Data screening method, data screening device, electronic equipment and storage medium
By processing the label sample data in model training with mask network and residual network, combining noise and pure label feature screening, the problem of label sample imbalance is solved, and the robustness and calculation accuracy of the model are improved.
Patent Information
- Application Number
- CN202210510328.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-11
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2042-05-11
AI Technical Summary
During the model training process, the random sampling method leads to uneven label sample data, affecting the robustness of the model.
By obtaining the original tag data, using a preset mask network to filter the pure image samples and noisy image samples, obtain the initial image samples, and extract the features through the residual network, combine noise and pure tag features for classification and calculation, and introduce linearity Filter and in-depth supervision to filter out target image samples.
The balance and calculation accuracy of label sample data are improved, and the target image samples with high quality are obtained, which enhances the robustness of the supervised learning model.
Smart Images

Figure CN114861799B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of artificial intelligence technology, and in particular to a data screening method, a data screening device, an electronic device and a storage medium. Background Art
[0002] At present, in the common model training process, random sampling is often used to select different types of labeled sample data for model training. This method often has the problem of imbalanced input labeled sample data, which affects the robustness of the model. Therefore, how to improve the balance of input labeled sample data has become a technical problem that needs to be solved urgently. Summary of the invention
[0003] The main purpose of the embodiments of the present application is to propose a data screening method, a data screening device, an electronic device and a storage medium, aiming to improve the balance of label sample data, thereby improving the robustness of the model.
[0004] To achieve the above object, a first aspect of an embodiment of the present application proposes a data screening method, the method comprising:
[0005] Acquire original label data, wherein the original label data includes clean image samples labeled with original clean labels and noisy image samples labeled with original noise labels;
[0006] Performing sample filtering processing on the clean image samples and the noisy image samples through a preset mask network to obtain initial image samples, wherein the initial image samples include samples screened from the clean image samples and samples screened from the noisy image samples, and the initial image samples are annotated with an initial clean label and an initial noise label, wherein the initial clean label is derived from the original clean label, and the initial noise label is derived from the original noise label;
[0007] Extracting features from the initial image sample using a preset residual network to obtain an initial sample feature map;
[0008] Performing feature mapping processing on the initial sample feature map according to the initial noise label to obtain a noise label feature, and performing classification score calculation on the noise label feature to obtain a noise label classification value;
[0009] The initial sample feature map is subjected to feature mapping processing according to the initial pure label to obtain a pure label feature, and a classification score is calculated for the pure label feature to obtain a pure label classification value; the initial image sample is subjected to sample screening processing according to the noise label classification value and the pure label classification value to obtain a target image sample, wherein the target image sample is annotated with a target pure label and a target noise label, wherein the target pure label is derived from the initial pure label, and the target noise label is derived from the initial noise label.
[0010] In some embodiments, the step of performing sample filtering processing on the clean image samples and the noisy image samples through a preset mask network to obtain initial image samples includes:
[0011] Get the preset mask sequence;
[0012] Randomly sorting the clean image samples and the noisy image samples to obtain an original image sample sequence;
[0013] Constructing a mapping relationship between the original image sample sequence and the mask sequence;
[0014] According to the mapping relationship, sample filtering is performed on the clean image samples and the noise image samples to obtain the initial image samples.
[0015] In some embodiments, the noise label feature includes a first noise feature and a second noise feature, and the step of performing feature mapping processing on the initial sample feature map according to the initial clean label to obtain a clean label feature, and performing classification score calculation on the clean label feature to obtain a clean label classification value includes:
[0016] According to the initial noise label, the initial sample feature map is input into a preset first mapping network, wherein the first mapping network includes a first multilayer perceptron and a second multilayer perceptron;
[0017] Performing mapping processing on the initial sample feature map through the first multilayer perceptron to obtain a first noise feature;
[0018] Calculating a classification score for the first noise feature using a preset first function to obtain a first classification value;
[0019] Mapping the first noise feature through the second multilayer perceptron to obtain a second noise feature;
[0020] Calculating the classification score of the second noise feature by using a preset second function to obtain a second classification value;
[0021] A weighted calculation is performed on the first classification value and the second classification value to obtain the noise label classification value.
[0022] In some embodiments, the clean label feature includes a first clean feature and a second clean feature, and the step of performing feature mapping processing on the initial sample feature map according to the initial clean label to obtain the clean label feature, and performing classification score calculation on the clean label feature to obtain the clean label classification value includes:
[0023] According to the initial clean label, inputting the initial sample feature map into a preset second mapping network, wherein the second mapping network includes a third multilayer perceptron and a fourth multilayer perceptron;
[0024] Mapping the initial sample feature map through the third multi-layer perceptron to obtain a first pure feature;
[0025] Calculate the classification score of the first pure feature by using the first function to obtain a third classification value;
[0026] Mapping the first pure feature through the fourth multi-layer perceptron to obtain a second pure feature, and fusing the second pure feature with the second noise feature to obtain a fused label feature;
[0027] Calculate the classification score of the fusion label feature by using the second function to obtain a fourth classification value;
[0028] The third classification value and the fourth classification value are weightedly calculated to obtain the clean label classification value.
[0029] In some embodiments, the step of performing sample screening processing on the initial image sample according to the noise label classification value and the clean label classification value to obtain the target image sample includes:
[0030] Comparing the noise label classification value with a preset first threshold, and selecting an initial image sample marked with an initial noise label whose noise label classification value is greater than or equal to the first threshold as the target image sample;
[0031] The clean label classification value is compared with a preset second threshold, and an initial image sample marked with an initial clean label whose clean label classification value is greater than or equal to the second threshold is selected as the target image sample.
[0032] In some embodiments, the step of obtaining the original label data includes:
[0033] Get the preset original label data source;
[0034] The original label data is crawled by a web crawler to obtain the original label data.
[0035] In some embodiments, before the step of obtaining the original label data, the method further includes:
[0036] According to a preset label ratio, extract reference clean label data from the reference clean label data set, and extract reference noisy label data from the reference noisy label data set;
[0037] An original label data source is constructed according to the reference noisy label data and the reference clean label data.
