A method, device, electronic device and storage medium for identifying sample noise

Abstract

The present application discloses a sample noise identification method, device, electronic device and storage medium. The embodiments of the present application can be applied to various scenarios such as cloud technology, artificial intelligence, smart transportation, and assisted driving. The method includes: obtaining a set of samples to be detected; the set of samples to be detected includes multiple samples to be detected and a category set corresponding to the multiple samples to be detected; obtaining a reference significant feature channel corresponding to each category in the category set; determining the feature vector of each sample to be detected; determining the credibility index of each sample to be detected based on the feature vector of each sample to be detected and the reference significant feature channel corresponding to the category of each sample to be detected; determining the wrong sample from multiple samples to be detected based on the credibility index of each sample to be detected. In this way, useful information for noise detection can be retained while effectively reducing the feature dimension, thereby not only improving the recognition speed of noise samples, but also improving the recognition accuracy.

Description

Technical Field

[0001] The present application relates to the technical field of machine learning, and in particular to a sample noise recognition method, device, electronic device and storage medium. Background Art

[0002] The training of deep neural networks usually requires a large number of correctly labeled samples (clean samples), but in actual application scenarios, obtaining large-scale, high-quality clean labels faces problems such as high labeling cost, high time expenditure, and low labeling quality.

[0003] For example, in crowdsourcing scenarios, companies usually ask several annotators to annotate a large number of unlabeled samples. However, due to the uneven annotation capabilities of different annotators and incorrect annotations, a large amount of annotation costs and time are spent to obtain a noisy labeled dataset. In addition, obtaining samples by searching keywords on the Internet is a low-cost method, but the resulting dataset also contains a large number of noisy samples.

[0004] Therefore, noise recognition has become an increasingly important topic worthy of research in practical application scenarios, among which the detection efficiency of noise samples needs to be improved urgently. Summary of the invention

[0005] In order to solve the problem of low detection efficiency of noise samples in the prior art, the present application provides a sample noise identification method, device, electronic device and storage medium:

[0006] According to a first aspect of the present application, a sample noise identification method is provided, comprising:

[0007] Obtain a set of samples to be detected; the set of samples to be detected includes multiple samples to be detected and category sets corresponding to the multiple samples to be detected; the category set includes the category of each sample to be detected in the multiple samples to be detected;

[0008] Obtain a reference significant feature channel corresponding to each category in the category set; the reference significant feature channel corresponding to each category is determined based on the feature vector of the clean sample corresponding to each category;

[0009] Determine the feature vector of each sample to be detected;

[0010] Determine the credibility index of each sample to be detected based on the feature vector of each sample to be detected and the reference significant feature channel corresponding to the category of each sample to be detected;

[0011] Based on the credibility index of each sample to be detected, an erroneous sample is determined from a plurality of samples to be detected.

[0012] According to a second aspect of the present application, a sample noise identification device is provided, comprising:

[0013] The first acquisition module is used to acquire a set of samples to be detected; the set of samples to be detected includes multiple samples to be detected and a category set corresponding to the multiple samples to be detected; the category set includes the category of each sample to be detected in the multiple samples to be detected;

[0014] The second acquisition module is used to acquire a reference significant feature channel corresponding to each category in the category set; the reference significant feature channel corresponding to each category is determined based on a feature vector of a clean sample corresponding to each category;

[0015] A first determination module is used to determine the feature vector of each sample to be detected;

[0016] A second determination module is used to determine the credibility index of each sample to be detected based on the feature vector of each sample to be detected and the reference significant feature channel corresponding to the category of each sample to be detected;

[0017] The third determination module is used to determine an erroneous sample from a plurality of samples to be detected based on the credibility index of each sample to be detected.

[0018] According to a third aspect of the present application, an electronic device is provided, the electronic device comprising a processor and a memory, the memory storing at least one instruction or at least one program, the at least one instruction or at least one program being loaded and executed by the processor to implement the sample noise identification method of the first aspect of the present application.

[0019] According to the fourth aspect of the present application, a computer storage medium is provided, in which at least one instruction or at least one program is stored, and the at least one instruction or at least one program is loaded and executed by a processor to implement the sample noise identification method of the first aspect of the present application.

[0020] According to the fifth aspect of the present application, a computer program product is provided, the computer program product comprising at least one instruction or at least one program, the at least one instruction or at least one program being loaded and executed by a processor to implement the sample noise identification method of the first aspect of the present application.

[0021] The present application provides a sample noise identification method, device, electronic device and storage medium, which have the following technical effects:

[0022] By obtaining a set of samples to be detected; the set of samples to be detected includes multiple samples to be detected and a set of categories corresponding to the multiple samples to be detected; the category set includes the categories of each sample to be detected in the multiple samples to be detected; obtaining a reference significant feature channel corresponding to each category in the category set; the reference significant feature channel corresponding to each category is determined based on the feature vector of the clean sample corresponding to each category; determining the feature vector of each sample to be detected; determining the credibility index of each sample to be detected based on the feature vector of each sample to be detected and the reference significant feature channel corresponding to the category of each sample to be detected; determining the wrong sample from multiple samples to be detected based on the credibility index of each sample to be detected. The present application is applicable to samples in the form of images or videos, and uses similar features of samples of the same type to screen noise samples in the samples to be detected, and only considers the most significant part of the features of the samples to be detected and the corresponding categories when comparing, so that the useful information for noise detection can be retained while effectively reducing the feature dimension, thereby not only improving the recognition speed of noise samples, but also improving the recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present application or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0024] Figure 1 is a schematic diagram of an application environment provided by an embodiment of the present application;

[0025] Figure 2 It is a flow chart of a sample noise identification method provided in an embodiment of the present application;

[0026] Figure 3 This is a flow chart of obtaining a reference significant feature channel corresponding to each category in a category set provided by an embodiment of the present application;

[0027] Figure 4 It is a schematic diagram of a flow chart for determining a significant feature channel of each clean sample provided by an embodiment of the present application;

[0028] Figure 5 This is a schematic diagram of a specific process for determining a significant feature channel of each clean sample provided by an embodiment of the present application;

[0029] Figure 6 It is a schematic diagram of a flow chart of determining a reference significant feature channel corresponding to each category provided by an embodiment of the present application;

[0030] Figure 7It is a schematic diagram of a flow chart of determining a comparison significant feature channel of each sample to be detected provided by an embodiment of the present application;

[0031] Figure 8 It is a schematic diagram of a process for determining the credibility index of each sample to be detected provided by an embodiment of the present application;

[0032] Fig. 9 It is a schematic diagram of a process for determining the credibility index of each sample to be detected provided by an embodiment of the present application;

[0033] Fig.10 It is a schematic diagram of a process for determining the credibility index of each sample to be detected provided by an embodiment of the present application;

[0034] Fig.11 It is a schematic diagram of a process for determining the credibility index of each sample to be detected provided by an embodiment of the present application;

[0035] Fig.12 is a schematic diagram of a flow chart of determining an error sample provided by an embodiment of the present application;

[0036] Fig.13 is a schematic diagram of a flow chart of determining an error sample provided by an embodiment of the present application;

[0037] Fig.14 It is a block diagram of a sample noise identification device provided in an embodiment of the present application;

[0038] Fig.15 It is a structural schematic diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0039] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0040] It should be noted that the terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. For example, a process, method, system, product or server that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0041] Before further describing the embodiments of the present application in detail, the nouns and terms involved in the embodiments of the present application are explained. The nouns and terms involved in the embodiments of the present application are subject to the following interpretations.

