Methods, devices and electronic equipment for annotating audio and video data
Through the use of the target generative adversarial network model, the problem of difficulty in labeling audio and video data is solved, and the feature parameters and labeling is automatically extracted, which reduces labor costs and improves efficiency.
Patent Information
- Application Number
- CN202111627068.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-28
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2041-12-28
AI Technical Summary
It is difficult to obtain a large amount of labeled audio and video data, and it is difficult for the prior art to effectively extract the characteristic parameters of audio and video data for classification and annotation.
The target generative adversarial network model is adopted to generate and restore controlled random data through generators and discriminators, and the characteristic parameters of the audio and video data to be labeled are automatically extracted according to the similarity parameters, realizing unsupervised learning and automatic labeling.
It reduces the cost of manual labeling, improves the labeling efficiency, and can realize automatic labeling of audio and video data without manual operation, simplifying the data acquisition process.
Smart Images

Figure CN114501164B_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates to the field of data processing technology, and in particular to a method, device and electronic device for labeling audio and video data. Background Art
[0002] Neural networks can be used to solve many problems. Neural networks are algorithms that simulate the behavioral characteristics of animal neural networks and perform distributed parallel information processing. Neural networks process information by adjusting the connections between a large number of internal nodes. If neural networks are to perform well, they must rely on a large amount of training audio and video data. However, it is usually difficult to obtain a large amount of annotated audio and video data. In the process of obtaining annotated audio and video data, how to extract the feature parameters corresponding to the audio and video data according to the classification criteria and annotate the data corresponding to the feature parameters is an urgent problem to be solved. Summary of the invention
[0003] In order to solve the above technical problems or at least partially solve the above technical problems, the present disclosure provides a method, device and electronic device for labeling audio and video data, which can automatically extract feature parameters corresponding to the audio and video data to be labeled, and label the classification of the audio and video data to be labeled.
[0004] In order to achieve the above objectives, the technical solutions provided by the embodiments of the present disclosure are as follows:
[0005] In a first aspect, a method for labeling audio and video data is provided, comprising:
[0006] Acquire first generated data according to the audio and video data to be annotated and the first controlled random data, where the first controlled random data is generated according to the characteristic distribution of the target audio and video data;
[0007] Determine a first similarity parameter and a second controlled random data corresponding to the first generated data, wherein the first similarity parameter is used to characterize the probability that the first generated data is the target audio and video data, and the second controlled random data is the first controlled random data restored according to the first generated data;
[0008] When the first similarity parameter is greater than or equal to a similarity threshold, the second controlled random data is determined to be a characteristic parameter of the audio and video data to be labeled, and the audio and video data to be labeled is labeled as the target audio and video data.
[0009] As an optional implementation of the embodiment of the present disclosure, the method is applied to a target generative adversarial network model, the target generative adversarial network model includes a generator and a discriminator, and the generator includes an encoder and a decoder;
[0010] The step of obtaining first generated data according to the audio and video data to be annotated and the first controlled random data includes:
[0011] Inputting the audio and video data to be annotated into the encoder to generate the encoded audio and video data to be annotated;
[0012] The encoded audio and video data to be marked and the first controlled random data are input into the decoder to obtain the first generated data.
[0013] As an optional implementation of the embodiment of the present disclosure, the generator is obtained after training according to the first objective function;
[0014] The first objective function is: max[logD fake +a*sum(logD condition )];
[0015] Among them, D fake logD is the probability that the similarity parameter output by the discriminator is the target audio and video data when the input signal of the discriminator is the generated data; condition is the similarity of the controlled random data corresponding to each generated data output by the generator, and a is a constant.
[0016] As an optional implementation of the embodiment of the present disclosure, the method is applied to a target generative adversarial network model, the target generative adversarial network model includes a generator and a discriminator, and the generator includes an encoder and a decoder;
[0017] The controller is specifically used for:
[0018] inputting the first generated data into the discriminator;
[0019] The first similarity parameter and the second controlled random data output by the discriminator are obtained.
[0020] As an optional implementation of the embodiment of the present disclosure, the discriminator is obtained after training according to the second objective function;
[0021] The second objective function is: max[logD real +log(1-D fake )+a*sum(log(1-D condition ))];
[0022] Among them, D real is the probability that the similarity parameter output by the discriminator is the target audio and video data when the input signal of the discriminator is real data; D fakeis the probability that the similarity parameter output by the discriminator is the target audio and video data when the input signal of the discriminator is the generated data; log (1-D condition ) is the similarity degree of the controlled random data corresponding to each generated data output by the generator, and a is a constant.
[0023] As an optional implementation of the embodiment of the present disclosure, before obtaining the first generated data according to the audio and video data to be annotated and the first controlled random data, the method further includes:
[0024] Determining the characteristic distribution of the target audio and video data according to a plurality of sample audio and video data, wherein the plurality of sample audio and video data are the target audio and video data;
[0025] The first controlled random data is determined according to the characteristic distribution of the target audio and video data.
[0026] In a second aspect, a device for labeling audio and video data is provided, characterized in that it includes:
[0027] A generation module, configured to obtain first generated data according to the audio and video data to be annotated and first controlled random data, wherein the first controlled random data is generated according to the characteristic distribution of the target audio and video data;
[0028] a discriminant module, configured to determine a first similarity parameter and a second controlled random data corresponding to the first generated data, wherein the first similarity parameter is used to characterize the probability that the first generated data is the target audio and video data, and the second controlled random data is the first controlled random data restored according to the first generated data;
[0029] The labeling module is used to determine that the second controlled random data is the characteristic parameter of the audio and video data to be labeled when the first similarity parameter is greater than or equal to the similarity threshold, and label the audio and video data to be labeled as the target audio and video data.
[0030] In a third aspect, an electronic device is provided, including: a controller, configured to:
[0031] Acquire first generated data according to the audio and video data to be annotated and the first controlled random data, where the first controlled random data is generated according to the characteristic distribution of the target audio and video data;
[0032] Determine a first similarity parameter and a second controlled random data corresponding to the first generated data, wherein the first similarity parameter is used to characterize the probability that the first generated data is the target audio and video data, and the second controlled random data is the first controlled random data restored according to the first generated data;
[0033] When the first similarity parameter is greater than or equal to a similarity threshold, the second controlled random data is determined to be a characteristic parameter of the audio and video data to be labeled, and the audio and video data to be labeled is labeled as the target audio and video data.