[0038] To achieve the above-mentioned purpose, a second aspect of an embodiment of the present application provides a data screening device, the device comprising:
[0039] A data acquisition module, used to acquire original label data, wherein the original label data includes clean image samples labeled with original clean labels and noisy image samples labeled with original noise labels;
[0040] a sample filtering module, configured to perform sample filtering processing on the clean image samples and the noisy image samples through a preset mask network to obtain initial image samples, wherein the initial image samples include samples screened out from the clean image samples and samples screened out from the noisy image samples, and the initial image samples are annotated with an initial clean label and an initial noisy label, wherein the initial clean label is derived from the original clean label, and the initial noisy label is derived from the original noisy label;
[0041] A feature extraction module, used to extract features from the initial image sample through a preset residual network to obtain an initial sample feature map;
[0042] A first calculation module is used to perform feature mapping processing on the initial sample feature map according to the initial noise label to obtain a noise label feature, and perform classification score calculation on the noise label feature to obtain a noise label classification value;
[0043] A second calculation module is used to perform feature mapping processing on the initial sample feature map according to the initial pure label to obtain a pure label feature, and perform classification score calculation on the pure label feature to obtain a pure label classification value;
[0044] A sample screening module is used to perform sample screening processing on the initial image sample according to the noise label classification value and the clean label classification value to obtain a target image sample, wherein the target image sample is annotated with a target clean label and a target noise label, wherein the target clean label is derived from the initial clean label, and the target noise label is derived from the initial noise label.
[0045] To achieve the above-mentioned purpose, the third aspect of an embodiment of the present application proposes an electronic device, which includes a memory, a processor, a program stored in the memory and executable on the processor, and a data bus for realizing connection and communication between the processor and the memory, and the program, when executed by the processor, implements the method described in the first aspect above.
[0046] To achieve the above-mentioned purpose, the fourth aspect of an embodiment of the present application proposes a storage medium, which is a computer-readable storage medium used for computer-readable storage, and the storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the method described in the first aspect above.
[0047] The data screening method, data screening device, electronic device and storage medium proposed in the present application obtain original label data, wherein the original label data includes clean image samples labeled with original clean labels and noisy image samples labeled with original noise labels. The clean image samples and the noisy image samples are subjected to sample filtering processing through a preset mask network to obtain initial image samples, the initial image samples include samples screened from the clean image samples and samples screened from the noisy image samples, the initial image samples are labeled with initial clean labels and initial noise labels, the initial clean labels are derived from the original clean labels, and the initial noise labels are derived from the original noise labels. A linear screening mechanism is introduced in the data sampling process, which can improve the accuracy and efficiency of data filtering. Furthermore, the initial image sample is feature extracted through the preset residual network to obtain the initial sample feature map, the initial sample feature map is feature mapped according to the initial noise label to obtain the noise label feature, and the noise label feature is classified and scored to obtain the noise label classification value; the initial sample feature map is feature mapped according to the initial clean label to obtain the clean label feature, and the clean label feature is classified and scored to obtain the clean label classification value. Deep supervision is introduced in the classification calculation process, which can extract the label features of the initial sample feature map at different mapping levels, and the classification calculation is performed based on the label features at different mapping levels, which can effectively improve the calculation accuracy. Finally, the initial image sample is sampled and processed according to the noise label classification value and the clean label classification value to obtain the target image sample. The target image sample is annotated with a target clean label and a target noise label. The target clean label comes from the initial clean label, and the target noise label comes from the initial noise label. This method can obtain a target image sample with higher quality, improve the balance of the label sample data of the input model, and thus use the target image sample to train the supervised learning model, which can improve the robustness of the supervised learning model. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] Figure 1 is a flow chart of a data screening method provided in an embodiment of the present application;
[0049] Figure 2 is another flow chart of the data screening method provided in an embodiment of the present application;
[0050] Figure 3 yes Figure 1 Flow chart of step S101 in FIG.
[0051] Figure 4 yes Figure 1 Flow chart of step S102 in FIG.
[0052] Figure 5 yes Figure 1 Flow chart of step S104 in FIG.
[0053] Figure 6 yes Figure 1 Flow chart of step S105 in FIG.
[0054] Figure 7 yes Figure 1 Flow chart of step S106 in FIG.
[0055] Figure 8 is a structural schematic diagram of a data screening device provided in an embodiment of the present application;
[0056] Fig. 9 It is a schematic diagram of the hardware structure of the electronic device provided in the embodiment of the present application. DETAILED DESCRIPTION
[0057] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.
[0058] It should be noted that, although the functional modules are divided in the device schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first", "second", etc. in the specification, claims and the above drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.
[0059] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.
[0060] First, some nouns involved in this application are analyzed:
[0061] Artificial Intelligence (AI): It is a new technical science that studies and develops theories, methods, technologies and application systems for simulating, extending and expanding human intelligence. AI is a branch of computer science. AI attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Research in this field includes robots, language recognition, image recognition, natural language processing and expert systems. AI can simulate the information process of human consciousness and thinking. AI is also a theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.
[0062] Natural Language Processing (NLP): NLP uses computers to process, understand, and apply human languages (such as Chinese, English, etc.). NLP is a branch of artificial intelligence and an interdisciplinary subject between computer science and linguistics. It is often referred to as computational linguistics. Natural language processing includes grammatical analysis, semantic analysis, and text understanding. Natural language processing is commonly used in technical fields such as machine translation, handwritten and printed character recognition, speech recognition and text-to-speech conversion, information intent recognition, information extraction and filtering, text classification and clustering, public opinion analysis, and opinion mining. It involves data mining related to language processing, machine learning, knowledge acquisition, knowledge engineering, artificial intelligence research, and linguistic research related to language computing.
[0063] Information Extraction: A text processing technology that extracts specified types of entity, relationship, event and other factual information from natural language text and forms structured data output. Information extraction is a technology that extracts specific information from text data. Text data is composed of some specific units, such as sentences, paragraphs, and chapters. Text information is composed of some small specific units, such as characters, words, phrases, sentences, paragraphs, or a combination of these specific units. Extracting noun phrases, names, place names, etc. from text data are all text information extraction. Of course, the information extracted by text information extraction technology can be various types of information.