[0042] Noise samples: samples that are incorrectly labeled in the training dataset.

[0043] Noisy learning: Learn a high-performance model on a dataset containing erroneous samples (noise samples).

[0044] Methods for solving the noisy learning problem can be roughly divided into two categories. One is to directly train a robust model in the presence of noisy labels. This type of method usually reduces the negative impact of overfitting to noise samples by designing a network structure that is robust to label noise or introducing a loss function that is robust to noise. The other method is to detect potential noise samples in the dataset. It first detects potential noise samples in the training set and removes them from the training set, and then uses the filtered training set for model training. In actual applications, the latter is more practical in the industry because it not only learns a robust deep learning model, but also provides a relatively clean dataset.

[0045] The embodiment of the present application provides a sample noise identification method, which uses an existing clean sample data set to select the most significant feature channels of clean samples for each category, and the screened out significant feature channels have a strong ability to distinguish clean noise samples. Then, the significant feature set of the sample to be detected is compared with the significant feature channels of its corresponding category, and the samples with relatively large differences are classified as noise samples. This method uses similar features of samples of the same type to screen noise samples in the samples to be detected, and only considers the most significant part of the features of the sample to be detected and the corresponding category when comparing. It can effectively reduce the feature dimension while retaining useful information for noise detection, thereby not only improving the recognition speed of noise samples, but also improving the recognition accuracy.

[0046] See also Figure 1 , Figure 1 1 is a schematic diagram of an application environment provided by an embodiment of the present application, and the application environment may include a client 10 and a server 20. The client 10 and the server 20 may be directly or indirectly connected via wired or wireless communication.

[0047] In some possible embodiments, the client 10 collects images or videos through a camera, and transmits a large number of collected images or videos to the server 20 through a network. After the images or videos are labeled by categories, they can be used as a set of samples to be detected. The server 20 provides a sample noise identification service, that is, to detect whether each sample to be detected in the set of samples to be detected belongs to the labeled category, and to identify the samples that are incorrectly labeled.

[0048] The embodiments of the present application can be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, smart transportation, assisted driving, etc. When applied to the field of smart transportation, the sample to be detected can be road environment information collected by road-side sensing equipment, such as road environment images or road environment videos; or, when applied to the field of assisted driving, the sample to be detected can be vehicle surrounding environment information collected by vehicle-side sensing equipment, such as vehicle surrounding environment images or vehicle surrounding environment videos.

[0049] The client 10 may be a physical device such as a smart phone, a computer (such as a desktop computer, a tablet computer, a laptop computer), an augmented reality (AR) / virtual reality (VR) device, a digital assistant, an intelligent voice interaction device (such as an intelligent speaker), an intelligent wearable device, an intelligent home appliance, a vehicle-mounted terminal, etc., or may be software running in a physical device, such as a computer program. The operating system corresponding to the client may be an Android system (Android system), an iOS system (a mobile operating system developed by Apple), a Linux system (an operating system), a Microsoft Windows system (Microsoft Windows operating system), etc.

[0050] The server 20 may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. The server may include a network communication unit, a processor, a memory, etc. The server may provide background services for the corresponding client.

[0051] It should be noted that Figure 1 This is just an example. For example, another application scenario may also only include the server 20. The server 20 obtains a set of samples to be detected from a database and identifies erroneous samples in the set of samples to be detected.

[0052] The following describes a specific embodiment of a sample noise identification method of the present application. Figure 2 It is a flowchart of a sample noise identification method provided in an embodiment of the present application. The present application provides method operation steps as described in the embodiment or flowchart, but may include more or fewer operation steps based on conventional or non-creative labor. The order of steps listed in the embodiment is only one way of executing the steps among many orders, and does not represent the only order of execution. When the actual system or product is executed, it can be executed in the order of the method shown in the embodiment or the drawings or in parallel (for example, in a parallel processor or multi-threaded processing environment). Specifically, Figure 2 As shown, the method may include:

[0053] S201: Obtain a set of samples to be detected.

[0054] The set of samples to be detected includes multiple samples to be detected and category sets corresponding to the multiple samples to be detected; the category set includes the category of each sample to be detected in the multiple samples to be detected.

[0055] In the embodiments of the present application, each sample to be detected can be a sample of any form, such as an image sample, a video sample, etc. The categories of the above-mentioned samples to be detected are labeled categories, that is, categories that have been manually labeled. Due to the uneven labeling capabilities of different labelers and incorrect labeling, the labeled categories of each sample to be detected in the set of samples to be detected may not be actual categories. Therefore, the present application identifies the samples that are incorrectly labeled in the set of samples to be detected through steps S201 to S209. The samples that are incorrectly labeled are the noise in the set of samples to be detected. After the noise is identified and deleted, the sample set can be used for model training.

[0056] S203: Obtain a reference significant feature channel corresponding to each category in the category set.

[0057] Among them, the reference salient feature channel corresponding to each category is determined based on the feature vector of the clean sample corresponding to each category.

[0058] In the embodiment of the present application, the feature vector is obtained after feature extraction of the sample. The feature vector records the feature information of the sample, and the feature information can be used to classify and identify the sample. In the present application, the feature vector is used to identify whether the sample is mislabeled. Since samples of the same type have similar features, the present application identifies the corresponding sample to be detected in the sample set to be detected based on the feature vector of the clean sample corresponding to each category.

[0059] In the related art, the feature vector obtained after the sample passes through the feature extraction layer in the deep neural network model usually has a high feature dimension, and the high feature dimension easily leads to the sample being too sparse in the high-dimensional feature space (this is also called the curse of dimensionality). This leads to the identification of wrong samples based on the similarity with clean samples. The distance between all sample points in the high-dimensional space and the category center feature value determined based on the clean samples is very far, that is, the similarity between all samples and the clean samples is relatively low, which makes the correct samples easy to be misjudged, resulting in a low accuracy rate of noise samples.