[0034] As an optional implementation of the embodiment of the present disclosure, the controller is built with a target generative adversarial network model, the target generative adversarial network model includes a generator and a discriminator, and the generator includes an encoder and a decoder;
[0035] The controller is specifically used for:
[0036] Inputting the audio and video data to be annotated into the encoder to generate the encoded audio and video data to be annotated;
[0037] The encoded audio and video data to be marked and the first controlled random data are input into the decoder to obtain the first generated data.
[0038] As an optional implementation of the embodiment of the present disclosure, the generator is obtained after training according to the first objective function;
[0039] The first objective function is: max[logD fake +a*sum(logD condition )];
[0040] Among them, D fake logD is the probability that the similarity parameter output by the discriminator is the target audio and video data when the input signal of the discriminator is the generated data; condition is the similarity of the controlled random data corresponding to each generated data output by the generator, and a is a constant.
[0041] As an optional implementation of the embodiment of the present disclosure, the controller is built with a target generative adversarial network model, the target generative adversarial network model includes a generator and a discriminator, and the generator includes an encoder and a decoder;
[0042] The controller is specifically used for:
[0043] inputting the first generated data into the discriminator;
[0044] The first similarity parameter and the second controlled random data output by the discriminator are obtained.
[0045] As an optional implementation of the embodiment of the present disclosure, the discriminator is obtained after training according to the second objective function;
[0046] The second objective function is: max[logD real +log(1-D fake )+a*sum(log(1-D condition ))];
[0047] Among them, D real is the probability that the similarity parameter output by the discriminator is the target audio and video data when the input signal of the discriminator is real data; D fake is the probability that the similarity parameter output by the discriminator is the target audio and video data when the input signal of the discriminator is the generated data; log (1-D condition ) is the similarity degree of the controlled random data corresponding to each generated data output by the generator, and a is a constant.
[0048] As an optional implementation of the embodiment of the present disclosure, the controller is further used for:
[0049] Before obtaining the first generated data according to the audio and video data to be labeled and the first controlled random data, determining the characteristic distribution of the target audio and video data according to a plurality of sample audio and video data, wherein the plurality of sample audio and video data are the target audio and video data;
[0050] The first controlled random data is determined according to the characteristic distribution of the target audio and video data.
[0051] In a fourth aspect, an electronic device is provided, comprising: a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, the method for labeling audio and video data as described in the first aspect or any one of its optional embodiments is implemented.
[0052] In a fifth aspect, a computer-readable storage medium is provided, comprising: a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the method for marking audio and video data as described in the first aspect or any one of its optional embodiments is implemented.
[0053] In a sixth aspect, a computer program product is provided, characterized in that it includes: when the computer program product is run on a computer, the computer implements the audio and video data labeling method as described in the first aspect or any one of its optional embodiments.
[0054] Compared with the prior art, the technical solution provided by the embodiment of the present disclosure has the following advantages: first controlled random data can be determined according to the characteristic distribution of target audio and video data, and the first controlled random data is used as noise to generate first generated data with the audio and video data to be labeled, so that when the generated data is obtained, constraints are imposed by the first controlled random data, and a first similarity parameter corresponding to the first generated data is calculated, as well as second controlled random data generated after audio and video data recovery based on the first controlled random data. Since the first similarity parameter is used to characterize the probability that the first generated data is the target audio and video data, when the first similarity parameter is greater than or equal to the similarity threshold, the first generated data is considered to be the target audio and video data. The second controlled random number recovered in this way can be considered to be the characteristic parameter of the audio and video data to be labeled, and it can be determined that the audio and video data to be labeled is the target audio and video data. BRIEF DESCRIPTION OF THE DRAWINGS
[0055] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure.
[0056] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.
[0057] Figure 1 A schematic diagram of the process of obtaining and applying annotated audio and video data in related technologies;
[0058] Figure 2 A schematic diagram of a process for obtaining and applying annotated audio and video data in an embodiment of the present disclosure;
[0059] Figure 3 Schematic diagram of the architecture of a GAN model in an embodiment of the present disclosure;
[0060] Figure 4 Schematic diagram of the architecture of an info GAN model in an embodiment of the present disclosure;
[0061] Figure 5 A schematic diagram of the architecture of a target generative adversarial network model provided in an embodiment of the present disclosure;
[0062] Fig. 6A is a hardware configuration block diagram of an electronic device 200 according to an embodiment of the present disclosure;
[0063] Figure 6B A software configuration block diagram of an electronic device 200 according to an embodiment of the present disclosure;
[0064] Figure 7 A flowchart of a method for labeling audio and video data in an embodiment of the present disclosure is shown;
[0065] Fig. 8A This is a schematic diagram of a process for obtaining first controlled random data in an embodiment of the present disclosure;
[0066] Figure 8B is a schematic diagram of characteristic distribution of a frequency range of human voice in an embodiment of the present disclosure;
[0067] Fig. 9 A schematic diagram of a flow chart of a method for labeling audio and video data based on a target generative adversarial network model in an embodiment of the present disclosure;
[0068] Fig.10 A schematic diagram of the architecture of a controller in an embodiment of the present disclosure;
[0069] Fig.11 A structural block diagram of a device for labeling audio and video data provided by an embodiment of the present disclosure. DETAILED DESCRIPTION
[0070] In order to more clearly understand the above-mentioned objectives, features and advantages of the present disclosure, the scheme of the present disclosure will be further described below. It should be noted that the embodiments of the present disclosure and the features in the embodiments can be combined with each other without conflict.
[0071] In the following description, many specific details are set forth to facilitate a full understanding of the present disclosure, but the present disclosure may also be implemented in other ways different from those described herein; it is obvious that the embodiments in the specification are only part of the embodiments of the present disclosure, rather than all of the embodiments.
[0072] The terms "first" and "second" in this disclosure are used to distinguish different objects rather than to describe a specific order of objects. For example, the first controlled random data and the second controlled random data are used to distinguish different controlled random data rather than to describe a specific order of controlled random data.