[0064] Supervised learning: The process of adjusting the parameters of a classifier to achieve the required performance using a set of samples of known categories, also known as supervised training or learning with a teacher. Supervised learning is the machine learning task of inferring a function from labeled training data. The training data consists of a set of training examples. In supervised learning, each instance consists of an input object (usually a vector) and a desired output value (also called a supervisory signal). The supervised learning algorithm analyzes the training data and produces an inferred function that can be used to map new instances.
[0065] Weakly supervised learning: Weakly supervised learning is a branch of machine learning. Compared with traditional supervised learning, it uses limited, noisy or inaccurately labeled data to train model parameters.
[0066] Multilayer Perceptron (MLP): A multilayer perceptron is a feed-forward artificial neural network model that maps multiple input data sets to a single output data set.
[0067] At present, in the common model training process, random sampling is often used to select different types of labeled data for model training. This method often has the problem of unbalanced input label data, which affects the robustness of the model. Therefore, how to improve the balance of input label data has become a technical problem that needs to be solved urgently.
[0068] Based on this, the embodiments of the present application provide a data screening method, a data screening device, an electronic device and a storage medium, aiming to improve the balance of label data and thus improve the robustness of the model.
[0069] The data screening method, data screening device, electronic device and storage medium provided in the embodiments of the present application are specifically described through the following embodiments. First, the data screening method in the embodiments of the present application is described.
[0070] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.
[0071] AI basic technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, mechatronics, etc. AI software technologies mainly include computer vision technology, robotics technology, biometrics technology, speech processing technology, natural language processing technology, and machine learning / deep learning.
[0072] The data screening method provided in the embodiment of the present application relates to the field of artificial intelligence technology. The data screening method provided in the embodiment of the present application can be applied to a terminal, can be applied to a server side, or can be software running in a terminal or a server side. In some embodiments, the terminal can be a smart phone, a tablet computer, a laptop computer, a desktop computer, etc.; the server side can be configured as 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, and big data and artificial intelligence platforms; the software can be an application that implements the data screening method, etc., but is not limited to the above forms.
[0073] The present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present application can also be practiced in distributed computing environments, in which tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media including storage devices.
[0074] Figure 1 is an optional flow chart of the data screening method provided in the embodiment of the present application. Figure 1 The method may include but is not limited to steps S101 to S106.
[0075] Step S101, obtaining original label data, wherein the original label data includes clean image samples labeled with original clean labels and noisy image samples labeled with original noise labels;
[0076] Step S102, performing sample filtering processing on the clean image samples and the noisy image samples through a preset mask network to obtain initial image samples, wherein the initial image samples include samples screened out from the clean image samples and samples screened out from the noisy image samples, and the initial image samples are annotated with an initial clean label and an initial noisy label, wherein the initial clean label is derived from the original clean label, and the initial noisy label is derived from the original noisy label;
[0077] Step S103, extracting features from the initial image sample using a preset residual network to obtain an initial sample feature map;
[0078] Step S104, performing feature mapping processing on the initial sample feature map according to the initial noise label to obtain noise label features, and performing classification score calculation on the noise label features to obtain noise label classification values;
[0079] Step S105, performing feature mapping processing on the initial sample feature map according to the initial pure label to obtain a pure label feature, and performing classification score calculation on the pure label feature to obtain a pure label classification value;
[0080] Step S106, performing sample screening processing on the initial image samples according to the noise label classification value and the clean label classification value to obtain target image samples, wherein the target image samples are annotated with a target clean label and a target noise label, wherein the target clean label is derived from the initial clean label, and the target noise label is derived from the initial noise label.
[0081] In the steps S101 to S106 shown in the embodiment of the present application, the clean image samples and the noise image samples are subjected to sample filtering processing through a preset mask network to obtain an initial image sample, the initial image sample includes a sample screened from the clean image sample and a sample screened from the noise image sample, the initial image sample is annotated with an initial clean label and an initial noise label, the initial clean label is derived from the original clean label, and the initial noise label is derived from the original noise label. A linear screening mechanism is introduced in the data sampling process, which can improve the accuracy and filtering efficiency of data filtering. The initial image sample is subjected to feature extraction through a preset residual network to obtain an initial sample feature map, the initial sample feature map is subjected to feature mapping processing according to the initial noise label to obtain a noise label feature, and the noise label feature is subjected to classification score calculation to obtain a noise label classification value; the initial sample feature map is subjected to feature mapping processing according to the initial clean label to obtain a clean label feature, and the clean label feature is subjected to classification score calculation to obtain a clean label classification value. Deep supervision is introduced in the classification calculation process, which can extract the label features of the initial sample feature map at different mapping levels, and the classification calculation is performed based on the label features at different mapping levels, which can effectively improve the calculation accuracy. Finally, the initial image samples are screened according to the noise label classification values and the pure label classification values to obtain the target image samples. This can obtain high-quality target image samples and improve the balance of the label sample data of the input model. The target image samples can then be used to train the supervised learning model, which can improve the robustness of the supervised learning model.
[0082] It should be explained that in actual applications, data labeling often requires a lot of manpower, material resources and financial resources. The quality of labeled data will be affected by human subjective factors to some extent, resulting in the fact that the labeled samples actually obtained often contain a variable proportion of label noise. These labeled data containing label deviations are called noisy label data. Noisy label data include noisy image samples and noise labels. Noise labels can include noise types such as environment, temperature, and pressure. Correspondingly, label data that does not contain label deviations is called pure label data.
[0083] See also Figure 2 In some embodiments, before step S101, the data screening method may include but is not limited to steps S201 to S202:
[0084] Step S201, extracting reference clean label data from a reference clean label data set and extracting reference noise label data from a reference noise label data set according to a preset label ratio;
[0085] Step S202: construct an original label data source according to the reference noisy label data and the reference clean label data.
[0086] In step S201 of some embodiments, the label ratio of the reference pure label data and the reference noise label data can be 1:9 or 1:4. It should be noted that the preset label ratio can be set according to actual business needs. Taking the label ratio of 1:4 and the total label data M of 20 as an example, 4 reference pure label data are extracted from the reference pure label data set, and 16 reference noise label data are extracted from the reference noise label data set. The reference pure label data includes pure image samples with reference pure labels, and the reference noise label data includes noise image samples with reference noise labels.