[0060] Based on this, the present application selects a more significant feature channel from the feature vector corresponding to each category in the category set as the reference significant feature channel corresponding to each category, and identifies the corresponding sample to be detected in the sample set to be detected based on the reference significant feature channel corresponding to each category. In this way, the dimension reduction of the feature can be achieved, and at the same time, the useful information for noise identification can be retained, which is conducive to improving the speed and accuracy of sample noise identification.

[0061] Among them, each feature channel in the feature vector represents the performance of the sample on this feature. For example, a feature channel of an image represents the performance of a certain image sample in color saturation. A more significant feature channel means that the sample has a more obvious performance on the corresponding feature, and the channel amplitude can be used to quantify the degree of performance of different samples on this feature.

[0062] In some possible embodiments, the reference salient feature channel corresponding to each category in the above-mentioned acquisition category set may include: Figure 3 The following steps are shown:

[0063] S301: Obtain multiple clean samples corresponding to each category in the category set.

[0064] Among them, clean samples refer to samples that are correctly labeled.

[0065] In a specific embodiment, the multiple clean samples corresponding to each category may be another sample set different from the sample set to be detected, and the category of each sample in the sample set is correctly labeled.

[0066] Alternatively, in another specific embodiment, the multiple clean samples corresponding to each category may also come from the set of samples to be detected after the previous round of noise screening; that is, the set of samples to be detected may be subjected to multiple rounds of noise identification and deletion to obtain a relatively clean set of samples, and steps S201 to S209 are performed in each round of noise identification; then, in each round of identification, the erroneous samples identified in the previous round are deleted to obtain a new set of samples to be detected, and the multiple clean samples corresponding to each category are determined from the new set of samples to be detected. In the first round of identification, each sample to be detected in the set of samples to be detected may be considered a clean sample.

[0067] S303: Determine a feature vector of each clean sample among a plurality of clean samples.

[0068] The feature vector of each clean sample is obtained after feature extraction of the clean sample.

[0069] See also Figure 4 , Figure 4 It is a flow chart of determining the significant feature channels of each clean sample provided by an embodiment of the present application. In a specific embodiment, the clean sample is input into the deep neural network model to obtain the feature vector output by the feature extraction layer. The feature vector of each clean sample is subjected to the subsequent step S305 to obtain the significant feature channel of each clean sample.

[0070] The above-mentioned deep neural network model may be a deep neural network model that has not completed training. The present application may train the deep neural network model based on the sample set to be detected, that is, the noise recognition of the sample set to be detected may be completed during the model training process.

[0071] S305: Determine the significant feature channel of each clean sample from the feature vector of each clean sample.

[0072] In a specific embodiment, this step may include: obtaining the amplitude of each feature channel in the feature vector of each clean sample; sorting the amplitude of each feature channel, and using the feature channel with the first preset number of bits before sorting as the significant feature channel of each clean sample.

[0073] As mentioned above, a more significant feature channel refers to a sample that performs more obviously on the corresponding feature, and the degree of performance of different samples on this feature can be quantified by the channel amplitude. Therefore, in this specific embodiment, for each clean sample, the amplitude of each feature channel in its feature vector is sorted, and the feature channel with the first preset number of bits before sorting is used as the significant feature channel of each clean sample; wherein the first preset number of bits can be determined according to the number of feature channels in actual applications. For example, when there are 2048 feature channels, the corresponding first preset number of bits can be 200.

[0074] In another specific embodiment, when the sample form is a video form, each clean sample may include multiple frames of video images, and accordingly, the feature vector of each clean sample may include a feature vector of each frame of video images in the multiple frames of video images;

[0075] Correspondingly, the above-mentioned determination of the significant feature channel of each clean sample from the feature vector of each clean sample may include the following steps:

[0076] S3051: Perform feature aggregation on multiple feature vectors corresponding to multiple frames of video images to obtain a first aggregated feature vector.

[0077] See also Figure 5 , Figure 5 1 is a schematic diagram of a specific process for determining a significant feature channel of each clean sample provided by an embodiment of the present application. Figure 5 As shown, each frame of video images in multiple frames of video images can obtain a corresponding feature vector through the feature extraction layer of the deep neural network model. For samples in video form, after the feature extraction layer extracts the features of each frame of image in the video, the deep neural network model will also aggregate the feature vectors corresponding to each frame of video image through a feature aggregation function to obtain an aggregated feature vector, i.e., a first aggregated feature vector, which is used to characterize the features of the video.

[0078] S3053: Calculate the inter-frame variance based on the feature vector of each frame of the video image to obtain a first variance vector.

[0079] S3055: Determine a feature channel corresponding to the first key channel information from the first aggregated feature vector, and use the feature channel corresponding to the first key channel information as a significant feature channel of the multi-frame video image.

[0080] This application considers that the temporal relationship between different frames in a video also contains information that is beneficial to noise detection. Therefore, Figure 5 As shown, the present application calculates the inter-frame variance based on the feature vector of each frame of the video image to obtain a first variance vector; the first variance vector contains the first key channel information; then, the first aggregate feature vector representing the video feature is screened according to the first key channel information, that is, the feature channel corresponding to the first key channel information is determined from the first aggregate feature vector, and the feature channel corresponding to the first key channel information is used as the significant feature channel of the video sample. In this way, in the process of reducing the dimension of the video feature, more representative significant feature channels can be obtained by using inter-frame features.

[0081] S307: Based on the salient feature channels of each clean sample, determine a reference salient feature channel corresponding to each category.

[0082] See also Figure 6 , Figure 6 This is a flow chart of determining a reference significant feature channel corresponding to each category provided by an embodiment of the present application. In a specific embodiment, this step may include: sorting multiple significant feature channels corresponding to multiple clean samples according to the number of occurrences, and using the significant feature channel with the second preset number of bits before sorting as the reference significant feature channel corresponding to each category.

[0083] In this specific embodiment, for each category, the number of occurrences of each significant feature channel of each clean sample under the category is counted, and then all significant feature channels are sorted according to the number of occurrences, and the significant feature channels with the second preset number of digits in front of the order are used as the reference significant feature channels corresponding to each category. For example, the number of significant feature channels of each clean sample is 200. Assuming that there are 10 clean samples under a certain category, the 2000 significant feature channels corresponding to the 10 clean samples are counted, and there are repeated significant feature channels in the 2000 significant feature channels, so as to determine the number of repetitions of the repeated significant feature channels, that is, the number of occurrences; the more the number of occurrences of the significant feature channel is, it means that different clean samples under the category all perform outstandingly on the significant feature channel, and the significant feature channel is representative for the category, so it can be used as the reference significant feature channel corresponding to the category. Among them, the second preset number of digits can be determined according to the number of feature channels and the number of significant feature channels of each clean sample in actual application. For example, when there are 2048 feature channels and 200 significant feature channels, the second preset number of digits can be 200.