[0073] Neural networks can refer to two types, one is biological neural network, and the other is artificial neural network (ANNs). Biological neural network generally refers to the neural network composed of brain neurons, cells, contacts, etc. of organisms, which is used to generate biological consciousness and help organisms think and act. Artificial neural network is also referred to as neural network (NNs), or artificial neural network is called connection model (Connection Model). Artificial neural network is an algorithmic mathematical model that imitates the behavioral characteristics of animal neural networks and performs distributed parallel information processing. This network relies on the complexity of the system to adjust the interconnected relationship between a large number of internal nodes to achieve the purpose of processing information. The neural networks involved in the following embodiments of the present disclosure are all artificial neural networks.
[0074] At present, artificial neural networks can be used to solve many problems. Artificial neural networks are an algorithm that simulates the behavioral characteristics of animal artificial neural networks and performs distributed parallel information processing. Artificial neural networks adjust the interconnected relationships between a large number of internal nodes to achieve the purpose of processing information. If artificial neural networks want to perform well, they must rely on a large amount of labeled data. A large amount of labeled data is the cornerstone of ensuring the performance of artificial neural networks, so obtaining labeled data has become a top priority for realizing artificial intelligence through artificial neural networks. In order to realize artificial intelligence, it is necessary to teach computers human understanding and judgment. Labeling data is a method that enables computers to recognize human language. Specifically, it is to label audio and video data that need to be recognized and distinguished by computers with feature tags, so that computers can continuously recognize these audio and video data, and finally realize that computers can recognize them autonomously. For display devices, it is usually necessary to distinguish some audio data based on certain features to identify audio data, so as to realize certain functions (such as displaying corresponding subtitle data). For example, when displaying on a display device, it may be necessary to distinguish between human voices in different languages (such as distinguishing between application human voices and Chinese human voices). If there is no labeled data, it is difficult to distinguish between human voices in different languages, because when the similarity between the two language systems is high, some wrong features may be selected for classification, such as classification based on the size of the elderly and children, men and women, and the volume of the speaking voice, because these differences are greater than the difference between the pronunciation of the two languages, and are easier to identify, so they will have a greater probability of becoming classification criteria. It can be seen from the above that the acquisition of labeled data is very important, but it is usually difficult to obtain a large amount of labeled audio and video data. At present, in the process of labeling audio and video data, how to extract the feature parameters corresponding to the audio and video data according to the classification criteria is an urgent problem to be solved.
[0075] For example, Figure 1As shown, it is a schematic diagram of the process of obtaining and applying annotated audio and video data in the related technology. After obtaining the audio and video data to be annotated, the audio and video data to be annotated is first annotated by manual annotation to obtain the annotated data, and then the annotated data is used as a training sample and input into the artificial neural network to train the artificial neural network.
[0076] Since the above-mentioned annotated data (annotated audio and video data) is obtained by manually annotating the audio and video data, a large amount of labor costs are required. In addition, when a large amount of manually annotated audio and video data needs to be obtained, it takes a long time to manually annotate the annotated audio and video data. In order to solve the above problems, the embodiments of the present disclosure provide an audio and video data annotation method, device and electronic device. For example, Figure 2 As shown, it is a schematic diagram of a process of obtaining and applying annotated audio and video data in an embodiment of the present disclosure. In the embodiment of the present disclosure, the target generative adversarial network can automatically extract the feature parameters corresponding to the audio and video data to be annotated, and determine that the audio and video data to be annotated is annotated as target audio and video data, obtain the annotated data, and use the annotated data as a training sample for the artificial neural network. The audio and video data involved in the embodiment of the present disclosure can be audio data and / or video data.
[0077] In the disclosed embodiment, the first generated data can be obtained by inputting the first controlled random data, and when the generated data is judged as real data, the second controlled anytime data that is used to restore the first controlled anytime data through the first generated data can be used as the characteristic parameter of the audio and video data to be annotated, and the audio and video data to be annotated is determined to be the same type of data as the target audio and video data represented by the first controlled random data. In this way, it is possible to obtain the annotated audio and video data without manual operation, which greatly reduces the labor cost and improves the annotation efficiency.
[0078] In the disclosed embodiment, unsupervised learning can be used to automatically extract feature parameters corresponding to audio and video data, and determine the classification of the audio and video data to be labeled. Unsupervised learning is an important branch of current artificial neural networks. Unsupervised learning essentially uses the features of audio and video data to cluster audio and video data according to certain rules. For audio data and video data, the biggest difficulty in classification is how to extract the corresponding features according to the classification criteria, because the classification of audio data and video data is often not directly based on the features of the information itself, but involves human understanding, such as the classification of apples and bananas in audio data and video data. Especially in complex backgrounds, it becomes more difficult to classify audio data and video data.
[0079] In the disclosed embodiment, based on the mutual information generative adversarial network (infoGAN) model in unsupervised learning, an improved target generative adversarial network model is obtained, through which the feature parameters corresponding to the audio and video data are automatically extracted. Among them, the infoGAN model is a model based on the evolution of the Generative Adversarial Networks (GAN) model.
[0080] The mathematical theory of the GAN model points out that the generator and the discriminator learn adversarially. Specifically, the generator generates fake samples that satisfy the positive sample distribution as much as possible, while the discriminator distinguishes true and false samples as much as possible. In this game process, the performance of both networks is getting better and better, thus achieving adversarial learning. The goal of the GAN model is actually to obtain a generated data (fake data) distribution that is consistent with the distribution of real data. However, since the input of the generator is a continuous random noise z and there is no constraint, the interpretability of GAN when it matches the specific dimension of z (which can be understood as incompressible noise) with the semantic features of the output is very poor. In other words, from the perspective of representation learning, the GAN model will find that since there is no restriction when the generator uses random noise z, it uses z in a highly mixed way, and no dimension of z clearly represents a feature. Therefore, in the process of data generation, we cannot know what kind of noise z can be used to generate what kind of results, which limits our use of GAN to some extent.
[0081] For example, Figure 3 As shown, it is a schematic diagram of the architecture of a GAN model. The GAN model includes a generator and a discriminator. After the generator inputs a random noise z, it will obtain generated data. After the generated data and the real data are input into the discriminator, the discriminator determines whether the generated data is real data and outputs a judgment result. When the judgment result is true, it indicates that the generated data is real data. When the judgment result is false, it indicates that the generated data is not real data.