[0087] In step S202 of some embodiments, the reference noisy label data and the reference clean label data are grouped into one set, thereby constructing an original label data source.
[0088] See also Figure 3 In some embodiments, step S101 may include but is not limited to steps S301 to S302:
[0089] Step S301, obtaining a preset original tag data source;
[0090] Step S302: crawling the original label data source through a web crawler to obtain the original label data.
[0091] In step S301 of some embodiments, the preset original label data source may be a pre-built local database, which can be directly called from the local end; the preset original label data source may also be a certain network platform, and the original label data source can be obtained by jumping to this network interface by inputting a network link.
[0092] In step S302 of some embodiments, the original label data source is crawled by means of a web crawler, and 2N image samples with different labels are extracted according to a preset label ratio to obtain original label data, wherein the original label data includes clean image samples labeled with original clean labels and noisy image samples labeled with original noise labels. For example, 2N image samples with different labels and corresponding original clean labels or original noise labels are selected from the original label data source in a ratio of 1:9 or 1:4. The original clean label data set can be recorded as Dc, and the original clean label data set includes clean image samples labeled with original clean labels; the original noise label data set can be recorded as Dn, and the original noise label data set includes noise image samples labeled with original noise labels.
[0093] See also Figure 4 In some embodiments, step S102 may include but is not limited to steps S401 to S404:
[0094] Step S401, obtaining a preset mask sequence;
[0095] Step S402, randomly sorting the clean image samples and the noisy image samples to obtain an original image sample sequence;
[0096] Step S403, constructing a mapping relationship between the original image sample sequence and the mask sequence;
[0097] Step S404: performing sample filtering on the pure image samples and the noise image samples according to the mapping relationship to obtain initial image samples.
[0098] In step S401 of some embodiments, the mask sequence preset in the mask network includes a first mask and a second mask, wherein the first mask may take a value of 0 and the second mask may take a value of 1. The first mask and the second mask are alternately arranged to obtain a mask sequence, and when the first mask takes a value of 0 and the second mask takes a value of 1, the obtained mask sequence may be expressed as {0, 1, 0, 1, 0, 1, ..., 0, 1, 0, 1, 0, 1}.
[0099] In step S402 of some embodiments, by randomly sorting the clean image samples and the noisy image samples, the order of 2N image samples and their corresponding original labels (i.e., the clean image samples labeled with the original clean labels and the noisy image samples labeled with the original noise labels) can be disrupted to obtain an original label data sequence. For example, N is 5, and the label ratio is 1:4, then the original label data includes 2 clean image samples labeled with the original clean labels and 8 noisy image samples labeled with the original noise labels. After random sorting, the original image sample sequence can be expressed as {Dn2, Dn7, Dn4, Dc1, Dn8, Dn6, Dn5, Dc2, Dn1, Dn3}.
[0100] In step S403 of some embodiments, each original label data in the original image sample sequence is matched with the mask in the mask sequence one by one, and a mapping relationship between the original image sample sequence and the mask sequence is constructed. For example, the original image sample sequence can be represented as {Dn2, Dn7, Dn4, Dc1, Dn8, Dn6, Dn5, Dc2, Dn1, Dn3}, and the mask sequence is {0, 1, 0, 1, 0, 1, 0, 1, 0, 1}, then the mapping relationship can be represented as {Dn2, 0}, {Dn7, 1}, {Dn4, 0}, {Dc1, 1}, {Dn8, 0}, {Dn6, 1}, {Dn5, 0}, {Dc2, 1}, {Dn1, 0}, {Dn3, 1}.
[0101] In step S404 of some embodiments, when performing sample filtering on the pure image samples and the noisy image samples according to the mapping relationship, the image samples corresponding to the first mask (pure image samples or noisy image samples) may be selected as the initial image samples, or the image samples (pure image samples or noisy image samples) may be selected as the initial image samples. For example, the image samples corresponding to the first code 0 are selected as the initial label data, and the initial label data include Dn2, Dn4, Dn8, Dn5, and Dn1, wherein the initial image samples are labeled with an initial pure label and an initial noisy label, the initial pure label is derived from the original pure label, and the initial noise label is derived from the original noise label.
[0102] The above steps S401 to S404 introduce a linear screening mechanism in the data sampling process, which can filter the selected 2N image samples with different labels into N image samples with different labels (including pure image samples with initial pure labels and noisy image samples with initial noise labels), and use these N image samples with different labels as the image sample set Ds, which can improve the accuracy and efficiency of data filtering.
[0103] In step S103 of some embodiments, the residual network includes multiple residual dense blocks, and the residual dense blocks are jump-connected. Each residual dense block includes multiple convolutional layers. The convolutional layers are used to extract features of the initial image samples to capture label features of the initial label data and obtain an initial sample feature map.
[0104] See also Figure 5 In some embodiments, the noise label feature includes a first noise feature and a second noise feature. Step S104 may also include but is not limited to steps S501 to S506:
[0105] Step S501, inputting the initial sample feature map into a preset first mapping network according to the initial noise label, wherein the first mapping network includes a first multilayer perceptron and a second multilayer perceptron;
[0106] Step S502, mapping the initial sample feature map through a first multilayer perceptron to obtain a first noise feature;
[0107] Step S503, calculating the classification score of the first noise feature by using a preset first function to obtain a first classification value;
[0108] Step S504, mapping the first noise feature through a second multilayer perceptron to obtain a second noise feature;
[0109] Step S505, calculating the classification score of the second noise feature by using a preset second function to obtain a second classification value;
[0110] Step S506: Perform weighted calculation on the first classification value and the second classification value to obtain a noise label classification value.
[0111] In step S501 of some embodiments, according to the different labels of image samples in the image sample set Ds, initial image samples with initial noise labels are selected and input into the first mapping network, wherein the first mapping network includes a first multilayer perceptron and a second multilayer perceptron. The initial image samples with initial noise labels can be mapped multiple times through the first mapping network to obtain noise label features corresponding to the initial image samples.