[0084] S205: Determine the feature vector of each sample to be detected.

[0085] In the embodiment of the present application, feature extraction can be performed on each sample to be detected in the sample set to be detected through the feature extraction layer in the deep neural network model to obtain a feature vector of each sample to be detected.

[0086] S207: Determine the credibility index of each sample to be detected based on the feature vector of each sample to be detected and the reference significant feature channel corresponding to the category of each sample to be detected.

[0087] In an embodiment of the present application, the credibility index of each sample to be detected refers to the degree of confidence that each sample to be detected is a clean sample of the labeled category; the credibility index can be calculated based on the feature vector of each sample to be detected and the reference significant feature channel corresponding to the category of each sample to be detected.

[0088] Two embodiments of step S207 are specifically described below. First, the first embodiment of step S207 is described.

[0089] In some possible embodiments, the above S207: determining the credibility index of each sample to be detected based on the feature vector of each sample to be detected and the reference significant feature channel corresponding to the category of each sample to be detected may include:

[0090] S2071: determining a comparison significant feature channel of each sample to be detected from a feature vector of each sample to be detected according to a reference significant feature channel corresponding to the category of each sample to be detected;

[0091] In a specific embodiment, step S2071 may include: Figure 7 The following steps are shown:

[0092] S701: Determine the channel identifier of the reference significant feature channel corresponding to the category of each sample to be detected.

[0093] S703: According to the channel identifier, determine the corresponding feature channel from the feature vector of each sample to be detected.

[0094] S705: Using the corresponding feature channel as a comparison significant feature channel of each sample to be detected.

[0095] In this specific embodiment, the data dimensions of each sample to be detected and the corresponding clean sample are the same, and the data dimensions of the feature vector of each sample to be detected and the feature vector of the corresponding clean sample are also the same; thus, Figure 8 As shown, Figure 8 This is a flow chart of determining the credibility index of each sample to be detected provided by the present application; for each sample to be detected, the feature channel corresponding to the channel identifier of the reference significant feature channel corresponding to its category can be directly determined from its feature vector to obtain a compared significant feature channel.

[0096] S2073: Determine the credibility index of each sample to be detected according to the compared significant feature channels of each sample to be detected.

[0097] In a specific embodiment, step S2073 may include: Fig. 9 The following steps are shown:

[0098] S901: Determine the number of samples to be tested corresponding to each category.

[0099] First, for each category in the category set, the number of samples to be detected under this category in the sample set to be detected is determined.

[0100] S903: Determine the category center feature value corresponding to each category according to the compared significant feature channels of each sample to be detected and the number of samples to be detected corresponding to each category.

[0101] Secondly, for each category, the following formula (1) can be used to determine the category center feature value corresponding to each category:

[0102]

[0103] Among them, c k represents the category center eigenvalue corresponding to category k; m k Indicates the number of samples to be tested corresponding to category k in the set of samples to be tested; Represents the comparison significant feature channel of the sample to be detected i marked as category k.

[0104] S905: Perform an inner product between the compared significant feature channels of each sample to be detected and the category center feature value corresponding to the category of each sample to be detected, so as to obtain a credibility index of each sample to be detected.

[0105] Secondly, for each sample to be tested, such as Figure 8 As shown, the inner product of the comparison significant feature channel of each sample to be detected and the category center feature value corresponding to the category of each sample to be detected is performed to obtain the credibility index of each sample to be detected; the credibility index of each sample to be detected can be determined using the following formula (2):

[0106]

[0107] in, represents the credibility index of the sample i to be tested which is labeled as category k; c k Represents the category center eigenvalue corresponding to category k; represents the comparison significant feature channel of the sample i to be tested which is marked as category k; · represents the inner product calculation.

[0108] A second embodiment of step S207 is described below.

[0109] In some possible embodiments, the above S207: determining the credibility index of each sample to be detected based on the feature vector of each sample to be detected and the reference significant feature channel corresponding to the category of each sample to be detected may also include:

[0110] S2071′: Determine the comparison significant feature channel of each sample to be detected from the feature vector of each sample to be detected.

[0111] In a specific embodiment, step S2071' may include: obtaining the amplitude of each feature channel in the feature vector of each sample to be detected; sorting the amplitude of each feature channel, and using the feature channel with the third preset number of bits before sorting as the comparison significant feature channel of each sample to be detected.

[0112] In this specific embodiment, the comparison significant feature channel of each sample to be detected is selected based on the amplitude of each feature channel of each sample to be detected, and reference can be made to the embodiment of selecting the significant feature channel of each clean sample based on the amplitude of each feature channel of each clean sample in step S305 above; wherein, since the comparison significant feature channel of each sample to be detected needs to be compared with the reference significant feature channel corresponding to the category of each sample to be detected in the subsequent steps, the third preset number of bits is equal to the second preset number of bits in the above text. This specific embodiment is applicable to different forms of samples to be detected, including image samples to be detected and video samples to be detected.

[0113] In another specific embodiment, when the sample form is a video form, each sample to be detected may include multiple frames of video images to be detected; accordingly, the feature vector of each sample to be detected may include the feature vector of each frame of the video images to be detected in the multiple frames of video images to be detected;

[0114] Correspondingly, the above-mentioned step S2071' may include: performing feature aggregation on multiple feature vectors corresponding to multiple frames of video images to be detected to obtain a second aggregated feature vector; performing inter-frame variance calculation based on the feature vector of each frame of video images to be detected to obtain a second variance vector; the second variance vector includes second key channel information; and according to the second key channel information, determining the comparison significant feature channel of each sample to be detected from the second aggregated feature vector.

[0115] In this specific embodiment, for determining the comparison significant feature channel for the sample to be detected in the form of a video, reference can be made to the embodiment of determining the significant feature channel for the clean sample in the form of a video in steps S3051 to S3055 above; no further details will be given here.

[0116] S2073': Determine the credibility index of each sample to be detected according to the compared significant feature channels of each sample to be detected and the reference significant feature channels corresponding to the category of each sample to be detected.

[0117] In a specific embodiment, step S2073' may include: Fig.10 The following steps are shown:

[0118] S1001: Determine the number of identical feature channels between the compared significant feature channels of each sample to be detected and the reference significant feature channels corresponding to the category of each sample to be detected.

[0119] S1003: Using the number of channels with the same characteristics as the reliability index of each sample to be detected.

[0120] like Fig.11 As shown, Fig.11This is another flow chart of determining the credibility index of each sample to be detected provided by an embodiment of the present application; solving the intersection between the comparison significant feature channel of each sample to be detected and the reference significant feature channel corresponding to the category of each sample to be detected, that is, the number of the same feature channels, and taking the number of the same feature channels as the credibility index of each sample to be detected. The credibility index of each sample to be detected can be determined according to the following formula (3):

[0121]

[0122] in, represents the credibility index of the sample i to be tested that is labeled as category k; Card(·) represents the number of elements in the calculation set; represents the comparison significant feature channel of the sample i to be tested which is marked as category k; L k represents the reference salient feature channel corresponding to category k.