[0082] In the infoGAN model, the input vector z is divided into two parts, random variable c and random noise z'. The random variable c can be understood as an interpretable hidden variable, and the random noise z' can be understood as incompressible noise. The infoGAN model hopes to constrain the relationship between c and the output signal so that the dimension of c corresponds to the semantic features of the output signal. Taking handwritten numbers as an example, these semantic features can include: the thickness of the strokes of the numbers, the inclination of the numbers, etc. The infoGAN model constrains c through mutual information, which can also be understood as a self-encoding process. The specific operation is that the discriminator also includes a recovery network Q. The output signal of the generator will pass through the recovery network Q to see if c can be recovered. The recovery process of the recovery network Q can actually be regarded as the inverse process of an autoencoder. The rest of the discriminator is the same as the conventional GAN model.
[0083] For example, Figure 4 As shown in the figure, it is a schematic diagram of the architecture of an infoGAN model. The infoGAN model includes a generator and a discriminator. The discriminator includes a generator network Q and a discriminator network D. After the random variable c and random noise z' are input into the generator, the generator will output a generated data. The generated data and the real data are input into the discriminator. The recovery network Q restores the random variable c according to the generated data, and the discriminator network D determines whether the generated data is real data. Among them, in the actual implementation process, the recovery network Q and the discriminator network D will share parameters. Only the last layer of the two is different. The recovery network Q is a vector, and the discriminator network D outputs a scalar.
[0084] like Figure 5 , which is a schematic diagram of the architecture of the target generative adversarial network model provided in an embodiment of the present disclosure. The architecture of the discriminator in the target generative adversarial network model is similar to Figure 4 The architecture diagram of the infoGAN model shown in FIG. 1 is similar, but the architecture of the generator is different. The generator includes an encoder and a decoder, and the input signal of the target generative adversarial network model in the embodiment of the present disclosure is the same as that of the above-mentioned Figure 4 The input signal of the infoGAN model in is also different. Figure 5In the process, the audio and video data to be labeled is input into the encoder to obtain the encoded audio and video data to be labeled, and then the encoded audio and video data and the first controlled random data are input into the decoder to obtain the first generated data, and then the discriminator compares the first generated data with the real data, and outputs the first similarity parameter to judge whether the first generated data is true or false, and the discriminator will also restore the first controlled random data through the first generated data to obtain the second controlled random data. When the first similarity parameter is greater than or equal to the similarity threshold, the second controlled random data is determined to be the feature parameter of the audio and video data to be labeled, and the audio and video data to be labeled is labeled as the target audio and video data.
[0085] The method for annotating audio and video data provided by the embodiment of the present disclosure can be implemented based on an electronic device, or a functional module or functional entity in the electronic device.
[0086] The electronic device may be a personal computer (PC), a server, a mobile phone, a tablet computer, a notebook computer, a mainframe computer, etc., which is not specifically limited in the embodiments of the present disclosure.
[0087] For example, Fig. 6A As shown in FIG. 1 , it is a hardware configuration block diagram of an electronic device 200 according to an embodiment of the present disclosure. Fig. 6A The electronic device 200 shown includes: a tuner 210, a communicator 220, a detector 230, an external device interface 240, a controller 250, a display 260, an audio output interface 270, a memory, a power supply, etc. Among them, the controller 250 includes a central processing unit, a video processor, an audio processor, a graphics processor, a RAM, a ROM, and a first interface to an nth interface for input / output. The display 260 can be at least one of a liquid crystal display, an OLED display, a touch display, and a projection display, and can also be a projection device and a projection screen. The tuner 210 receives broadcast television signals through wired or wireless reception, and demodulates audio and video signals from multiple wireless or wired broadcast television signals, such as and EPG audio and video data signals. The detector 230 is used to collect signals of the external environment or interact with the outside. The controller 250 and the tuner 210 can be located in different split devices, that is, the tuner 210 can also be in an external device of the main device where the controller 250 is located, such as an external set-top box.
[0088] In some embodiments, the electronic device is a terminal device with a display function, such as a television, a mobile phone, a computer, a learning machine, etc.
[0089] In some embodiments, the controller 250 controls the operation of the electronic device and responds to the user's operation through various software control programs stored in the memory. The controller 250 controls the overall operation of the electronic device 200. The user can input a user command in a graphical user interface (GUI) displayed on the display 260, and the user input interface receives the user input command through the graphical user interface (GUI). Alternatively, the user can input a user command by inputting a specific sound or gesture, and the user input interface recognizes the sound or gesture through a sensor to receive the user input command.
[0090] An output interface (display 260 and / or audio output interface 270), configured to output user interaction information;
[0091] The communicator 220 is used to communicate with the server 400 or other devices.
[0092] In some embodiments, the controller 250 can automatically extract characteristic parameters of the audio and video data to be annotated, and annotate the audio and video data to be annotated. The implementation method includes: obtaining first generated data according to the audio and video data to be annotated and the first controlled random data, wherein the first controlled random data is generated according to the characteristic distribution of the target audio and video data; determining a first similarity parameter and a second controlled random data corresponding to the first generated data, wherein the first similarity parameter is used to characterize the probability that the first generated data is the target audio and video data, and the second controlled random data is the first controlled random data restored according to the first generated data; when the first similarity parameter is greater than or equal to the similarity threshold, determining that the second controlled random data is the characteristic parameter of the audio and video data to be annotated, and annotating the audio and video data to be annotated as the target audio and video data.
[0093] In some embodiments, the controller 250 has a built-in target generative adversarial network model, which includes a generator and a discriminator, and the generator includes an encoder and a decoder; the controller 250 specifically extracts the feature parameters of the audio frequency data to be labeled through the generator and the discriminator;
[0094] The implementation methods in some embodiments include: the controller 250 may specifically input the audio and video data to be annotated into the encoder to generate the encoded audio and video data to be annotated;
[0095] The encoded audio and video data to be annotated and the first controlled random data are input into the decoder to obtain the first generated data. The generator is obtained after training according to the first objective function;
[0096] The first objective function is: max[logDfake +a*sum(logD condition )] (1);
[0097] In the above formula (1), D fake logD is the probability that the similarity parameter output by the discriminator is the target audio and video data when the input signal of the discriminator is the generated data; condition is the similarity of the controlled random data corresponding to each generated data output by the generator, and a is a constant.
