A method and apparatus for identifying

Abstract

The application provides a recognition method and device. The method comprises: processing target recognition model on to-be-recognized information to obtain a plurality of target feature values. The target recognition model comprises a plurality of binary classifiers corresponding to a plurality of categories. Each binary classifier is used to output a target feature value of to-be-recognized information. The target feature value is used to represent the consistency degree between the real category of to-be-recognized information and the category corresponding to the binary classifier. For each binary classifier, a first reference probability that to-be-recognized information belongs to the category corresponding to the binary classifier is determined on a corresponding first probability density distribution curve, and a second reference probability that to-be-recognized information does not belong to the category corresponding to the binary classifier is determined on a corresponding second probability density distribution curve. The recognition result of to-be-recognized information is determined according to the first reference probability and the second reference probability. The application can improve the recognition accuracy of to-be-recognized information.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and more specifically, to a recognition method and apparatus. Background Technology

[0002] With the rapid development of artificial intelligence technology, deep learning models are widely used in information recognition, such as image recognition and speech recognition. However, the output layer of current recognition models typically uses a softmax function with an independence assumption for classification calculation. This independence assumption method has the drawback of failing to correctly distinguish between overlapping categories. While the independence assumption may hold true for discrete problems, such as defining heads and tails of a coin or the six sides of a die, it cannot correctly distinguish between similar problems, such as the distinction between the front and back nasal consonants "en" and "eng". Therefore, using the independence assumption to handle similarity problems can lead to biased recognition results and reduced accuracy. Summary of the Invention

[0003] The embodiments of this application provide an identification method and apparatus, and the technical solution provided by this application can improve the identification accuracy of the information to be identified.

[0004] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part from practice of this application.

[0005] According to a first aspect of the embodiments of this application, an identification method is provided, the method comprising: acquiring information to be identified; processing the information to be identified using a trained target identification model to obtain multiple target feature values ​​of the information to be identified, wherein the target identification model includes multiple binary classifiers corresponding one-to-one with multiple categories, each binary classifier outputting a target feature value of the information to be identified, the target feature value being used to characterize the degree of consistency between the true category of the information to be identified and the category corresponding to the binary classifier; for each binary classifier, determining, on a preset first probability density distribution curve corresponding to the binary classifier, the probability corresponding to the target feature value output by the binary classifier, as a first reference probability that the information to be identified belongs to the category corresponding to the binary classifier, and on a preset second probability density distribution curve corresponding to the binary classifier, determining, as a second reference probability that the information to be identified does not belong to the category corresponding to the binary classifier; determining, based on the first reference probability and the second reference probability, the target probability that the information to be identified belongs to the category corresponding to the binary classifier; and determining, based on the target probability of each of the multiple binary classifiers, the identification result of the information to be identified.

[0006] In some embodiments of this application, based on the foregoing scheme, the method further includes: acquiring training data, the training data including various sample information; training a pre-constructed first recognition model based on the training data to obtain a target recognition model, wherein the first recognition model includes a first output layer, the first output layer including multiple initial binary classifiers corresponding one-to-one with various categories, each initial binary classifier being used to output feature values ​​of sample information.

[0007] In some embodiments of this application, based on the foregoing scheme, before training the pre-built first recognition model, the method further includes: setting a matching first reference feature value and second reference feature value for each initial binary classifier, so that the feature value output by the initial binary classifier is located within the interval formed by the first reference feature value and the second reference feature value, wherein the first reference feature value is the feature value that the initial binary classifier should output when the true category of the sample information is consistent with the category corresponding to the initial binary classifier, and the second reference feature value is the feature value that the initial binary classifier should output when the true category of the sample information is inconsistent with the category corresponding to the initial binary classifier.

[0008] In some embodiments of this application, based on the foregoing scheme, the first recognition model further includes a second output layer parallel to the first output layer. The second output layer includes a multi-classifier, which outputs the posterior probability of sample information belonging to each of multiple categories. Training the pre-constructed first recognition model to obtain a target recognition model includes: inputting training data into the first recognition model to obtain a first output result output by the first output layer and a second output result output by the second output layer; determining a first loss data of the first recognition model based on the first output result using a probabilistic loss function, and determining a second loss data of the first recognition model based on the second output result using a distributed loss function; determining the target loss data of the first recognition model based on the first loss data and the second loss data; updating the model parameters of the first recognition model in reverse based on the target loss data to obtain a new first recognition model; returning to the step of obtaining training data until the first recognition model converges to obtain a second recognition model; and removing or freezing the first output layer in the second recognition model to obtain the target recognition model.