[0123] S209: Determine an erroneous sample from a plurality of samples to be detected based on the credibility index of each sample to be detected.

[0124] In the embodiment of the present application, after determining the credibility index of each sample to be detected, the error samples in the set of samples to be detected are determined based on the credibility index of each sample to be detected. The error samples are subsequently deleted to obtain a relatively clean sample set, and the above steps S201 to S209 can be repeatedly performed until a clean sample set is obtained, and the clean sample set can be used for subsequent model training.

[0125] In some possible embodiments, in order to balance the differences in the credibility indicators of the samples to be detected in different categories in the category set, in this step, the present application may normalize the credibility indicators of the samples to be detected in each category.

[0126] Thus, the above-mentioned method of determining the error sample from multiple samples to be detected based on the credibility index of each sample to be detected may include: Fig.12 The following steps are shown:

[0127] S1201: Based on the credibility index of each sample to be detected, determine the average credibility index corresponding to each category in the set of samples to be detected.

[0128] S1203: normalizing the credibility index of each sample to be detected based on the average credibility index corresponding to each category to obtain the processed credibility index of each sample to be detected.

[0129] In this step, the credibility index of each sample to be tested after processing can be obtained according to the following formula (4):

[0130]

[0131] in, Represents the credibility index of each sample to be tested after processing; Represents the average credibility index corresponding to category k.

[0132] S1205: Fitting a Gaussian mixture model based on the processed credibility index of each sample to be detected to obtain a category deviation index of each sample to be detected.

[0133] S1207: Determine the samples to be detected whose category deviation index is greater than or equal to the preset index among the multiple samples to be detected as erroneous samples.

[0134] In this step, if Fig.13 As shown, Fig.13 The figure is a flow chart of determining an erroneous sample provided by an embodiment of the present application. A two-component Gaussian mixture model π(·) is fitted using the credibility index of all samples to be detected in the sample set to be detected, and the category deviation index of the sample to be detected is defined. Therefore, according to the fitting results, the category deviation index of each sample to be detected can be obtained, and the category deviation index is the possibility that the sample to be detected is noise. The larger the category deviation index, the greater the possibility that the sample to be detected is noise. Therefore, the sample to be detected whose category deviation index is greater than or equal to the preset index among multiple samples to be detected is determined as an erroneous sample; wherein the preset index is the maximum tolerable category deviation index, which can be 0.5; therefore, when the category deviation index is less than 0.5, the sample to be detected is a clean sample that is correctly labeled; when the category deviation index is greater than or equal to 0.5, the sample to be detected is an erroneous sample that is incorrectly labeled.

[0135] In the embodiments of the present application, the above-mentioned sample noise recognition method was experimentally tested based on an image dataset and a video dataset, respectively.

[0136] First, for the image data set, the noise in the image data set is identified by using the sample noise identification method of determining the comparison of significant feature channels based on amplitude in step S2071' in the above embodiment and determining the credibility index based on intersection in steps S1001 to S1003. During the experiment, PreAct-ResNet18 was selected as the model network structure of the experiment, and experiments were conducted on CIFAR10 and CIFAR100 image classification data sets. The noise in the two data sets is artificially generated and controllable, so in this experiment, two types of noise are artificially constructed: symmetric noise and asymmetric noise.

[0137] Symmetric noise is caused by each sample in the sample set being independently assigned to a random label instead of its true label, and the probability of symmetric noise is uniformly distributed. In this experiment, symmetric noise is defined in both CIFAR10 and CIFAR100, and the noise ratio is assigned to 20%, 50%, 80%, and 90%.

[0138] Asymmetric noise is caused by the fact that all samples in a class can only be assigned to a specific class other than the true label. In this experiment, asymmetric noise is defined on CIFAR10, and the probability of a sample being mislabeled is set to 40%.

[0139] In this experiment, based on the CIFAR10 and CIFAR100 data sets, the sample noise recognition method of the embodiment of the present application and the existing noise recognition algorithm were verified respectively, and the test classification accuracy obtained by verification is shown in the following Table 1:

[0140]

[0141]

[0142] Table 1 Test accuracy of noise recognition experiment on CIFAR dataset

[0143] Table 1 above records the best accuracy achieved by the method on the dataset in all rounds during training, where A represents asymmetric noise, S represents symmetric noise, and - represents the absence of such experimental results.

[0144] It can be seen from the results in Table 1 above that under the sample noise recognition method provided in the embodiment of the present application, the obtained model has the highest accuracy and is better than the existing noise recognition algorithm. This proves that the sample noise recognition method provided in the embodiment of the present application, when applied to the image sample, can effectively identify the noise in the image sample set to be detected, thereby helping to improve the accuracy of the image classification model in the subsequent image classification model training process.

[0145] For the video data set, the present application selects three embodiments 1 to 3 based on the above, embodiment 1: a sample noise identification method with clean sample guidance, which determines and compares significant feature channels based on the amplitude in step S2071' and determines the credibility index based on the intersection in steps S1001 to S1003; embodiment 2: a sample noise identification method without clean sample guidance, which determines and compares significant feature channels based on the amplitude in step S2071' and determines the credibility index based on the inner product of steps S901 to S905; embodiment 3: a sample noise identification method without clean sample guidance, which determines and compares significant feature channels based on the inter-frame variance in step S2071' and determines the credibility index based on the inner product of steps S901 to S905;

[0146] In this experiment, TSM-ResNet50 is selected as the basic structure. Experiments are conducted on three large-scale video classification datasets, namely Mini-kinetics (K200), kinetics (K400) and Something-Something-V1 (SthV1). Two types of noise are constructed in this experiment: symmetric noise and asymmetric noise. In symmetric noise, the noise ratio is distributed as 20%, 40%, 60% and 80%; in asymmetric noise, the probability of sample mislabeling is set to 10%, 20%, 40%.

[0147] In this experiment, based on the K200, K400 and SthV1 data sets, the sample noise recognition method of the embodiment of the present application and the existing noise recognition algorithm were verified respectively. The existing noise recognition algorithm includes Co-teaching, Topofilter, and M-correction. The test classification accuracy obtained by verification is shown in Tables 2 to 4 below:

[0148]

[0149] Table 2 Test accuracy of K200 dataset

[0150]

[0151] Table 3 Test accuracy of K400 dataset

[0152]

[0153]

[0154] Table 4 Test accuracy of SthV1 dataset

[0155] Tables 2 to 4 above record the optimal accuracy achieved by the method on the data set in all rounds during the training period. From Tables 2 to 4 above, it can be seen that under the sample noise recognition method provided in the embodiment of the present application, the obtained model accuracy is the highest, which is better than the existing noise recognition algorithm. This proves that the sample noise recognition method provided in the embodiment of the present application, when applied to video samples, can effectively identify the noise in the set of video samples to be detected, thereby helping to improve the accuracy of the video classification model in the subsequent video classification model training process.