[0098] The implementation methods in some embodiments include: the controller 250 may specifically input the first generated data into the discriminator; and obtain the first similarity parameter and the second controlled random data output by the discriminator.
[0099] The above discriminator is obtained after training according to the second objective function;
[0100] The second objective function is:
[0101] max[logD real +log(1-D fake )+a*sum(log(1-D condition ))] (2);
[0102] In the above formula (2), D real is the probability that the similarity parameter output by the discriminator is the target audio and video data when the input signal of the discriminator is real data; D fake is the probability that the similarity parameter output by the discriminator is the target audio and video data when the input signal of the discriminator is the generated data; log(1-D condition ) is the similarity degree of the controlled random data corresponding to each generated data output by the generator, and a is a constant.
[0103] In some embodiments, the controller 250 is further configured to determine the characteristic distribution of the target audio and video data according to a plurality of sample audio and video data before obtaining the first generated data according to the audio and video data to be labeled and the first controlled random data, wherein the plurality of sample audio and video data are the target audio and video data;
[0104] The first controlled random data is determined according to the characteristic distribution of the target audio and video data.
[0105] like Figure 6B As shown, Figure 6B FIG. 1 is a schematic diagram of software configuration in an electronic device 200 according to one or more embodiments of the present application. Figure 6BAs shown, the system is divided into four layers, from top to bottom, respectively, the application layer (Applications) (referred to as "application layer"), the application framework layer (Application Framework) (referred to as "framework layer"), the Android runtime (Android runtime) and the system library layer (referred to as "system runtime library layer"), and the kernel layer. The kernel layer includes at least one of the following drivers: audio driver, display driver, Bluetooth driver, camera driver, WIFI driver, USB driver, HDMI driver, sensor driver (such as fingerprint sensor, temperature sensor, pressure sensor, etc.), and power driver, etc. The annotation method of audio and video data provided in the embodiment of the present application can be implemented based on the above-mentioned electronic device.
[0106] The method for labeling audio and video data provided by the embodiment of the present disclosure can be implemented based on the above-mentioned electronic device. The method can determine first controlled random data according to the characteristic distribution of target audio and video data, and use the first controlled random data as noise to generate first generated data with the audio and video data to be labeled, and calculate the first similarity parameter corresponding to the first generated data, as well as the second controlled random data generated after the audio and video data is restored based on the first controlled random data. Since the first similarity parameter is used to characterize the probability that the first generated data is the target audio and video data, when the first similarity parameter is greater than or equal to the similarity threshold, the first generated data is considered to be the target audio and video data. The second controlled random number restored in this way can be considered to be the characteristic parameter of the target audio and video data. The scheme can automatically extract the characteristic parameters of the audio and video data to be labeled based on the characteristic distribution of the target audio and video data, and label the audio and video data to be labeled as the target audio and video data.
[0107] In order to explain this scheme in more detail, the following will be combined in an exemplary manner Figure 7 To explain, it is understandable that Figure 7 The steps involved may include more steps or fewer steps in actual implementation, and the order of these steps may also be different, so as to implement the audio and video data marking method provided in the embodiment of the present application.
[0108] like Figure 7 FIG. 1 is a flow chart of a method for labeling audio and video data in an embodiment of the present disclosure. Figure 1 The method for labeling audio and video data includes the following steps S701 to S704:
[0109] S701. Obtain first generated data according to the audio and video data to be annotated and the first controlled random data.
[0110] The method for labeling audio and video data provided in the embodiment of the present application can be applied to Figure 5 The target generative adversarial network model shown includes a generator and a discriminator, and the generator includes an encoder and a decoder.
[0111] In the above S701, according to the audio and video data to be labeled and the first controlled random data, the first generated data can be obtained by first inputting the audio and video data to be labeled and the first controlled random data into a generator; and obtaining the first generated data output by the generator.
[0112] In an embodiment of the present disclosure, the audio and video data to be labeled is input into an encoder in a generator of a target generative adversarial network model. The encoder encodes the audio and video data to be labeled to obtain a coded signal and inputs the coded signal into a decoder in the generator. The decoder decodes the coded signal based on the first controlled random data to obtain the first generated data.
[0113] Among them, the above-mentioned first controlled random data is generated according to the characteristic distribution of the target audio and video data.
[0114] In some embodiments, obtaining the first controlled random data may be achieved by the following steps:
[0115] A. Obtain a small amount of sample audio and video data and manually annotate them.
[0116] For example, Fig. 8A A small amount of sample audio and video data can be obtained from 100,000 samples, for example, 1,000 sample audio and video data can be obtained, and the sample audio and video data marked as the same type of audio and video data (which can be the target type of audio and video data) can be input into Fig. 8A In the classifier (classifer) shown in , a feature distribution function for the audio and video data of this type is obtained, and the feature distribution function is used to characterize the feature distribution of the target audio and video data.
[0117] The samples may be samples of audio and video data, and the audio and video data may be audio data and / or video data. The audio and video data types are marked for these small samples.
[0118] The above classifier can use a conventional full-link classifier.
[0119] B. Obtain sample audio and video data marked as target audio and video data from the manually labeled sample audio and video data.
[0120] The manually labeled sample audio and video data may all be sample audio and video data labeled as target-type audio and video data, or may have a portion of sample audio and video data labeled as target-type audio and video data.
[0121] For example, with respect to image data, assuming that there are 1,000 sample audio and video data labeled as apple images, and the feature of the apple images is the aspect ratio of the apple, then the classifier can generate a feature distribution of the apple images' aspect ratio based on these 1,000 samples. Suppose that the feature of the apple images is the color of the apple, then the classifier can generate a feature distribution of the apple images' color based on these 1,000 samples, and so on.
[0122] For example, for audio data, assuming that there are 1,000 audio data labeled as human voice samples, and the characteristic of the human voice is the frequency range, then the classifier can generate a characteristic distribution of the frequency range of the human voice based on these 1,000 samples. Figure 8B The figure shows a schematic diagram of a characteristic distribution of a frequency range of human voice, and the characteristic distribution is in the form of a normal distribution. Figure 8B Assuming that the horizontal axis represents the frequency range f and the vertical axis represents the probability P, we can get Figure 8B The curve shown in shows that the frequency range of human voice is between f1 and f2, among which the probability of being at f0 is the highest.