[0009] In some embodiments of this application, based on the foregoing scheme, the method further includes: acquiring test data, the test data including various test information; inputting the test data into a target recognition model to obtain multiple sets of test results corresponding one-to-one with multiple binary classifiers, the test results including feature values ​​of various test information output by the binary classifiers; for each binary classifier, extracting first result data and second result data from the test results corresponding to that binary classifier, the first result data including the feature value output by the binary classifier when the true category of the test information is consistent with the category corresponding to that binary classifier, the second result data including the feature value output by the binary classifier when the true category of the test information is inconsistent with the category corresponding to that binary classifier; determining a first probability density distribution curve corresponding to that binary classifier based on the first result data, and determining a second probability density distribution curve corresponding to that binary classifier based on the second result data.

[0010] In some embodiments of this application, based on the aforementioned scheme, determining the first probability density distribution curve corresponding to the binary classifier based on the first result data includes: obtaining an interval consisting of a first reference feature value and a second reference feature value corresponding to the binary classifier, as a feature value interval; dividing the feature value interval into multiple feature value sub-intervals; determining the number of feature values ​​in the first result data located in each feature value sub-interval, obtaining multiple first quantities that correspond one-to-one with the multiple feature value sub-intervals; performing curve fitting on the relationship between the multiple feature value sub-intervals and the multiple first quantities to obtain a first reference curve; and normalizing the first reference curve to obtain the first probability density distribution curve corresponding to the binary classifier.

[0011] In some embodiments of this application, based on the aforementioned scheme, determining the second probability density distribution curve corresponding to the binary classifier based on the second result data includes: determining the number of feature values ​​located in each feature value sub-interval in the second result data to obtain a plurality of second quantities corresponding one-to-one with the plurality of feature value sub-intervals; performing curve fitting on the relationship between the plurality of feature value sub-intervals and the plurality of second quantities to obtain a second reference curve; and performing normalization processing on the second reference curve to obtain the second probability density distribution curve corresponding to the binary classifier.

[0012] In some embodiments of this application, based on the foregoing scheme, determining the target probability that the information to be identified belongs to the category corresponding to the binary classifier according to the first reference probability and the second reference probability includes: calculating the target probability using the following formula:

[0013]

[0014] Where P0 represents the target probability, P1 represents the first reference probability, P2 represents the second reference probability, a is a first set parameter greater than or equal to 0 and less than or equal to 1, and n is a second set parameter greater than a.

[0015] In some embodiments of this application, based on the foregoing scheme, the target recognition model includes an acoustic model for performing speech recognition.

[0016] According to a second aspect of the embodiments of this application, an identification device is provided, the device comprising: an acquisition unit for acquiring information to be identified; a processing unit for processing the information to be identified using a trained target identification model to obtain multiple target feature values ​​of the information to be identified, wherein the target identification model includes multiple binary classifiers corresponding one-to-one with multiple categories, each binary classifier outputting a target feature value of the information to be identified, the target feature value being used to characterize the degree of consistency between the true category of the information to be identified and the category corresponding to the binary classifier; and a first determination unit for determining, for each binary classifier, on a preset first probability density distribution curve corresponding to the binary classifier. The system comprises three parts: a first reference probability unit, a second reference probability unit, and a third reference probability unit. The first reference probability unit determines the probability that the information to be identified belongs to the category corresponding to the binary classifier, and the second reference probability unit determines the probability that the information to be identified does not belong to the category corresponding to the binary classifier, based on the first and second reference probabilities. The second determination unit determines the target probability that the information to be identified belongs to the category corresponding to the binary classifier based on the first and second reference probabilities.