[0112] In step S502 of some embodiments, mapping processing is performed on the initial image samples with the initial noise labels by a first multilayer perceptron, and the initial image samples with the initial noise labels are mapped to a preset first vector space to obtain a first noise feature.
[0113] In step S503 of some embodiments, the first function is a cross entropy function, and the cross entropy function can be used to conveniently calculate the loss of the first noise feature to obtain a first loss value, thereby using the first loss value as the first classification value. Specifically, the cross entropy function can be expressed as shown in formula (1):
[0114]
[0115] Among them, y is a binary label, which can be 0, y i Is the probability value of the output belonging to the y label, y i The first noise feature can be obtained by classifying and calculating it through the sigmoid classification function, where N is the number of samples.
[0116] In step S504 of some embodiments, the first noise feature is mapped to a preset second vector space by a second multilayer perceptron to obtain a second noise feature.
[0117] In step S505 of some embodiments, the second function is a sigmoid classification function. By classifying and calculating the second noise feature, the obtained second classification value can be compressed to between (0, 1). The second classification value is the probability value that the input initial image sample with the initial noise label belongs to the corresponding noise label category.
[0118] In step S506 of some embodiments, the first classification value and the second classification value are weighted according to a preset weight parameter to obtain a final noise label classification value. In some specific embodiments, in order to improve calculation accuracy, the weight parameters of the first classification value and the second classification are both 0.5.
[0119] See also Figure 6 In some embodiments, the clean tag feature includes a first clean feature and a second clean feature, and step S105 further includes but is not limited to steps S601 to S606:
[0120] Step S601, inputting the initial sample feature map into a preset second mapping network according to the initial clean label, wherein the second mapping network includes a third multilayer perceptron and a fourth multilayer perceptron;
[0121] Step S602, mapping the initial sample feature map through a third multi-layer perceptron to obtain a first pure feature;
[0122] Step S603, calculating the classification score of the first pure feature by using the first function to obtain a third classification value;
[0123] Step S604, mapping the first pure feature through a fourth multi-layer perceptron to obtain a second pure feature, and fusing the second pure feature with the second noise feature to obtain a fused label feature;
[0124] Step S605, calculating the classification score of the fused label feature by using the second function to obtain a fourth classification value;
[0125] Step S606: Perform weighted calculation on the third classification value and the fourth classification value to obtain a pure label classification value.
[0126] In step S601 of some embodiments, according to the different labels of image samples in the image sample set Ds, initial image samples with initial pure labels are selected and input into the second mapping network, wherein the second mapping network includes a third multilayer perceptron and a fourth multilayer perceptron. The second mapping network can perform multiple mappings on the initial image samples with initial pure labels to obtain pure label features corresponding to the initial image samples.
[0127] In step S602 of some embodiments, the initial image samples with the initial clean labels are mapped to a preset first vector space by a third multi-layer perceptron to obtain a first clean feature.
[0128] In step S603 of some embodiments, the first function is a cross entropy function, and the cross entropy function can be used to conveniently calculate the loss of the first pure feature to obtain a second loss value, thereby using the second loss value as the second classification value. Specifically, the cross entropy function can be expressed as shown in the above formula (1).
[0129] In step S604 of some embodiments, the first pure feature is mapped to a preset second vector space by a fourth multilayer perceptron to obtain a second pure feature. Further, the second pure feature and the second noise feature are feature fused. Specifically, the second pure feature and the second noise feature are first vectorized, and then the second pure feature and the second noise feature in vector form are vectorized to obtain a fused label feature. In this way, the mapping between the noise label feature space and the pure label feature space can interact, which can improve the accuracy of data screening and the robustness of the model in the subsequent model training process.
[0130] In step S605 of some embodiments, the second function is a sigmoid classification function. By classifying and calculating the second fused label feature, the obtained fourth classification value can be compressed to between (0, 1). The fourth classification value is the probability value that the input initial image sample with the initial pure label belongs to the corresponding pure label category.
[0131] In step S606 of some embodiments, the third classification value and the fourth classification value are weighted according to a preset weight parameter to obtain a final noise label classification value. In some specific embodiments, in order to improve calculation accuracy, the weight parameters of the third classification value and the fourth classification are both 0.5.
[0132] The above steps S501 to S506 and steps S601 to S606 introduce deep supervision during the classification calculation. The same classification value calculation is added to different extraction levels of the initial feature map, and the classification values of different extraction levels are summed to obtain the final result. This can effectively reduce the possibility of errors in the calculated scores. At the same time, by summing the classification values of different extraction levels to obtain the final result, the classification calculation process can also be regularized to avoid the occurrence of data overfitting.
[0133] See also Figure 7 In some embodiments, step S106 may also include but is not limited to steps S701 to S702:
[0134] Step S701, comparing the noise label classification value with a preset first threshold, and selecting an initial image sample marked with an initial noise label whose noise label classification value is greater than or equal to the first threshold as a target image sample;
[0135] Step S702 , comparing the clean label classification value with a preset second threshold, and selecting an initial image sample marked with an initial clean label whose clean label classification value is greater than or equal to the second threshold as a target image sample.
[0136] In step S701 of some embodiments, the first threshold can be set according to actual business needs, and the noise label classification value and the first threshold are compared. The higher the noise label classification value, the better the stability of the corresponding initial image sample annotated with the initial pure label. Therefore, the initial image sample annotated with the initial pure label whose noise label classification value is greater than or equal to the first threshold is selected as the target image sample, wherein the target image sample is annotated with the target noise label.
[0137] In step S702 of some embodiments, the second threshold can be set according to actual business needs. The pure label classification value and the second threshold are compared. The higher the pure label classification value, the better the stability of the corresponding initial image sample annotated with the initial pure label. Therefore, the initial image sample annotated with the initial pure label whose pure label classification value is greater than or equal to the second threshold is selected as the target image sample, wherein the target image sample is annotated with the target pure label.
[0138] The above-mentioned target image samples can be used for the training of supervised learning models, especially the training of weakly supervised learning models. The target image samples can improve the balance between pure input and noisy input in model training, thereby improving the robustness of the supervised learning model and making the supervised learning model compatible in single-label and multi-label application scenarios, thereby improving the applicability of the model.