[0123] For example, for audio data, assuming that the characteristics of human voice are the phonemes of vowels, clustering can be performed based on the factors of vowels, because the frequency distribution of vowel factors is generally relatively stable and it is relatively easy to calculate its probability distribution. In this case, even existing human voice samples can be used for training to improve efficiency.
[0124] For example, with respect to audio data, assuming the characteristics of human voice, the tones of the language can also be clustered. The five types of tones, namely, first tone, second tone, third tone, fourth tone and light tone, can be clustered according to probability distribution. If the probability of the sound to be labeled being in these five tones is very low, it means that the possibility of it being a human voice is relatively low.
[0125] C. Input the sample of the target audio and video data into the classifier, and obtain the feature distribution function of the target audio and video data output by the classifier.
[0126] D. Obtain the first controlled random data from the characteristic distribution function of the target audio and video data.
[0127] In the disclosed embodiment, any value can be taken from the characteristic distribution function of the target audio and video data to obtain the first controlled random data (Conditional Random data). Conditional Random data can be used to mark the attribute characteristics of audio and video data. Usually, one audio and video data may have multiple attribute characteristics.
[0128] There are L attribute features in an audio and video data, denoted as: c1, c2, ... cL; then assuming that these L attribute features are independent of each other, the feature of the audio and video data can be globally expressed as:
[0129]
[0130] In traditional infoGAN, random vectors c are usually random noises that conform to normal distribution, or linear noises within a certain range. This allows each c to automatically match a feature of the image or signal. For example, for the MNIST handwritten character set, two random vectors c with the same probability distribution are used, one of which will get the thickness of the handwritten strokes, and the other will get the direction of character rotation.
[0131] Among them, the MNIST handwritten character set is a very classic data set in the field of machine learning. The MNIST handwritten character set is an entry-level computer vision data set that contains various handwritten digital image samples. Each sample is a 28*28 pixel grayscale handwritten digital image, and each pixel is an integer from 0 to 255.
[0132] The difference between the target generative adversarial network model in the disclosed embodiment and the traditional infoGAN model is that the goal of the traditional infoGAN model is to control the generated data according to certain rules, while the goal of the target generative adversarial network model in the disclosed embodiment is to classify unlabeled data. InfoGAN generates data from random vectors, which are featureless random variables, and we need to extract features from the original signal to generate controlled ConditionalRandom data (i.e., the process of generating the first controlled audio and video data in steps A to D above).
[0133] For image signals or audio signals with complex backgrounds, the random vector c usually input can only capture certain features of the entire signal. Therefore, if you want to classify a part of the signal, you must also analyze the features of that part and then derive the data range of the corresponding feature parameters, so that you can get the probability distribution of the feature.
[0134] For example, to identify apples on a fruit tree, we can use different Conditional Random data to mark different features, and we can give its data range from different features such as the color range, shape range, and size range of the apple. This work requires analysis of sample audio and video data, which is similar to the concept of a weak label. However, for data labeling, the number of data sets is often in the order of tens of thousands, hundreds of thousands, millions, or even tens of millions, which makes the labeling work extremely laborious. The above method in the disclosed embodiment can greatly save the time of data labeling.
[0135] Exemplary, combined Fig. 8A and Figure 5 ,like Fig. 9 FIG. 1 is a flow chart of a method for labeling audio and video data based on a target generative adversarial network model in an embodiment of the present disclosure, from Fig. 9 It can be seen that based on Fig. 8A The determined first controlled random data is input into the generator as random noise, and the first controlled random data is used to process the audio and video data to be annotated input by the generator.
[0136] based on Fig. 9 From the process shown, it can be known that the controller 250 in the electronic device 200 of the embodiment of the present disclosure will not only have a built-in target generative adversarial network model, but also a classifier. For example, as shown in Figure 10, it is a schematic diagram of the architecture of a controller in the embodiment of the present disclosure, and the controller has a built-in target generative adversarial network model and a classifier to implement the audio and video data labeling method provided by the embodiment of the present disclosure.
[0137] In the above embodiment, after manual labeling based on a small number of samples, a classifier can be used to generate first controlled random data, and the audio and video data to be labeled can be processed based on the first controlled random data to finally determine the feature parameters corresponding to the target audio and video data. This method can automatically extract the feature parameters of the target audio and video data through a small amount of manual labeling.
[0138] In the embodiment of the present disclosure, the generator in the target generative adversarial network model is obtained after training according to the first objective function. Since the first controlled random data is input into the generator as noise in the embodiment of the present disclosure, the first objective function in the embodiment of the present disclosure is improved on the basis of the traditional infoGAN model, and the first objective function is optimized as follows:
[0139] max[logD fake +a*sum(logD condition )] (1);
[0140] Among them, D fake logD is the probability that the similarity parameter output by the discriminator is the target audio and video data when the input signal of the discriminator is the generated data; condition is the similarity of the controlled random data corresponding to each generated data output by the generator, and a is a constant. In actual implementation, the above a can be debugged according to empirical values.
[0141] S702: Determine a first similarity parameter and second controlled random data corresponding to the first generated data.
[0142] Among them, the first similarity parameter is used to characterize the probability that the first generated data is the target type audio and video data, and the second controlled random data is the first controlled random data restored according to the first generated data.
[0143] In some embodiments, determining the first similarity parameter and the second controlled random data corresponding to the first generated data includes: inputting the first generated data into a discriminator; and obtaining the first similarity parameter and the second controlled random data output by the discriminator.
[0144] For example, Figure 5 As shown, the first generated data can be input into the discriminator, and the discriminator will discriminate whether the judgment result of the first generated data is true or false (true / false). If the discriminator determines that the first generated data is real data, the output judgment result is true; if the discriminator determines that the first generated data is not real data, the output judgment result is false, and the judgment result can be a similarity parameter, that is, a similarity parameter between the first generated data and the real data. The result of true / false can be obtained based on the size of the similarity parameter. At the same time, the recovery network in the discriminator can also simultaneously output the second controlled random data obtained when the first controlled random data is recovered based on the first generated data.