[0017] In the technical solution of this application, when identifying the information to be identified, the probability of the information to be identified belonging to each of the multiple categories is determined by a target recognition model including multiple binary classifiers and the probability density distribution curve corresponding to each binary classifier. This abandons the ∑P in the traditional recognition model. i =1 constraint, where P i The posterior probability of the information to be identified belonging to the i-th category among multiple categories is used to predict the probability of the information to be identified belonging to each category among multiple categories by using the probability density distribution curve corresponding to the binary classifier. In this way, the independent conditional probability is transformed into a non-independent distribution, and the hard boundary problem is transformed into a soft boundary problem. For categories with similarity, there is a problem of overlapping features. Since the identification method of this application performs independent calculations for each category, the calculation results will not conflict due to similarity, thereby improving the identification accuracy of the information to be identified.

[0018] Furthermore, in traditional recognition models, output modeling requires modeling the entire set, which makes them susceptible to uneven data distribution. Once they encounter out-of-set issues, they will inevitably output incorrect results. However, in the technical solution of this application, the problem of the entire set is transformed into an independent problem of whether the information to be identified belongs to each category. In the technical solution of this application, even for untrained data, the probability of each target can be calculated based on the similarity with the features of each dimension of the existing category, making it more robust to out-of-set data.

[0019] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0020] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. In the drawings:

[0021] Figure 1 A flowchart illustrating an identification method according to an embodiment of this application is shown;

[0022] Figure 2 A flowchart illustrating the process of training a target recognition model according to an embodiment of this application is shown;

[0023] Figure 3 A schematic diagram of the structure of a first identification model according to an embodiment of this application is shown;

[0024] Figure 4 A detailed flowchart illustrating the training process for obtaining a target recognition model according to an embodiment of this application is shown.

[0025] Figure 5 A detailed flowchart illustrating the construction of a first probability density distribution curve and a second probability density distribution curve according to an embodiment of this application is shown.

[0026] Figure 6 A detailed flowchart illustrating the process of constructing a first probability density distribution curve according to an embodiment of this application is shown.

[0027] Figure 7 A schematic diagram of a first reference curve and a second reference curve according to an embodiment of this application is shown;

[0028] Figure 8 A block diagram of an identification device according to an embodiment of this application is shown. Detailed Implementation

[0029] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided to make this application more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art.

[0030] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0031] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0032] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0033] It should be noted that "multiple" in this article refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0034] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such uses of these terms can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described.

[0035] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0036] The following detailed description of some embodiments of this application will be provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0037] See Figure 1 The diagram illustrates a flowchart of an identification method according to an embodiment of this application, specifically including the following steps S110 to S160:

[0038] S110, Obtain the information to be identified.

[0039] In this application, the information to be identified can be image information, acoustic feature information of the speech to be identified, or other information with recognition requirements; this application does not limit the scope of the information. It is understood that the type of information to be identified is adapted to the target recognition model.

[0040] See also Figure 1 S120, The trained target recognition model is used to process the information to be recognized to obtain multiple target feature values ​​of the information to be recognized. The target recognition model includes multiple binary classifiers that correspond one-to-one with multiple categories. Each binary classifier is used to output the target feature value of the information to be recognized. The target feature value is used to characterize the degree of consistency between the true category of the information to be recognized and the category corresponding to the binary classifier.

[0041] It should be noted that the target feature values ​​output by the binary classifier of the target recognition model are not probability values, but target feature values ​​of the information to be recognized. Therefore, the multiple target feature values ​​output by the target recognition model are non-normalized, that is, the sum of multiple target feature values ​​is not 1. The recognition result of the information to be recognized cannot be directly determined based on multiple target feature values.

[0042] It should also be noted that the target recognition model can be an acoustic model for speech recognition, an image recognition model for image recognition, or a recognition model for other purposes. This application does not limit the specific model.

[0043] In some implementations, it can be achieved by performing, such as Figure 2 The steps shown yield the target recognition model.

[0044] See Figure 2Specifically, this includes the following S210 to S220:

[0045] S210, Obtain training data, which includes various sample information.

[0046] The training data also includes category labels that correspond one-to-one with various sample information.