[0156] In summary, a sample noise recognition method provided in an embodiment of the present application is suitable for samples in the form of images or videos. It uses similar features of samples of the same type to screen noise samples in the samples to be detected, and only considers the most significant features of the samples to be detected and the corresponding categories during comparison. In this way, it is possible to effectively reduce the feature dimension while retaining useful information for noise detection, thereby improving not only the recognition speed of noise samples but also the recognition accuracy.

[0157] The present application also provides a sample noise identification device, such as Fig.14 As shown, the sample noise identification device 140 includes:

[0158] The first acquisition module 1401 is used to acquire a set of samples to be detected; the set of samples to be detected includes multiple samples to be detected and a category set corresponding to the multiple samples to be detected; the category set includes the category of each sample to be detected in the multiple samples to be detected;

[0159] The second acquisition module 1402 is used to acquire a reference significant feature channel corresponding to each category in the category set; the reference significant feature channel corresponding to each category is determined based on a feature vector of a clean sample corresponding to each category;

[0160] The first determination module 1403 is used to determine the feature vector of each sample to be detected;

[0161] A second determination module 1404 is used to determine the credibility index of each sample to be detected based on the feature vector of each sample to be detected and the reference significant feature channel corresponding to the category of each sample to be detected;

[0162] The third determination module 1405 is used to determine an erroneous sample from a plurality of samples to be detected based on the credibility index of each sample to be detected.

[0163] In some possible embodiments, the second acquisition module 1402 is also used to obtain multiple clean samples corresponding to each category in the category set; determine the feature vector of each clean sample in the multiple clean samples; determine the significant feature channel of each clean sample from the feature vector of each clean sample; and determine the reference significant feature channel corresponding to each category based on the significant feature channel of each clean sample.

[0164] In some possible embodiments, the second acquisition module 1402 is further used to obtain the amplitude of each feature channel in the feature vector of each clean sample; sort the amplitude of each feature channel, and use the feature channel with the first preset number of bits before sorting as the significant feature channel of each clean sample.

[0165] In some possible embodiments, each clean sample includes multiple frames of video images; the feature vector of each clean sample includes a feature vector of each frame of video images in the multiple frames of video images;

[0166] The second acquisition module 1402 is also used to perform feature aggregation on multiple feature vectors corresponding to multiple frames of video images to obtain a first aggregated feature vector; perform inter-frame variance calculation based on the feature vector of each frame of video image to obtain a first variance vector; the first variance vector includes first key channel information; determine the feature channel corresponding to the first key channel information from the first aggregated feature vector, and use the feature channel corresponding to the first key channel information as a significant feature channel of the multiple frames of video images.

[0167] In some possible embodiments, the second acquisition module 1402 is further used to sort the multiple significant feature channels corresponding to the multiple clean samples according to the number of occurrences, and use the significant feature channels with the second preset number of bits before sorting as reference significant feature channels corresponding to each category.

[0168] In some possible embodiments, the second determination module 1404 is also used to determine the channel identification of the reference significant feature channel corresponding to the category of each sample to be detected; determine the corresponding feature channel from the feature vector of each sample to be detected according to the channel identification; use the corresponding feature channel as the comparison significant feature channel of each sample to be detected; and determine the credibility index of each sample to be detected according to the comparison significant feature channel of each sample to be detected.

[0169] In some possible embodiments, the second determination module 1404 is also used to determine the number of samples to be detected corresponding to each category; determine the category center feature value corresponding to each category based on the compared significant feature channels of each sample to be detected and the number of samples to be detected corresponding to each category; perform an inner product between the compared significant feature channels of each sample to be detected and the category center feature value corresponding to the category of each sample to be detected to obtain a credibility index of each sample to be detected.

[0170] In some possible embodiments, the second determination module 1404 is also used to determine the comparison significant feature channel of each sample to be detected from the feature vector of each sample to be detected; determine the number of identical feature channels between the comparison significant feature channel of each sample to be detected and the reference significant feature channel corresponding to the category of each sample to be detected; and use the number of identical feature channels as a credibility indicator of each sample to be detected.

[0171] In some possible embodiments, the second determination module 1404 is further used to obtain the amplitude of each feature channel in the feature vector of each sample to be detected; sort the amplitude of each feature channel, and use the feature channel with the third preset number of bits before sorting as the comparison significant feature channel of each sample to be detected.

[0172] In some possible embodiments, each sample to be detected includes multiple frames of video images to be detected; the feature vector of each sample to be detected includes the feature vector of each frame of video images to be detected in the multiple frames of video images to be detected;

[0173] The second determination module 1404 is also used to perform feature aggregation on multiple feature vectors corresponding to multiple frames of video images to be detected to obtain a second aggregated feature vector; perform inter-frame variance calculation based on the feature vector of each frame of video images to be detected to obtain a second variance vector; the second variance vector includes second key channel information; and determine the comparison significant feature channel of each sample to be detected from the second aggregated feature vector based on the second key channel information.

[0174] In some possible embodiments, the third determination module 1405 is also used to determine the average credibility index corresponding to each category in the set of samples to be detected based on the credibility index of each sample to be detected; normalize the credibility index of each sample to be detected based on the average credibility index corresponding to each category to obtain the credibility index of each processed sample to be detected; fit a Gaussian mixture model based on the credibility index of each processed sample to be detected to obtain the category deviation index of each sample to be detected; determine the sample to be detected whose category deviation index is greater than or equal to the preset index among multiple samples to be detected as an erroneous sample.

[0175] In some possible embodiments, the first determination module 1403 is further configured to extract features from each sample to be detected in the sample set to be detected, so as to obtain a feature vector of each sample to be detected.

[0176] It should be noted that the device in the device embodiment and the method embodiment are based on the same inventive concept.

[0177] An embodiment of the present application provides an electronic device, which includes a processor and a memory, wherein the memory stores at least one instruction or at least one program, and the at least one instruction or the at least one program is loaded and executed by the processor to implement the sample noise identification method provided in the above method embodiment.

[0178] Further, Fig.15 The hardware structure diagram of an electronic device for implementing the sample noise identification method provided in the embodiment of the present application is shown. The electronic device may participate in forming or include the sample noise identification device provided in the embodiment of the present application. Fig.15As shown, the electronic device 100 may include one or more (1002a, 1002b, ..., 1002n are used to illustrate) processors 1002 (the processor 1002 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 1004 for storing data, and a transmission device 1006 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the I / O interface), a network interface, a power supply and / or a camera. It can be understood by those skilled in the art that Fig.15 The structure shown is only for illustration and does not limit the structure of the above electronic device. Fig.15 More or fewer components as shown, or with Fig.15 Different configurations shown.