[0145] Among them, the discriminator in the above-mentioned target generative adversarial network model is obtained after training according to the second objective function. Since the first controlled random data is input into the generator as noise in the embodiment of the present disclosure, the second objective function in the embodiment of the present disclosure is also improved accordingly on the basis of the traditional infoGAN model, and the second objective function is optimized as follows:
[0146] max[logD real +log(1-D fake )+a*sum(log(1-D condition ))] (2);
[0147] Among them, D real is the probability that the similarity parameter output by the discriminator is the target audio and video data when the input signal of the discriminator is real data; D fakeis the probability that the similarity parameter output by the discriminator is the target audio and video data when the input signal of the discriminator is the generated data; log(1-D condition ) is the similarity of the controlled random data corresponding to each generated data output by the generator, and a is a constant. In actual implementation, the above a can be debugged according to empirical values.
[0148] S703. When the first similarity parameter is greater than or equal to the similarity threshold, the second controlled random data is determined as the characteristic parameter of the audio and video data to be annotated. The preset similarity threshold may be a threshold set. When the first similarity parameter is greater than or equal to the threshold, it indicates that the first generated data is determined to be real data, and the second controlled random data can be determined as the characteristic parameter of the audio and video data to be annotated. Correspondingly, when the first similarity parameter is less than the threshold, it indicates that the first generated data is determined to be generated data, and the second controlled random data will not be determined as the characteristic parameter of the audio and video data to be annotated.
[0149] In the disclosed embodiment, after obtaining the first similarity parameter, it is first determined whether the first similarity parameter is greater than or equal to the similarity threshold. When the first similarity parameter is greater than or equal to the similarity threshold, the first generated data is considered to be real data, and the second controlled random data is determined as the characteristic parameter of the audio and video data to be labeled. When it is determined that the first similarity parameter is less than the similarity threshold, the first generated data is considered not to be real data, and the second controlled random data is discarded.
[0150] S704: Label the audio and video data to be labeled as target audio and video data.
[0151] Since the second controlled random data is the data obtained by recovering the first controlled random number data based on the first generated data, the data type represented by the second controlled random data should be the same as that of the first controlled random data. Then, when the first controlled random data is the characteristic distribution of the target audio and video data, the second controlled random data also represents the characteristic distribution of the target audio and video data. Furthermore, since the second controlled random data is determined as the characteristic parameter of the audio and video data to be labeled when the first similarity parameter is greater than or equal to the similarity threshold, it can be determined that the audio and video data to be labeled is the target audio and video data. At this time, the audio and video data to be labeled can be labeled as the target audio and video data.
[0152] The method for labeling audio and video data provided by the embodiment of the present disclosure can determine the first controlled random data according to the characteristic distribution of the target audio and video data, and use the first controlled random data as noise to generate the first generated data with the audio and video data to be labeled, and calculate the first similarity parameter corresponding to the first generated data, as well as the second controlled random data generated after the audio and video data is restored based on the first controlled random data. Since the first similarity parameter is used to characterize the probability that the first generated data is the target audio and video data, when the first similarity parameter is greater than or equal to the similarity threshold, the first generated data is considered to be the target audio and video data. The second controlled random number recovered in this way can be considered to be the characteristic parameter of the audio and video data to be labeled, and it can be determined that the audio and video data to be labeled is the target audio and video data. Fig.11 FIG. 1 is a block diagram of a device for annotating audio and video data provided by an embodiment of the present disclosure, the device comprising:
[0153] A generating module 1101 is used to obtain first generated data according to the audio and video data to be labeled and the first controlled random data, where the first controlled random data is generated according to the characteristic distribution of the target audio and video data;
[0154] A discriminating module 1102 is configured to determine a first similarity parameter and a second controlled random data corresponding to the first generated data, wherein the first similarity parameter is used to characterize the probability that the first generated data is the target audio and video data, and the second controlled random data is the first controlled random data restored according to the first generated data;
[0155] The labeling module 1103 is used to determine that the second controlled random data is a characteristic parameter of the audio and video data to be labeled when the first similarity parameter is greater than or equal to a similarity threshold, and label the audio and video data to be labeled as the target audio and video data.
[0156] As an optional implementation of the embodiment of the present disclosure, the device is built with a target generative adversarial network model, the target generative adversarial network model includes a generator and a discriminator, and the generator includes an encoder and a decoder;
[0157] The generating module 1101 is specifically used for:
[0158] Inputting the audio and video data to be annotated into the encoder to generate the encoded audio and video data to be annotated;
[0159] The encoded audio and video data to be marked and the first controlled random data are input into the decoder to obtain the first generated data.
[0160] As an optional implementation of the embodiment of the present disclosure, the generator is obtained after training according to the first objective function;
[0161] The first objective function is: max[logD fake +a*sum(logD condition )] (1);
[0162] Among them, D fake logD is the probability that the similarity parameter output by the discriminator is the target audio and video data when the input signal of the discriminator is the generated data; condition is the similarity of the controlled random data corresponding to each generated data output by the generator, and a is a constant.
[0163] As an optional implementation of the embodiment of the present disclosure, the device is built with a target generative adversarial network model, the target generative adversarial network model includes a generator and a discriminator, and the generator includes an encoder and a decoder;
[0164] The identification module 1102 is specifically used for:
[0165] inputting the first generated data into the discriminator;
[0166] The first similarity parameter and the second controlled random data output by the discriminator are obtained.
[0167] As an optional implementation of the embodiment of the present disclosure, the discriminator is obtained after training according to the second objective function;
[0168] The second objective function is:
[0169] max[logD real +log(1-D fake )+a*sum(log(1-D condition ))] (2);
[0170] Among them, D real is the probability that the similarity parameter output by the discriminator is the target audio and video data when the input signal of the discriminator is real data; D fake is the probability that the similarity parameter output by the discriminator is the target audio and video data when the input signal of the discriminator is the generated data; log (1-D condition ) is the similarity degree of the controlled random data corresponding to each generated data output by the generator, and a is a constant.
[0171] As an optional implementation of the embodiment of the present disclosure, Fig.11 As shown, the device also includes:
[0172] A determination module 1104 is configured to obtain the audio and video data to be annotated before the first generated data according to the audio and video data to be annotated and the first controlled random data, and determine the characteristic distribution of the target audio and video data according to a plurality of sample audio and video data, wherein the plurality of sample audio and video data are the target audio and video data;
[0173] The first controlled random data is determined according to the characteristic distribution of the target audio and video data.