[0047] For example, if it is necessary to train an acoustic model for speech recognition, taking the use of pinyin as the modeling unit of the acoustic model as an example, a pronunciation dictionary from single characters to pinyin can be constructed first based on expert knowledge. Then, based on the pronunciation dictionary, the text annotation of the speech data is translated into a dictionary, converting the text sequence into the corresponding pinyin sequence. After that, the speech data is processed by segmentation, the acoustic features are extracted from each frame, and each frame is labeled with the pinyin category corresponding to the current frame according to the pinyin sequence, so as to obtain training data in which the sound features and pinyin categories correspond one-to-one frame.

[0048] See also Figure 2 S220, Based on the training data, the pre-constructed first recognition model is trained to obtain the target recognition model. The first recognition model includes a first output layer, which includes multiple initial binary classifiers that correspond one-to-one with multiple categories. Each initial binary classifier is used to output the feature value of the sample information.

[0049] When training the first recognition model, a function without a summation constraint of 1 can be selected as the loss function, that is, a distributed loss function, such as the MSE loss function, can be selected.

[0050] In some implementations, the first recognition model may consist of only a first output layer.

[0051] In other embodiments, the first recognition model may further include a second output layer in parallel with the first output layer, the second output layer including a multi-classifier, which outputs the posterior probability of sample information belonging to each of a variety of categories.

[0052] This multi-classifier calculates posterior probabilities using the Softmax function. The posterior probabilities calculated by the Softmax function are subject to probability normalization, meaning that for each sample information, the sum of the posterior probabilities output by the second output layer is 1.

[0053] Specifically, the model structure of the first recognition model can be as follows: Figure 3 As shown. The first output layer and the second output layer run in parallel, and the data output from the embedding layer is synchronously input to the first output layer and the second output layer respectively.

[0054] In some implementations, the following steps may also be performed before performing S220 above:

[0055] For each initial binary classifier, a matching first reference feature value and a second reference feature value are set so that the feature value output by the initial binary classifier is located within the interval formed by the first reference feature value and the second reference feature value. The first reference feature value is the feature value that the initial binary classifier should output when the true category of the sample information is consistent with the category corresponding to the initial binary classifier. The second reference feature value is the feature value that the initial binary classifier should output when the true category of the sample information is inconsistent with the category corresponding to the initial binary classifier.

[0056] The first reference eigenvalue can be set to any positive number, and the second reference eigenvalue can be set to 0 or a negative number close to 0.

[0057] It should be noted that the first reference feature value for each initial binary classifier can be the same or different. Similarly, the second reference feature value for each initial binary classifier can be the same or different.

[0058] It is understandable that setting a first reference feature value and a second reference feature value for each initial binary classifier is ideal. Ideally, the initial binary classifier will output the first reference feature value when the true category of the sample information matches the category corresponding to the initial binary classifier, and the second reference feature value when the true category of the sample information does not match the category corresponding to the initial binary classifier. However, since the computational effect of the initial binary classifier is unlikely to reach an absolutely ideal level, the output feature value will be within the interval formed by the first reference feature value and the second reference feature value.

[0059] Furthermore, the binary classifier of the target recognition model obtained after subsequent training, after processing the information to be recognized, outputs target feature values ​​that are also constrained within the interval formed by the first reference feature value and the second reference feature value.

[0060] For example, for the initial binary classifier 1, the first reference feature value can be set to 100, and the second reference feature value can be set to 0. If the initial binary classifier corresponds to category 1, when the true category of the input sample information is category 1, the initial binary classifier 1 should output 100, and when the true category of the input sample information is not category 1, the initial binary classifier 1 should output 0.

[0061] In this embodiment, because the deep learning training model is ultimately trained according to the definition of the category labels of the sample information, for the training of Pinyin, assuming there are a total of 100 Pinyin characters, even with a uniform distribution, the ratio of positive to negative samples will be 1:99. In this case, the first recognition model is prone to convergence to predicting all negative samples due to the uneven distribution of positive and negative samples, which will severely affect the model's performance. Therefore, for the training loss function with the non-independence assumption, when training category y... i When initializing the binary classifier, it can be set that when the class label is class y... i Initially, the feature value output by the binary classifier is defined as 100. When the class label is not class y... i At this point, the feature value output by the initial binary classifier is defined as 0. After training in this way, because the loss function for positive samples is 100 times that for negative samples, the first recognition model has a greater optimization force on positive samples, which can alleviate the problem caused by data imbalance.