[0139] The data screening method of the embodiment of the present application obtains original label data, wherein the original label data includes clean image samples labeled with original clean labels and noisy image samples labeled with original noise labels. The clean image samples and the noisy image samples are subjected to sample filtering processing through a preset mask network to obtain initial image samples, the initial image samples include samples screened from the clean image samples and samples screened from the noisy image samples, the initial image samples are labeled with initial clean labels and initial noise labels, the initial clean labels are derived from the original clean labels, and the initial noise labels are derived from the original noise labels. A linear screening mechanism is introduced in the data sampling process, which can improve the accuracy and efficiency of data filtering. Furthermore, the initial image sample is feature extracted through the preset residual network to obtain the initial sample feature map, the initial sample feature map is feature mapped according to the initial noise label to obtain the noise label feature, and the noise label feature is classified and scored to obtain the noise label classification value; the initial sample feature map is feature mapped according to the initial clean label to obtain the clean label feature, and the clean label feature is classified and scored to obtain the clean label classification value. Deep supervision is introduced in the classification calculation process, which can extract the label features of the initial sample feature map at different mapping levels, and the classification calculation is performed based on the label features at different mapping levels, which can effectively improve the calculation accuracy. Finally, the initial image sample is sampled and processed according to the noise label classification value and the clean label classification value to obtain the target image sample. The target image sample is annotated with a target clean label and a target noise label. The target clean label comes from the initial clean label, and the target noise label comes from the initial noise label. This method can obtain a target image sample with higher quality, improve the balance of the label sample data of the input model, and thus use the target image sample to train the supervised learning model, which can improve the robustness of the supervised learning model.
[0140] See also Figure 8 The embodiment of the present application also provides a data screening device, which can implement the above data screening method, and the device includes:
[0141] The data acquisition module 801 is used to acquire original label data, wherein the original label data includes clean image samples labeled with original clean labels and noisy image samples labeled with original noise labels;
[0142] A sample filtering module 802 is used to perform sample filtering processing on the clean image samples and the noisy image samples through a preset mask network to obtain initial image samples, where the initial image samples include samples screened from the clean image samples and samples screened from the noisy image samples, and the initial image samples are annotated with an initial clean label and an initial noisy label, where the initial clean label is derived from the original clean label, and the initial noisy label is derived from the original noisy label;
[0143] The feature extraction module 803 is used to extract features from the initial image sample through a preset residual network to obtain an initial sample feature map;
[0144] The first calculation module 804 is used to perform feature mapping processing on the initial sample feature map according to the initial noise label to obtain the noise label feature, and perform classification score calculation on the noise label feature to obtain the noise label classification value;
[0145] The second calculation module 805 is used to perform feature mapping processing on the initial sample feature map according to the initial pure label to obtain a pure label feature, and perform classification score calculation on the pure label feature to obtain a pure label classification value;
[0146] The sample screening module 806 is used to perform sample screening processing on the initial image samples according to the noise label classification value and the pure label classification value to obtain the target image samples. The target image samples are annotated with a target pure label and a target noise label. The target pure label is derived from the initial pure label, and the target noise label is derived from the initial noise label.
[0147] In some embodiments, the data acquisition module 801 includes:
[0148] A data source acquisition unit, used to acquire a preset original tag data source;
[0149] The data crawling unit is used to crawl the original label data source through a network crawler to obtain the original label data.
[0150] In some embodiments, the sample filtering module 802 includes:
[0151] A mask sequence acquisition unit, used to acquire a preset mask sequence;
[0152] A random sorting unit, used for randomly sorting the pure image samples and the noise image samples to obtain an original image sample sequence;
[0153] A construction unit, used to construct a mapping relationship between an original image sample sequence and a mask sequence;
[0154] The filtering unit is used to perform sample filtering on the pure image samples and the noise image samples according to the mapping relationship to obtain initial image samples.
[0155] In some embodiments, the noise tag feature includes a first noise feature and a second noise feature, and the first calculation module 804 includes:
[0156] A first input unit, used to input the initial sample feature map into a preset first mapping network according to the initial noise label, wherein the first mapping network includes a first multilayer perceptron and a second multilayer perceptron;
[0157] A first mapping unit, configured to perform mapping processing on the initial sample feature map through a first multi-layer perceptron to obtain a first noise feature;
[0158] A first classification value calculation unit, configured to calculate a classification score for the first noise feature using a preset first function to obtain a first classification value;
[0159] A second mapping unit is used to perform mapping processing on the first noise feature through a second multi-layer perceptron to obtain a second noise feature;
[0160] A second classification value calculation unit, used to calculate the classification score of the second noise feature by using a preset second function to obtain a second classification value;
[0161] The first weighting unit is used to perform weighted calculation on the first classification value and the second classification value to obtain a noise label classification value.
[0162] In some embodiments, the clean tag feature includes a first clean feature and a second clean feature, and the second calculation module 805 includes:
[0163] A second input unit is used to input the initial sample feature map into a preset second mapping network according to the initial clean label, wherein the second mapping network includes a third multilayer perceptron and a fourth multilayer perceptron;
[0164] A third mapping unit is used to perform mapping processing on the initial sample feature map through a third multi-layer perceptron to obtain a first pure feature;
[0165] A third classification value calculation unit, configured to calculate a classification score for the first pure feature using the first function to obtain a third classification value;
[0166] a fourth mapping unit, configured to perform mapping processing on the first pure feature through a fourth multi-layer perceptron to obtain a second pure feature, and perform feature fusion on the second pure feature and the second noise feature to obtain a fused label feature;
[0167] A fourth classification value calculation unit, used to calculate the classification score of the fused label feature by using the second function to obtain a fourth classification value;
[0168] The second weighting unit is used to perform weighted calculation on the third classification value and the fourth classification value to obtain a pure label classification value.
[0169] In some embodiments, the sample screening module 806 includes:
[0170] A first comparison unit is used to compare the noise label classification value with a preset first threshold, and select an initial image sample marked with an initial noise label whose noise label classification value is greater than or equal to the first threshold as a target image sample;
[0171] The second comparison unit is used to compare the clean label classification value with a preset second threshold, and select the initial image sample marked with the initial clean label whose clean label classification value is greater than or equal to the second threshold as the target image sample.