[0179] It should be noted that the one or more processors 1002 and / or other data processing circuits described above may generally be referred to herein as "data processing circuits". The data processing circuits may be embodied in whole or in part as software, hardware, firmware, or any other combination thereof. In addition, the data processing circuit may be a single independent processing module, or may be incorporated in whole or in part into any of the other components in the electronic device 100 (or mobile device). As described in the embodiments of the present application, the data processing circuit acts as a processor control (e.g., selection of a variable resistor terminal path connected to an interface).

[0180] The memory 1004 can be used to store software programs and modules of application software, such as program instructions / data storage devices corresponding to the sample noise identification method described in the embodiment of the present application. The processor 1002 executes various functional applications and data processing by running the software programs and modules stored in the memory 1004, that is, to implement the above-mentioned sample noise identification method. The memory 1004 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 1004 may further include a memory remotely arranged relative to the processor 1002, and these remote memories may be connected to the electronic device 100 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0181] The transmission device 1006 is used to receive or send data via a network. The specific example of the above network may include a wireless network provided by a communication provider of the electronic device 100. In one example, the transmission device 1006 includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one embodiment, the transmission device 1006 can be a radio frequency (Radio Frequency, RF) module, which is used to communicate with the Internet wirelessly.

[0182] The display may be, for example, a touch screen liquid crystal display (LCD) that enables a user to interact with a user interface of the electronic device 100 (or mobile device).

[0183] An embodiment of the present application also provides a computer-readable storage medium, which can be set in an electronic device to store at least one instruction or at least one program related to implementing a sample noise identification method in a method embodiment. The at least one instruction or the at least one program is loaded and executed by the processor to implement the sample noise identification method provided in the above method embodiment.

[0184] Optionally, in this embodiment, the storage medium may be located in at least one of the multiple network servers of the computer network. Optionally, in this embodiment, the storage medium may include, but is not limited to, various media that can store program codes, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk, or an optical disk.

[0185] It should be noted that the above-mentioned sequence of the embodiments of the present application is for description only and does not represent the advantages and disadvantages of the embodiments. The above-mentioned specific embodiments of the present application are described. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0186] Each embodiment in this application is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device and electronic device embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.

[0187] A person skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware or by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, and the above-mentioned storage medium may be a read-only memory, a disk or an optical disk, etc.

[0188] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A sample noise identification method, characterized in that: include: Acquire a set of samples to be detected; the set of samples to be detected includes a plurality of samples to be detected and a category set corresponding to the plurality of samples to be detected; the category set includes the category of each sample to be detected in the plurality of samples to be detected; Obtaining a reference significant feature channel corresponding to each category in the category set; The reference significant feature channel corresponding to each category is determined based on the feature vector of the clean sample corresponding to each category; The obtaining of a reference significant feature channel corresponding to each category in the category set includes: Acquire multiple clean samples corresponding to each category in the category set; determine the feature vector of each clean sample in the multiple clean samples; obtain the amplitude of each feature channel in the feature vector of each clean sample; sort the amplitude of each feature channel, and use the feature channel with the first preset number of bits before sorting as the significant feature channel of each clean sample; determine the reference significant feature channel corresponding to each category based on the significant feature channel of each clean sample; each clean sample includes multiple frames of video images; Determine the feature vector of each sample to be detected; Determining the credibility index of each sample to be detected based on the feature vector of each sample to be detected and the reference significant feature channel corresponding to the category of each sample to be detected; Based on the credibility indicators of the samples to be detected, an erroneous sample is determined from the multiple samples to be detected.

2. The sample noise identification method according to claim 1, characterized in that: The feature vector of each clean sample includes a feature vector of each frame of video image in the multiple frames of video images; Determining the significant feature channel of each clean sample from the feature vector of each clean sample includes: Performing feature aggregation on a plurality of feature vectors corresponding to the plurality of frames of video images to obtain a first aggregated feature vector; Based on the feature vector of each frame of the video image, inter-frame variance calculation is performed to obtain a first variance vector; the first variance vector includes first key channel information; A feature channel corresponding to the first key channel information is determined from the first aggregated feature vector, and the feature channel corresponding to the first key channel information is used as a significant feature channel of the multiple frames of video images.

3. The sample noise identification method according to claim 1 or 2, characterized in that: The determining, based on the salient feature channels of the clean samples, the reference salient feature channels corresponding to each category comprises: The plurality of significant feature channels corresponding to the plurality of clean samples are sorted according to the number of occurrences, and the significant feature channels with a second preset number of digits before sorting are used as reference significant feature channels corresponding to each category.

4. The sample noise identification method according to claim 1, characterized in that: The determining of the credibility index of each sample to be detected based on the feature vector of each sample to be detected and the reference significant feature channel corresponding to the category of each sample to be detected includes: Determine the channel identifier of the reference significant feature channel corresponding to the category of each sample to be detected; According to the channel identifier, determining a corresponding feature channel from the feature vectors of each sample to be detected; Using the corresponding characteristic channels as the comparison significant characteristic channels of each sample to be detected; The credibility index of each sample to be detected is determined according to the compared significant feature channels of each sample to be detected.

5. The sample noise identification method according to claim 4, characterized in that: Determining the credibility index of each sample to be detected according to the compared significant feature channels of each sample to be detected includes: Determine the number of samples to be tested corresponding to each category; Determine the category center feature value corresponding to each category according to the compared significant feature channels of each sample to be detected and the number of samples to be detected corresponding to each category; The inner product of the compared significant feature channel of each sample to be detected and the category center feature value corresponding to the category of each sample to be detected is performed to obtain the credibility index of each sample to be detected.

6. The sample noise identification method according to claim 1, characterized in that: The determining of the credibility index of each sample to be detected based on the feature vector of each sample to be detected and the reference significant feature channel corresponding to the category of each sample to be detected includes: Determining the comparison significant feature channel of each sample to be detected from the feature vector of each sample to be detected; Determining the number of identical feature channels between the compared significant feature channels of each of the samples to be detected and the reference significant feature channels corresponding to the category of each of the samples to be detected; The number of the same characteristic channels is used as a credibility indicator of each sample to be detected.

7. The sample noise identification method according to claim 6, characterized in that: Determining the comparison significant feature channel of each sample to be detected from the feature vector of each sample to be detected includes: Obtaining the amplitude of each feature channel in the feature vector of each sample to be detected; The characteristic channels are sorted by amplitude, and the characteristic channels with the third preset number of bits before sorting are used as the comparison significant characteristic channels of the samples to be detected.