[0062] In S220, the pre-built first recognition model is trained based on the training data to obtain the target recognition model. The specific implementation method can be as follows: Figure 4 Perform the steps shown:

[0063] See Figure 4 Specifically, this includes the following S221 to S226:

[0064] S221, input the training data into the first recognition model to obtain the first output result output by the first output layer and the second output result output by the second output layer.

[0065] Understandably, the first output includes the posterior probability of each sample in the training data belonging to each of the multiple categories. The second output includes the feature values ​​output by each initial binary classifier for each sample in the training data.

[0066] See also Figure 4 S222, Based on the first output result, determine the first loss data of the first recognition model through a probabilistic loss function, and based on the second output result, determine the second loss data of the first recognition model through a distributed loss function.

[0067] In this embodiment, the probabilistic loss function can be the cross-entropy loss function, and the distributed loss function can be the MSE loss function.

[0068] For example, in Figure 3The diagram illustrates n initial binary classifiers corresponding to n categories. These n initial binary classifiers constitute the first output layer. After the sample information is input into the first recognition model, the vector output by the embedding layer is input into both the first output layer and the second output layer. In the first output layer, the n initial binary classifiers output n feature values, including S1 to Sn; the second output layer outputs n posterior probabilities, including P1 to Pn.

[0069] Furthermore, using a probabilistic loss function, n second loss values, including L11 to L1n, are calculated that correspond one-to-one with the n posterior probabilities; using a distributed loss function, n first loss values, including L21 to L2n, are calculated that correspond one-to-one with the n feature values.

[0070] See also Figure 4 S223, Based on the first loss data and the second loss data, determine the target loss data of the first identification model.

[0071] In this embodiment, the specific implementation method may be as follows: for each category among multiple categories, determine the first loss value and the second loss value corresponding to that category from the first loss data and the second loss data respectively, and add the first loss value and the second loss value corresponding to that category to obtain the target loss value corresponding to that category; use the target loss value of each category among multiple categories as the target loss data.

[0072] For example, in Figure 3 In the first output layer, the sample information is processed by a distributed loss function to obtain the first loss data, which includes S1-Sn. The sample information is processed by a second output layer to obtain the second loss data, which includes P1-Pn, which includes P1-Pn. Here, S1 / P1 is the loss value obtained for category 1, S2 / P2 is the loss value obtained for category 2, ..., and further, S1+P1 is the target loss value for category 1, and S2+P2 is the target loss value for category 2.

[0073] See also Figure 4 S224, Based on the target loss data, the model parameters of the first recognition model are updated in reverse to obtain a new first recognition model.

[0074] See also Figure 4 S225, return to the step of obtaining training data until the first recognition model converges and the second recognition model is obtained.

[0075] It is understandable that after each acquisition of training data, a new first recognition model is trained, and the first recognition model is continuously optimized. After the model converges, it is used as the second recognition model.

[0076] See also Figure 4S226, Remove or freeze the first output layer in the second recognition model to obtain the target recognition model.

[0077] Understandably, the second output layer is essentially an external layer used to train the target recognition model. It assists in optimizing the training of the first recognition model through the second output layer and a probabilistic loss function. This probabilistic loss function is easy to converge and can improve the final convergence effect of the first recognition model. After training is completed, the second output layer is no longer used in actual application. Therefore, the second output layer can be removed or frozen so that when the target recognition model is processing the information to be recognized, the data will not flow to the second output layer, but will only be processed through the first output layer. The final data only contains the target feature values ​​output by each binary classifier in the target recognition model.

[0078] It's worth noting that traditional recognition models previously placed high demands on the orthogonality of the output categories, or required sequential-biased calculation methods for differentiation. However, both are relatively difficult to implement. The former requires rigorous mathematical analysis of the original problem, which is impossible in typical training. The latter still doesn't address the issue of data overlap; when using sequential methods for direct classification, overlapping information is easily overlooked, leading to overfitting.

