaussian distribution probability acquisition unit 140 is used to acquire the distribution weight corresponding to each position weight in the position weight set based on Gaussian distribution and taking the symmetry center as the center to obtain the sample Gaussian distribution probability.

[0075] Specifically, Gaussian distribution is a weighted distribution mode. The closer to the center position, the larger the value, and the farther away from the center position, the smaller the value. Therefore, if described in a one-dimensional coordinate system, based on Gaussian distribution, with the origin coordinate as the center, the distribution weights corresponding to each coordinate in the sample label coordinate are obtained to obtain the sample Gaussian distribution probability. Then, based on the position relationship of each non-sample class label value from the origin coordinate from near to far, Gaussian probability distribution is introduced into the sample initial label so that the non-sample class labels have different weights, and the non-sample class labels that are closer to the sample class labels have relatively larger weights. The distribution probability density function of the Gaussian distribution can be adopted as shown in the above formula (1), where x is the coordinates in the sample label coordinates, f(x) is used to describe the probability value of identifying the sample as the category corresponding to the x coordinate, c is the coordinate of the sample's own label (that is, the origin coordinate, which is also the preset reference value) or the mean of the absolute values ​​of the coordinates in the sample label coordinates, e is a constant parameter, and the variance σ is a constant, which can be taken as an empirical value of 1.5, and the probability value of each coordinate in the sample label coordinates can be obtained. For example, for the above image blur six-classification sample initial label [0, 0, 0, 1, 0, 0], the corresponding sample label coordinates are [-3, -2, -1, 0, 1, 2], the variance is 1.5, c is the coordinate of the sample class label and is set to 0. Substituting the coordinates of each position of the sample label coordinate [-3, -2, -1, 0, 1, 2] into formula (1), the probability values ​​of each position category are: 0.0360, 0.1093, 0.2130, 0.2660, 0.2130, 0.1093, and the sample Gaussian distribution probability is [0.0360, 0.1093, 0.2130, 0.2660, 0.2130, 0.1093].

[0076] The probability value normalization unit 150 is used to normalize the probability values ​​of other positions in the sample Gaussian probability distribution except the origin position to obtain corresponding normalized probability values.

[0077] Specifically, the probability values ​​of other positions except the origin position in the sample Gaussian probability distribution are normalized to obtain corresponding normalized probability values. For example, for the sample Gaussian distribution probability [0.0360, 0.1093, 0.2130, 0.2660, 0.2130, 0.1093], the five probability values ​​of other positions except 0.2660 are added and summed to obtain 0.6806, and the five probability values ​​of other positions are normalized to obtain normalized probability values ​​of 0.06, 0.16, 0.31, 0.31, and 0.16.

[0078] The target smooth label acquisition unit 160 is used to normalize all the normalized probability values ​​as a whole with the sample smoothed class label value to obtain the weight of the whole, and based on the weight, obtain the target probability value corresponding to each normalized probability value, and replace the non-sample smoothed class label value at the corresponding position in the sample initial smooth label with the target probability value to obtain the target smooth label of the training sample.

[0079] Specifically, all the normalized probability values ​​are taken as a whole and normalized with the sample smoothed class label value, that is, 1-the sample smoothed class label value is taken as the sum of all the normalized probability values, and the sum is evenly divided into each of the normalized probability values, and the calculated result is taken as the target probability value corresponding to each of the normalized probability values, thereby obtaining the target probability value corresponding to each of the normalized probability values, and replacing the non-sample smoothed class label value at the corresponding position in the sample initial smooth label with the target probability value to obtain the target smoothed label of the training sample, so that the target smoothed label has both the normalized target and the different weights of other classification categories according to the distance from the sample smoothed class label value, thereby avoiding the one-size-fits-all approach to other labels and adding more reasonable weights based on equal trust.

[0080] Furthermore, the target smoothing label acquisition unit 160 is specifically used for:

[0081] The weight is equally divided by all the normalized probability values ​​to obtain a sub-weight of each normalized probability value in the weight, and the sub-weight is used as the target probability value corresponding to the normalized probability value.

[0082] Specifically, all the normalized probability values ​​are taken as a whole and normalized with the sample smoothed class label value to obtain the weights occupied by all the normalized probability values, and then the weights are evenly distributed to each of the normalized probability values ​​to obtain the sub-weight of each normalized probability value, and the sub-weight is used as the target probability value corresponding to the normalized probability value, and then the target probability value replaces the non-sample smoothed class label value at the corresponding position in the sample initial smooth label to obtain the target smooth label of the training sample. For example, for the above normalized probability values ​​of 0.06, 0.16, 0.31, 0.31, 0.16, it is necessary to consider normalizing them again with the sample smoothed class label value 0.9, that is, the above normalized probability values ​​0.06, 0.16, 0.31, 0.31, 0.16 need to be normalized to 0.1, that is, the probability values ​​of other positions except the origin position in the Gaussian distribution probability distribution of the sample are normalized twice, so as to obtain [0.006, 0.016, 0.031, 0.031, 0.016] corresponding to the above normalized probability values ​​0.06, 0.16, 0.31, 0.31, 0.16 normalized to 0.1, and then [0.006, 0.016, 0.031, 0.031, 0.016] replaces the corresponding non-sample smoothed class label values ​​at each position except 0.9 in the sample initial smoothed labels [0.02, 0.02, 0.02, 0.9, 0.02, 0.02] to obtain the target smoothed labels of the training samples [0.006, 0.016, 0.031, 0.9, 0.031, 0.016].