8. The sample noise identification method according to claim 6, characterized in that: Each sample to be detected includes multiple frames of video images to be detected; the feature vector of each sample to be detected includes the feature vector of each frame of video images to be detected in the multiple frames of video images to be detected; Determining the comparison significant feature channel of each sample to be detected from the feature vector of each sample to be detected includes: Performing feature aggregation on a plurality of feature vectors corresponding to the plurality of frames of video images to be detected to obtain a second aggregated feature vector; Based on the feature vector of each frame of the video image to be detected, inter-frame variance calculation is performed to obtain a second variance vector; the second variance vector includes second key channel information; According to the second key channel information, the comparison significant feature channel of each sample to be detected is determined from the second aggregated feature vector.

9. The sample noise identification method according to claim 1, characterized in that: The determining of an erroneous sample from the plurality of samples to be detected based on the credibility index of each sample to be detected comprises: Based on the credibility index of each sample to be detected, determine the average credibility index corresponding to each category in the set of samples to be detected; Normalizing the credibility index of each sample to be detected based on the average credibility index corresponding to each category to obtain the processed credibility index of each sample to be detected; Based on the Gaussian mixture model, the credibility index of each sample to be detected after the processing is fitted to obtain the category deviation index of each sample to be detected; The samples to be detected whose category deviation index is greater than or equal to a preset index among the multiple samples to be detected are determined as erroneous samples.

10. A sample noise identification device, characterized in that: include: A first acquisition module is used to acquire a set of samples to be detected; the set of samples to be detected includes a plurality of samples to be detected and a category set corresponding to the plurality of samples to be detected; the category set includes the category of each sample to be detected in the plurality of samples to be detected; A second acquisition module is used to acquire a reference significant feature channel corresponding to each category in the category set; The reference significant feature channel corresponding to each category is determined based on the feature vector of the clean sample corresponding to each category; the second acquisition module is also used to obtain multiple clean samples corresponding to each category in the category set; determine the feature vector of each clean sample in the multiple clean samples; obtain the amplitude of each feature channel in the feature vector of each clean sample; sort the amplitude of each feature channel, and use the feature channel with the first preset number of bits before sorting as the significant feature channel of each clean sample; Based on the salient feature channels of each clean sample, determine the reference salient feature channels corresponding to each category; Each clean sample includes multiple frames of video images; A first determination module, used to determine the feature vector of each sample to be detected; A second determination module, configured to determine the credibility index of each sample to be detected based on the feature vector of each sample to be detected and the reference significant feature channel corresponding to the category of each sample to be detected; The third determination module is used to determine an erroneous sample from the multiple samples to be detected based on the credibility index of each sample to be detected.

11. The device according to claim 10, characterized in that The feature vector of each clean sample includes the feature vector of each frame of video images in the multiple frames of video images; The second acquisition module is also used to perform feature aggregation on multiple feature vectors corresponding to multiple frames of video images to obtain a first aggregated feature vector; perform inter-frame variance calculation based on the feature vector of each frame of video image to obtain a first variance vector; the first variance vector includes first key channel information; determine the feature channel corresponding to the first key channel information from the first aggregated feature vector, and use the feature channel corresponding to the first key channel information as a significant feature channel of the multiple frames of video images.

12. The device according to claim 10 or 11, characterized in that The second acquisition module is further used to sort the multiple significant feature channels corresponding to the multiple clean samples according to the number of occurrences, and use the significant feature channels with a second preset number of bits before sorting as reference significant feature channels corresponding to each category.

13. The device according to claim 10, characterized in that The second determination module is also used to determine the channel identification of the reference significant feature channel corresponding to the category of each sample to be detected; according to the channel identification, determine the corresponding feature channel from the feature vector of each sample to be detected; use the corresponding feature channel as the comparison significant feature channel of each sample to be detected; and determine the credibility index of each sample to be detected according to the comparison significant feature channel of each sample to be detected.

14. The device according to claim 13, characterized in that The second determination module is also used to determine the number of samples to be detected corresponding to each category; determine the category center feature value corresponding to each category based on the compared significant feature channels of each sample to be detected and the number of samples to be detected corresponding to each category; perform an inner product between the compared significant feature channels of each sample to be detected and the category center feature value corresponding to the category of each sample to be detected to obtain a credibility index of each sample to be detected.

15. The device according to claim 10, characterized in that The second determination module is also used to determine the comparison significant feature channels of each sample to be detected from the feature vector of each sample to be detected; determine the number of identical feature channels between the comparison significant feature channels of each sample to be detected and the reference significant feature channels corresponding to the category of each sample to be detected; and use the number of identical feature channels as a credibility indicator of each sample to be detected.

16. The device according to claim 15, characterized in that The second determination module is also used to obtain the amplitude of each feature channel in the feature vector of each sample to be detected; sort the amplitude of each feature channel, and use the feature channel with the third preset number of digits before sorting as the comparison significant feature channel of each sample to be detected.

17. The device according to claim 15, characterized in that Each sample to be detected includes multiple frames of video images to be detected; the feature vector of each sample to be detected includes the feature vector of each frame of video images to be detected in the multiple frames of video images to be detected; The second determination module is also used to perform feature aggregation on multiple feature vectors corresponding to multiple frames of video images to be detected to obtain a second aggregated feature vector; perform inter-frame variance calculation based on the feature vector of each frame of video images to be detected to obtain a second variance vector; the second variance vector includes second key channel information; based on the second key channel information, determine the comparison significant feature channel of each sample to be detected from the second aggregated feature vector.

18. The device according to claim 10, characterized in that The third determination module is also used to determine the average credibility index corresponding to each category in the set of samples to be detected based on the credibility index of each sample to be detected; normalize the credibility index of each sample to be detected based on the average credibility index corresponding to each category to obtain the credibility index of each sample to be detected after processing; fit a Gaussian mixture model based on the credibility index of each sample to be detected after processing to obtain the category deviation index of each sample to be detected; determine the sample to be detected whose category deviation index is greater than or equal to the preset index among multiple samples to be detected as an erroneous sample.

19. An electronic device, characterized in that: The electronic device includes a processor and a memory, wherein the memory stores at least one instruction or at least one program, and the at least one instruction or the at least one program is loaded and executed by the processor to implement the sample noise identification method according to any one of claims 1 to 9.

20. A computer storage medium, characterized in that The storage medium stores at least one instruction or at least one program, and the at least one instruction or at least one program is loaded and executed by the processor to implement the sample noise identification method according to any one of claims 1 to 9.

21. A computer program product, characterized in that The computer program product includes at least one instruction or at least one program, and the at least one instruction or the at least one program is loaded and executed by a processor to implement the sample noise identification method according to any one of claims 1 to 9.

Images