[0079] Compared to sequence classification, as mentioned earlier, sequence classification methods are often inaccurate in calculating the timestamps of the returned results, especially for speech problems. Common end-to-end solutions such as CTC and RNNT are prone to shifting the timing of spikes. Furthermore, traditional recognition models generally require a relatively long input time window, causing the location of spikes to frequently become a performance bottleneck in scenarios with high latency requirements. However, under the assumption of non-probabilistic independence, the granularity of the entire input length for the decision can be minimized, thereby fundamentally and more easily improving response speed in low-latency scenarios.

[0080] See also Figure 1 S130, for each binary classifier, on a preset first probability density distribution curve corresponding to the binary classifier, determine the probability corresponding to the target feature value output by the binary classifier as a first reference probability that the information to be identified belongs to the category corresponding to the binary classifier, and on a preset second probability density distribution curve corresponding to the binary classifier, determine the probability corresponding to the target feature value output by the binary classifier as a second reference probability that the information to be identified does not belong to the category corresponding to the binary classifier.

[0081] In this application, because the first output layer of the target recognition model is not trained based on a probabilistic loss function, the overall distribution of the final model cannot be guaranteed to satisfy the normalized probability distribution function. Therefore, after the target recognition model is trained, it is necessary to perform probability statistics on the binary classifier corresponding to each category on a development set and fit the corresponding probability density distribution curve.

[0082] It should be noted that in the target recognition model, each binary classifier has a corresponding first probability density distribution curve and a second probability density distribution curve.

[0083] In some implementations, a first probability density distribution curve and a second probability density distribution curve of the binary classifier are constructed. Specific implementations can be as follows: Figure 5 Perform the steps shown:

[0084] See Figure 5 The diagram illustrates a detailed process for constructing a first probability density distribution curve and a second probability density distribution curve according to an embodiment of this application, specifically including the following steps S510 to S540:

[0085] S510 acquires test data, which includes various test information.

[0086] The test data also includes category labels that correspond one-to-one with various test information.

[0087] In this embodiment, if the target recognition model is an acoustic model, a batch of audio can be compiled and labeled in accordance with the above step S210 to obtain test data.

[0088] See also Figure 5 S520 inputs the test data into the target recognition model to obtain multiple sets of test results corresponding one-to-one with multiple binary classifiers. The test results include feature values ​​of various test information output by the binary classifiers.

[0089] It is understandable that after each type of test information in the test data is input into the target recognition model, each binary classifier in the target recognition model will output the feature value corresponding to that test information.

[0090] For example, suppose the test data includes 100 types of test information, and the target recognition model includes 10 binary classifiers. When the 100 types of test information are input into the target recognition model, each binary classifier will output 100 feature values. That is, a set of test results includes 100 feature values, and with 10 binary classifiers, 10 sets of test results will be obtained.

[0091] See also Figure 5S530, for each binary classifier, extract first result data and second result data from the test results corresponding to the binary classifier. The first result data includes the feature value output by the binary classifier when the true category of the test information is consistent with the category corresponding to the binary classifier. The second result data includes the feature value output by the binary classifier when the true category of the test information is inconsistent with the category corresponding to the binary classifier.

[0092] For example, suppose the test data includes test information 1 belonging to category 1, test information 2 belonging to category 2, and test information 3 belonging to category 3, and the target recognition model includes binary classifier 1 corresponding to category 1, binary classifier 2 corresponding to category 2, and binary classifier 3 corresponding to category 3.

[0093] Furthermore, after inputting the test data into the target recognition model, binary classifier 1 outputs feature value 11 for test information 1, feature value 12 for test information 2, and feature value 13 for test information 3; binary classifier 2 outputs feature value 21 for test information 1, feature value 22 for test information 2, and feature value 23 for test information 3; binary classifier 3 outputs feature value 31 for test information 1, feature value 32 for test information 2, and feature value 33 for test information 3.

[0094] Furthermore, for binary classifier 1, since the initial binary classifier 1 corresponds to category 1, the true category of test information 1 is consistent with the category corresponding to binary classifier 1, and the first result data corresponding to binary classifier 1 includes feature value 11; the true categories of test information 2 and test information 3 are inconsistent with the category corresponding to binary classifier 1, and the second result data of binary classifier 1 includes feature value 12 and feature value 13.