[0083] The sample label smoothing device of the multi-level classification neural network can be implemented in the form of a computer program. The computer program can be used in Figure 4 Runs on the computer device shown.

[0084] See also Figure 4 , Figure 4 is a schematic block diagram of a computer device provided by an embodiment of the present invention. The computer device may be a sample label smoothing method for executing a multi-level classification neural network.

[0085] See also Figure 4 The computer device 500 includes a processor 502 , a memory and a network interface 505 connected via a system bus 501 , wherein the memory may include a storage medium 503 and an internal memory 504 .

[0086] The storage medium 503 may store an operating system 5031 and a computer program 5032. When the computer program 5032 is executed, the processor 502 may execute a sample label smoothing method of a multi-level classification neural network, wherein the storage medium 503 may be a volatile storage medium or a non-volatile storage medium.

[0087] The processor 502 is used to provide computing and control capabilities to support the operation of the entire computer device 500 .

[0088] The internal memory 504 provides an environment for the operation of the computer program 5032 in the storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute the sample label smoothing method of the multi-layer classification neural network.

[0089] The network interface 505 is used for network communication, such as providing data information transmission, etc. Those skilled in the art will understand that Figure 4 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device 500 to which the solution of the present invention is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0090] The processor 502 is used to run the computer program 5032 stored in the memory to implement the corresponding functions in the sample label smoothing method of the multi-level classification neural network.

[0091] Those skilled in the art will understand that Figure 4 The embodiments of the computer device shown in the figure do not constitute a limitation on the specific composition of the computer device. In other embodiments, the computer device may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently. For example, in some embodiments, the computer device may only include a memory and a processor. In such embodiments, the structure and function of the memory and the processor are the same as those of the embodiment of the present invention. Figure 4 The embodiments shown are consistent and will not be described again here.

[0092] It should be understood that in the embodiment of the present invention, the processor 502 may be a central processing unit (CPU), and the processor 502 may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0093] In another embodiment of the present invention, a computer-readable storage medium is provided. The computer-readable storage medium may be a volatile or non-volatile computer-readable storage medium. The computer-readable storage medium stores a computer program, wherein when the computer program is executed by a processor, the steps included in the sample label smoothing method of the multi-level classification neural network are implemented.

[0094] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, the specific working process of the above-described equipment, devices and units can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here. Those of ordinary skill in the art can appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented with electronic hardware, computer software or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to the function. Whether these functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0095] In the several embodiments provided by the present invention, it should be understood that the disclosed equipment, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. Units with the same function may also be combined into one unit. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices or units, or may be an electrical, mechanical or other form of connection.

[0096] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the embodiments of the present invention.

[0097] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0098] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a computer-readable storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned computer-readable storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a magnetic disk or an optical disk.

[0099] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A sample label smoothing method of a multi-level classification neural network, applied to image blur classification, characterized in that: include: Obtaining a sample initial label corresponding to a training sample, wherein the sample initial label is a one-dimensional matrix, wherein the one-dimensional matrix includes multiple values, each of which is used to describe the probability that the training sample is identified as a classification category corresponding to the value, wherein the multiple values ​​include a sample class label value and a non-sample class label value, wherein the sample class label value is used to describe the probability that the training sample is identified as a classification category with the highest probability, and the non-sample class label value is used to describe the probability that the training sample is identified as a classification category other than the classification category with the highest probability; The sample initial label is smoothed by using a preset initial label smoothing method to obtain the sample initial smoothed label, wherein the sample class label value is correspondingly smoothed to the sample class smoothed label value; Taking the position of the sample class label value in the sample initial label as the symmetry center, assigning position weights from small to large to the positions of the values ​​according to the order from near to far of the position of the values, and obtaining a position weight set including all the position weights; Based on Gaussian distribution, taking the symmetry center as the center, obtaining the distribution weight corresponding to each position weight in the position weight set, and obtaining the sample Gaussian distribution probability; Normalizing the probability values ​​of other positions in the sample Gaussian probability distribution except the origin position to obtain corresponding normalized probability values; All the normalized probability values ​​are taken as a whole and normalized with the sample smoothed class label value to obtain the weight of the whole. Based on the weight, the target probability value corresponding to each normalized probability value is obtained, and the target probability value replaces the non-sample smoothed class label value at the corresponding position in the sample initial smooth label to obtain the target smooth label of the training sample.