[0172] The specific implementation of the data screening device is basically the same as the specific implementation of the above-mentioned data screening method, and will not be repeated here.
[0173] The embodiment of the present application also provides an electronic device, the electronic device comprising: a memory, a processor, a program stored in the memory and executable on the processor, and a data bus for realizing connection and communication between the processor and the memory, and the above-mentioned data screening method is realized when the program is executed by the processor. The electronic device can be any intelligent terminal including a tablet computer, a vehicle-mounted computer, etc.
[0174] See also Fig. 9 , Fig. 9 The hardware structure of an electronic device of another embodiment is illustrated, and the electronic device includes:
[0175] The processor 901 may be implemented by a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (Application Specific Integrated Circuit, ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present application;
[0176] The memory 902 can be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 902 can store an operating system and other application programs. When the technical solution provided in the embodiment of this specification is implemented by software or firmware, the relevant program code is stored in the memory 902, and the processor 901 calls and executes the data screening method of the embodiment of this application;
[0177] Input / output interface 903, used to implement information input and output;
[0178] Communication interface 904, used to realize communication interaction between the device and other devices, which can be realized by wired mode (such as USB, network cable, etc.) or wireless mode (such as mobile network, WIFI, Bluetooth, etc.);
[0179] A bus 905 that transmits information between various components of the device (e.g., the processor 901, the memory 902, the input / output interface 903, and the communication interface 904);
[0180] The processor 901 , the memory 902 , the input / output interface 903 and the communication interface 904 are connected to each other in communication within the device via a bus 905 .
[0181] An embodiment of the present application also provides a storage medium, which is a computer-readable storage medium used for computer-readable storage. The storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the above-mentioned data screening method.
[0182] The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory may optionally include a memory remotely disposed relative to the processor, and these remote memories may be connected to the processor 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.
[0183] The data screening method, data screening device, electronic device and storage medium provided by the embodiments of the present application obtain original label data, wherein the original label data includes clean image samples labeled with original clean labels and noisy image samples labeled with original noise labels. The clean image samples and the noisy image samples are subjected to sample filtering processing through a preset mask network to obtain initial image samples, the initial image samples include samples screened from the clean image samples and samples screened from the noisy image samples, the initial image samples are labeled with initial clean labels and initial noise labels, the initial clean labels are derived from the original clean labels, and the initial noise labels are derived from the original noise labels. A linear screening mechanism is introduced in the data sampling process, which can improve the accuracy and efficiency of data filtering. Furthermore, the initial image sample is feature extracted through the preset residual network to obtain the initial sample feature map, the initial sample feature map is feature mapped according to the initial noise label to obtain the noise label feature, and the noise label feature is classified and scored to obtain the noise label classification value; the initial sample feature map is feature mapped according to the initial clean label to obtain the clean label feature, and the clean label feature is classified and scored to obtain the clean label classification value. Deep supervision is introduced in the classification calculation process, which can extract the label features of the initial sample feature map at different mapping levels, and the classification calculation is performed based on the label features at different mapping levels, which can effectively improve the calculation accuracy. Finally, the initial image sample is sampled and processed according to the noise label classification value and the clean label classification value to obtain the target image sample. The target image sample is annotated with a target clean label and a target noise label. The target clean label comes from the initial clean label, and the target noise label comes from the initial noise label. This method can obtain a target image sample with higher quality, improve the balance of the label sample data of the input model, and thus use the target image sample to train the supervised learning model, which can improve the robustness of the supervised learning model.
[0184] The embodiments described in the embodiments of the present application are intended to more clearly illustrate the technical solutions of the embodiments of the present application and do not constitute a limitation on the technical solutions provided in the embodiments of the present application. Those skilled in the art will appreciate that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.
[0185] It can be understood by those skilled in the art that Figure 1-7 The technical solutions shown in the figure do not constitute a limitation on the embodiments of the present application, and may include more or fewer steps than those shown in the figure, or a combination of certain steps, or different steps.
[0186] The device embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separated, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
[0187] Those skilled in the art will appreciate that all or some of the steps in the methods disclosed above, and the functional modules / units in the systems and devices may be implemented as software, firmware, hardware, or a suitable combination thereof.
[0188] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising 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.
[0189] It should be understood that in the present application, "at least one (item)" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the objects associated before and after are in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.
[0190] In the several embodiments provided in the present application, it should be understood that the disclosed 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 above units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.
[0191] The units described above 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 solution of this embodiment.
[0192] In addition, each functional unit in each embodiment of the present application 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.
[0193] 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 application, 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 storage medium, including multiple instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, referred to as ROM), random access memory (Random Access Memory, referred to as RAM), disk or optical disk and other media that can store programs.
[0194] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but the scope of the rights of the present invention is not limited thereto. Any modification, equivalent substitution and improvement made by a person skilled in the art without departing from the scope and essence of the present invention should be within the scope of the rights of the present invention.
Claims
1. A data screening method, characterized in that: The method comprises: Acquire original label data, wherein the original label data includes clean image samples labeled with original clean labels and noisy image samples labeled with original noise labels; Performing sample filtering processing on the clean image samples and the noisy image samples through a preset mask network to obtain initial image samples, wherein the initial image samples include samples screened from the clean image samples and samples screened from the noisy image samples, and the initial image samples are annotated with an initial clean label and an initial noise label, wherein the initial clean label is derived from the original clean label, and the initial noise label is derived from the original noise label; Extracting features from the initial image sample using a preset residual network to obtain an initial sample feature map; Performing feature mapping processing on the initial sample feature map according to the initial noise label to obtain a noise label feature, and performing classification score calculation on the noise label feature to obtain a noise label classification value; Performing feature mapping processing on the initial sample feature map according to the initial pure label to obtain a pure label feature, and performing classification score calculation on the pure label feature to obtain a pure label classification value; Performing sample screening processing on the initial image sample according to the noise label classification value and the clean label classification value to obtain a target image sample, wherein the target image sample is annotated with a target clean label and a target noise label, wherein the target clean label is derived from the initial clean label, and the target noise label is derived from the initial noise label; The noise label feature includes a first noise feature and a second noise feature; the initial sample feature map is subjected to feature mapping processing according to the initial noise label to obtain a noise label feature, and a classification score is calculated for the noise label feature to obtain a noise label classification value, including: According to the initial noise label, the initial sample feature map is mapped by a first multilayer perceptron of a preset first mapping network to obtain a first noise feature; a classification score is calculated by a preset first function to obtain a first classification value; the first noise feature is mapped by a second multilayer perceptron of the first mapping network to obtain a second noise feature; a classification score is calculated by a preset second function to obtain a second classification value for the second noise feature; and the first classification value and the second classification value are weighted to obtain a classification value for the noise label; The performing sample screening processing on the initial image sample according to the noise label classification value and the clean label classification value to obtain the target image sample includes: An initial image sample marked with an initial noise label and having a noise label classification value greater than or equal to a preset first threshold is selected as the target image sample.