[0095] Similarly, the first result data of binary classifier 2 includes feature value 22, and the second result data includes feature value 21 and feature value 23; the first result data of binary classifier 3 includes feature value 33, and the second result data includes feature value 31 and feature value 32.

[0096] See also Figure 5 S540, determine the first probability density distribution curve corresponding to the binary classifier based on the first result data, and determine the second probability density distribution curve corresponding to the binary classifier based on the second result data.

[0097] In S540, a first probability density distribution curve corresponding to the binary classifier is determined based on the first result data. At least two specific implementation methods exist:

[0098] In the first implementation, curve fitting is performed on the first result data, and the curve obtained by curve fitting is normalized to obtain the first probability density distribution curve corresponding to the binary classifier.

[0099] The second implementation method can be as follows: Figure 6 Perform the steps shown.

[0100] See Figure 6 The diagram illustrates a detailed process for constructing a first probability density distribution curve according to an embodiment of this application, specifically including the following steps S541 to S545:

[0101] S541, obtain the interval formed by the first reference feature value and the second reference feature value corresponding to the binary classifier, as the feature value interval.

[0102] Assuming the first reference feature value of the binary classifier is 100 and the second reference feature value is 0, then the feature value range is [0, 100].

[0103] See also Figure 6 S542 divides the eigenvalue interval into multiple eigenvalue sub-intervals.

[0104] In this embodiment, the feature value interval can be divided equally with a step size of 0.05.

[0105] It's important to note that dividing the feature value interval equally is necessary for statistical binning of the first result data from the binary classifier. Since the probability density function itself may not have an analytical solution that can be regressed, statistical binning of the entire data is required. Independent probability statistics are performed within each bin, and the probability distribution within each bin can be approximated using a uniform distribution or other distributions. This yields an approximate numerical solution that approximates the actual distribution. The smaller the bin intervals, the closer the overall distribution fits the data; as the intervals approach zero, the fitted curve closely resembles the actual data distribution.

[0106] See also Figure 6 S543, determine the number of feature values ​​in each feature value sub-interval in the first result data, and obtain multiple first quantities that correspond one-to-one with multiple feature value sub-intervals.

[0107] It is understandable that since the determined first result data includes multiple feature values, each feature value will fall within a different feature value sub-interval. Therefore, by counting the feature values ​​that fall within each feature value sub-interval, the corresponding first quantity can be obtained.

[0108] For example, if the number of feature values ​​falling into the [0,0.5] feature value sub-interval in the first result data is 5, then the first quantity corresponding to the [0,0.5] feature value sub-interval is 5.

[0109] See also Figure 6 S544, curve fitting is performed on the relationship between multiple feature value sub-intervals and multiple first quantities to obtain the first reference curve.

[0110] In this embodiment, the correspondence between multiple feature value sub-intervals and multiple first quantities can be represented in the form of a bar chart.

[0111] For example, a first probability density distribution curve is constructed for binary classifier 1 in the target recognition model. After equalizing and statistically processing the first result data of binary classifier 1, the following curve is formed: Figure 7 The data shown is composed of bars in light gray, where... Figure 7 In the diagram, the horizontal axis represents the eigenvalues, with each bar corresponding to a sub-interval of eigenvalues, and the vertical axis represents the quantity. Of course, the vertical axis corresponding to the bars in the light gray area represents the first quantity.

[0112] Furthermore, through the analysis of Figure 7 Curve fitting of the light gray histogram yields the first reference curve corresponding to binary classifier 1, i.e. Figure 7 Curve 1 is shown in the figure.

[0113] See also Figure 6 S545, normalize the first reference curve to obtain the first probability density distribution curve corresponding to the binary classifier.

[0114] It should be noted that the normalization of the first reference curve is to ensure that the curve conforms to the distribution law of the probability distribution curve.

[0115] For example, for Figure 7 Normalization of curve 1 can be achieved by taking the point corresponding to the maximum ordinate value on curve 1 as the center point and reprocessing the ordinate of the center point to 1. Furthermore, the points distributed on both sides of the center point on curve 1 can be processed accordingly, so that the probability determined when the feature value is the x-coordinate corresponding to the center point is 1.