[0174] The above-mentioned audio and video data annotation device can implement each process executed by the above-mentioned audio and video data annotation method, and can achieve the same technical effect. To avoid repetition, it will not be repeated here.
[0175] An embodiment of the present application provides an electronic device, including: a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program implements the above-mentioned audio and video data marking method when executed by the processor.
[0176] An embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the various processes of the above-mentioned audio and video data annotation method are implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.
[0177] The computer readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.
[0178] The present application provides a computer program product, characterized in that it includes: when the computer program product is run on a computer, the computer is enabled to implement the above-mentioned audio and video data labeling method.
[0179] For the convenience of explanation, the above description has been made in conjunction with specific embodiments. However, the above discussion in some embodiments is not intended to be exhaustive or limit the embodiments to the specific forms disclosed above. According to the above teachings, various modifications and variations can be obtained. The selection and description of the above embodiments are to better explain the principles and practical applications, so that those skilled in the art can better use the embodiments and various different variations of the embodiments suitable for specific use considerations.
Claims
1. A method for labeling audio and video data, characterized in that: include: Acquire first generated data according to the audio and video data to be labeled and the first controlled random data, wherein the first controlled random data is generated according to the characteristic distribution function of the target audio and video data; the characteristic distribution function of the target audio and video data is the output of the classifier when a sample of the target audio and video data is input into the classifier; Determine a first similarity parameter and a second controlled random data corresponding to the first generated data, wherein the first similarity parameter is used to characterize the probability that the first generated data is the target audio and video data, and the second controlled random data is the first controlled random data restored according to the first generated data; When the first similarity parameter is greater than or equal to a similarity threshold, the second controlled random data is determined to be a characteristic parameter of the audio and video data to be labeled, and the audio and video data to be labeled is labeled as the target audio and video data.
2. The method according to claim 1, characterized in that: The method is applied to a target generative adversarial network model, wherein the target generative adversarial network model includes a generator and a discriminator, and the generator includes an encoder and a decoder; The step of obtaining first generated data according to the audio and video data to be annotated and the first controlled random data includes: Inputting the audio and video data to be annotated into the encoder to generate the encoded audio and video data to be annotated; The encoded audio and video data to be marked and the first controlled random data are input into the decoder to obtain the first generated data.
3. The method according to claim 2, characterized in that The generator is obtained after training according to the first objective function; The first objective function is: max[logD fake +a*sum(logD condition )]; Among them, D fake logD is the probability that the similarity parameter output by the discriminator is the target audio and video data when the input signal of the discriminator is the generated data; condition is the similarity of the controlled random data corresponding to each generated data output by the generator, and a is a constant.
4. The method according to claim 1, characterized in that: The method is applied to a target generative adversarial network model, wherein the target generative adversarial network model includes a generator and a discriminator, and the generator includes an encoder and a decoder; The determining of the first similarity parameter and the second controlled random data corresponding to the first generated data includes: inputting the first generated data into the discriminator; The first similarity parameter and the second controlled random data output by the discriminator are obtained.
5. The method according to claim 4, characterized in that The discriminator is obtained after training according to the second objective function; The second objective function is: max[logD real +log(1-D fake )+a*sum(log(1-D condition ))]; Among them, D real is the probability that the similarity parameter output by the discriminator is the target audio and video data when the input signal of the discriminator is real data; D fake is the probability that the similarity parameter output by the discriminator is the target audio and video data when the input signal of the discriminator is the generated data; log(1-D condition ) is the similarity degree of the controlled random data corresponding to each generated data output by the generator, and a is a constant.
6. The method according to claim 1, characterized in that Before obtaining the first generated data according to the audio and video data to be annotated and the first controlled random data, the method further includes: Determining the characteristic distribution of the target audio and video data according to a plurality of sample audio and video data, wherein the plurality of sample audio and video data are the target audio and video data; The first controlled random data is determined according to the characteristic distribution of the target audio and video data.
7. A device for labeling audio and video data, characterized in that: include: A generation module, configured to obtain first generated data according to the audio and video data to be labeled and first controlled random data, wherein the first controlled random data is generated according to a characteristic distribution function of the target audio and video data; the characteristic distribution function of the target audio and video data is the output of the classifier when a sample of the target audio and video data is input into the classifier; a discriminant module, configured to determine a first similarity parameter and a second controlled random data corresponding to the first generated data, wherein the first similarity parameter is used to characterize the probability that the first generated data is the target audio and video data, and the second controlled random data is the first controlled random data restored according to the first generated data; The labeling module is used to determine that the second controlled random data is the characteristic parameter of the audio and video data to be labeled when the first similarity parameter is greater than or equal to the similarity threshold, and label the audio and video data to be labeled as the target audio and video data.
8. An electronic device, characterized in that: The method comprises: a controller, configured to obtain first generated data according to the audio and video data to be labeled and first controlled random data, wherein the first controlled random data is generated according to a characteristic distribution function of the target audio and video data; the characteristic distribution function of the target audio and video data is the output of the classifier when a sample of the target audio and video data is input into the classifier; Determine a first similarity parameter and a second controlled random data corresponding to the first generated data, wherein the first similarity parameter is used to characterize the probability that the first generated data is the target audio and video data, and the second controlled random data is the first controlled random data restored according to the first generated data; When the first similarity parameter is greater than or equal to a similarity threshold, the second controlled random data is determined to be a characteristic parameter of the audio and video data to be labeled, and the audio and video data to be labeled is labeled as the target audio and video data.
9. A computer-readable storage medium, characterized in that: include: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for marking audio and video data according to any one of claims 1 to 6 is implemented.
10. A computer program product, characterized in that include: When the computer program product runs on a computer, the computer implements the method for marking audio and video data according to any one of claims 1 to 6.
Citation Information
Patent Citations
Neural network training method and device and image generation method and device
CN110634167A
Model training method and device, recognition method, storage medium and electronic equipment
CN110738227A
Configured artificial intelligence scene application research and development method and system
CN112529023A
CGAN-based environment monitoring method, device and system and storage medium
CN112712002A
Image processing model training method and device
CN113160027A