2. The sample label smoothing method of the multi-level classification neural network according to claim 1 is characterized in that: The position of the sample class label value in the sample initial label is taken as the symmetry center, and the position of each value is configured with a position weight from small to large in order from the nearest to the symmetry center, so as to obtain a position weight set containing all the position weights, including: Converting the initial labels of the samples into a one-dimensional coordinate system to obtain a one-dimensional coordinate system; Taking the symmetry center as the origin of the one-dimensional coordinate system, taking the origin coordinate of the origin as the preset reference value, and according to the order from near to far of the position of each non-sample class label value from the origin coordinate, increasing the absolute value of the coordinate corresponding to the position of each non-sample class label value in sequence, to obtain the coordinate corresponding to the position of the non-sample class label value; The coordinates are used as position weights, and the sample label coordinates corresponding to the sample initial labels are obtained.

3. The sample label smoothing method of the multi-level classification neural network according to claim 2 is characterized in that: The step of increasing the absolute values ​​of the coordinates corresponding to the positions of the non-sample class label values ​​in sequence according to the order of the positions of the non-sample class label values ​​from the origin coordinates from near to far includes: Taking the preset step length as the coordinate unit, according to the order from near to far of the position of each non-sample class label value from the origin coordinate, the absolute value of the coordinate corresponding to the position of each non-sample class label value is increased in sequence.

4. The sample label smoothing method of the multi-level classification neural network according to claim 1, characterized in that: The obtaining, based on the weight, a target probability value corresponding to each normalized probability value includes: All the normalized probability values ​​divide the weight equally to obtain a sub-weight of each normalized probability value accounting for the weight, and the sub-weight is used as the target probability value corresponding to the normalized probability value.

5. The sample label smoothing method of the multi-level classification neural network according to claim 3 is characterized in that: The step of increasing the absolute values ​​of the coordinates corresponding to the positions of the non-sample class label values ​​in sequence according to the order of the positions of the non-sample class label values ​​from the origin coordinates from near to far includes: If it is determined that the non-sample class label value is moved one category to the right, the corresponding coordinate is increased by 1; If it is determined that the label value of the non-sample class is moved to the left by one category, the corresponding coordinate is reduced by 1.

6. The sample label smoothing method of the multi-level classification neural network according to claim 1, characterized in that: The preset initial label smoothing method adopts a multi-layer classification neural network, and the multi-layer classification neural network is used to classify categories in which there is a layer-by-layer progressive relationship between multiple categories of hierarchical classification.

7. The sample label smoothing method of the multi-level classification neural network according to claim 2, characterized in that: The distribution probability density function of the Gaussian distribution is: Among them, x is the coordinates in the sample label coordinates, c is the mean of the absolute values ​​of the coordinates in the sample label coordinates, and the variance σ is a constant.

8. A sample label smoothing device for a multi-level classification neural network, applied to image blur classification, characterized in that: include: A sample initial label acquisition unit is used to acquire a sample initial label corresponding to a training sample, wherein the sample initial label is a one-dimensional matrix, wherein the one-dimensional matrix includes a plurality of values, each of which is used to describe the probability that the training sample is identified as a classification category corresponding to the value, wherein the plurality of values ​​include a sample class label value and a non-sample class label value, wherein the sample class label value is used to describe the probability that the training sample is identified as a classification category with the highest probability, and the non-sample class label value is used to describe the probability that the training sample is identified as a classification category other than the classification category with the highest probability; A label smoothing unit is used to smooth the sample initial label by using a preset initial label smoothing method to obtain the sample initial smoothed label, wherein the sample class label value is correspondingly smoothed to the sample smoothed class label value; A position weight set acquisition unit is used to configure position weights from small to large for the positions of the values ​​according to the order from near to far of the positions of the values ​​in the symmetry center, taking the position of the sample class label value in the sample initial label as the symmetry center, and obtaining a position weight set including all the position weights; A Gaussian distribution probability acquisition unit, configured to acquire, based on Gaussian distribution and with the symmetry center as the center, the distribution weight corresponding to each position weight in the position weight set, and obtain the sample Gaussian distribution probability; A probability value normalization unit, used to normalize the probability values ​​of other positions in the sample Gaussian probability distribution except the origin position to obtain corresponding normalized probability values; The target smooth label acquisition unit is used to normalize all the normalized probability values ​​as a whole with the sample smoothed class label value to obtain the weight of the whole, and based on the weight, obtain the target probability value corresponding to each normalized probability value, and replace the non-sample smoothed class label value at the corresponding position in the sample initial smooth label with the target probability value to obtain the target smooth label of the training sample.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the sample label smoothing method of the multi-layer classification neural network according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the sample label smoothing method of the multi-layer classification neural network according to any one of claims 1 to 7 is implemented.