2. The data screening method according to claim 1, characterized in that: The step of performing sample filtering processing on the clean image sample and the noise image sample through a preset mask network to obtain an initial image sample includes: Get the preset mask sequence; Randomly sorting the clean image samples and the noisy image samples to obtain an original image sample sequence; Constructing a mapping relationship between the original image sample sequence and the mask sequence; According to the mapping relationship, sample filtering is performed on the clean image samples and the noise image samples to obtain the initial image samples.
3. The data screening method according to claim 1, characterized in that: The clean label feature includes a first clean feature and a second clean feature. The step of performing feature mapping processing on the initial sample feature map according to the initial clean label to obtain the clean label feature, and performing classification score calculation on the clean label feature to obtain the clean label classification value includes: According to the initial clean label, inputting the initial sample feature map into a preset second mapping network, wherein the second mapping network includes a third multilayer perceptron and a fourth multilayer perceptron; Mapping the initial sample feature map through the third multi-layer perceptron to obtain a first pure feature; Calculate the classification score of the first pure feature by using the first function to obtain a third classification value; Mapping the first pure feature through the fourth multi-layer perceptron to obtain a second pure feature, and fusing the second pure feature with the second noise feature to obtain a fused label feature; Calculate the classification score of the fusion label feature by using the second function to obtain a fourth classification value; The third classification value and the fourth classification value are weightedly calculated to obtain the clean label classification value.
4. The data screening method according to any one of claims 1 to 3, characterized in that: The step of performing sample screening processing on the initial image sample according to the noise label classification value and the clean label classification value to obtain the target image sample comprises: The clean label classification value is compared with a preset second threshold, and an initial image sample marked with an initial clean label whose clean label classification value is greater than or equal to the second threshold is selected as the target image sample.
5. The data screening method according to any one of claims 1 to 3, characterized in that: The step of obtaining the original label data includes: Get the preset original label data source; The original label data is crawled by a web crawler to obtain the original label data.
6. The data screening method according to claim 5, characterized in that: Before the step of obtaining the original label data, the method further includes: According to a preset label ratio, extract reference clean label data from the reference clean label data set, and extract reference noisy label data from the reference noisy label data set; An original label data source is constructed according to the reference noisy label data and the reference clean label data.
7. A data screening device, characterized in that: The device comprises: A data acquisition module, used to acquire original label data, wherein the original label data includes clean image samples labeled with original clean labels and noisy image samples labeled with original noise labels; a sample filtering module, configured to perform sample filtering processing on the clean image samples and the noisy image samples through a preset mask network to obtain initial image samples, wherein the initial image samples include samples screened out from the clean image samples and samples screened out from the noisy image samples, and the initial image samples are annotated with an initial clean label and an initial noisy label, wherein the initial clean label is derived from the original clean label, and the initial noisy label is derived from the original noisy label; A feature extraction module, used to extract features from the initial image sample through a preset residual network to obtain an initial sample feature map; A first calculation module is used to perform feature mapping processing on the initial sample feature map according to the initial noise label to obtain a noise label feature, and perform classification score calculation on the noise label feature to obtain a noise label classification value; A second calculation module is used to perform feature mapping processing on the initial sample feature map according to the initial pure label to obtain a pure label feature, and perform classification score calculation on the pure label feature to obtain a pure label classification value; a sample screening module, configured to perform sample screening processing on the initial image sample according to the noise label classification value and the clean label classification value to obtain a target image sample, wherein the target image sample is annotated with a target clean label and a target noise label, wherein the target clean label is derived from the initial clean label, and the target noise label is derived from the initial noise label; The noise label feature includes a first noise feature and a second noise feature; the feature mapping process is performed on the initial sample feature map according to the initial noise label to obtain the noise label feature, and the classification score is calculated for the noise label feature to obtain the noise label classification value, which includes: According to the initial noise label, the initial sample feature map is mapped by a first multilayer perceptron of a preset first mapping network to obtain a first noise feature; a classification score is calculated for the first noise feature by a preset first function to obtain a first classification value; the first noise feature is mapped by a second multilayer perceptron of the first mapping network to obtain a second noise feature; a classification score is calculated for the second noise feature by a preset second function to obtain a second classification value; a weighted calculation is performed on the first classification value and the second classification value to obtain a classification value of the noise label The performing sample screening processing on the initial image sample according to the noise label classification value and the clean label classification value to obtain the target image sample includes: An initial image sample marked with an initial noise label and having a noise label classification value greater than or equal to a preset first threshold is selected as the target image sample.
8. An electronic device, characterized in that: The electronic device includes a memory, a processor, a program stored in the memory and executable on the processor, and a data bus for realizing connection and communication between the processor and the memory. When the program is executed by the processor, the steps of the data screening method as described in any one of claims 1 to 6 are realized.
9. A storage medium, the storage medium being a computer-readable storage medium, used for computer-readable storage, characterized in that: The storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the steps of the data screening method according to any one of claims 1 to 6.
Citation Information
Patent Citations
Sample data cleaning method, device, computer device, and storage medium
CN109241903A
Beauty prediction method and device based on migration and weak supervision, and storage medium
WO2021258482A1