[0116] It is understandable that since the ordinate value corresponding to the center point is the largest, it means that the number of feature values ​​falling into the feature value sub-interval corresponding to the center point in the first result data of binary classifier 1 is the largest. If the feature value sub-interval corresponding to the center point is defined as the target feature value sub-interval, then when the feature value output by binary classifier 1 for the test information falls into the target feature value sub-interval, it means that the true category of the test information is consistent with the category corresponding to binary classifier 1.

[0117] In summary, through the above steps S541 to S545, the first probability density distribution curve corresponding to each binary classifier in the target recognition model can be determined. The first probability density distribution curve records the correspondence between feature values ​​and probabilities.

[0118] In step S540 above, a second probability density distribution curve corresponding to the binary classifier is determined based on the second result data. Specific implementation methods include at least the following two:

[0119] In the first implementation, curve fitting is performed on the second result data, and the curve obtained by curve fitting is normalized to obtain the second probability density distribution curve corresponding to the binary classifier.

[0120] The second implementation method can be performed according to the following steps: determine the number of feature values ​​in each feature value sub-interval in the second result data to obtain multiple second quantities that correspond one-to-one with multiple feature value sub-intervals; perform curve fitting on the relationship between multiple feature value sub-intervals and multiple second quantities to obtain a second reference curve; and normalize the second reference curve to obtain a second probability density distribution curve corresponding to the binary classifier.

[0121] Specifically, the execution logic of this embodiment is the same as that of steps S543 to S545 above, and will not be repeated here.

[0122] For example, a second probability density distribution curve corresponding to binary classifier 1 in the target recognition model is constructed. After equalizing and statistically processing the second result data of binary classifier 1, the following curve is formed: Figure 7 The data shown is composed of bars in dark gray, where... Figure 7 In the diagram, the horizontal axis represents the eigenvalues, with each bar corresponding to a sub-interval of eigenvalues. The vertical axis represents the quantity; of course, the vertical axis corresponding to the bars in the dark gray area represents the second quantity.

[0123] Furthermore, through the analysis of Figure 7 Curve fitting of the histogram in the medium-dark gray area yields a second reference curve corresponding to binary classifier 1, i.e. Figure 7 Curve 2 is shown in the figure.

[0124] Understandably, the final determined second probability density distribution curve records the correspondence between eigenvalues ​​and probabilities.

[0125] In summary, in some embodiments of this application, based on test data and the feature values ​​output by each binary classifier in the target recognition model for the test data, a corresponding first probability density distribution curve and a second probability density distribution curve can be constructed for each binary classifier.

[0126] Furthermore, in S130 above, after each binary classifier of the target recognition model outputs a target feature value for the information to be recognized, the probability corresponding to the target feature value output by the binary classifier can be determined on the first probability density distribution curve corresponding to the binary classifier, as a first reference probability. And the probability corresponding to the target feature value output by the binary classifier can be determined on the second probability density distribution curve corresponding to the binary classifier, as a second reference probability.

[0127] Understandably, the determined first reference probability reflects the likelihood that the information to be identified belongs to the category corresponding to the binary classifier. The higher the first reference probability, the greater the likelihood that the information to be identified belongs to the category corresponding to the binary classifier.

[0128] The determined second reference probability reflects the degree of likelihood that the information to be identified does not belong to the category corresponding to the binary classifier. The higher the second reference probability, the greater the likelihood that the information to be identified does not belong to the category corresponding to the binary classifier.

[0129] See also Figure 1 S140, Based on the first reference probability and the second reference probability, determine the target probability that the information to be identified belongs to the category corresponding to the binary classifier.

[0130] In this embodiment, the specific implementation methods include at least the following three:

[0131] The first implementation method is to directly use the determined first reference probability as the target probability that the information to be identified belongs to the category corresponding to the binary classifier.

[0132] The second implementation method: If the difference between the sum of the first reference probability and the second reference probability and 1 is less than a preset threshold, then the first reference probability is taken as the target probability that the information to be identified belongs to the category corresponding to the binary classifier. The preset threshold can be set to a value that is close to 0, such as 0.01 or 0.02.

[0133] In the third implementation, since the first probability density distribution curve and the second probability density distribution curve corresponding to the binary classifier are not normalized during calculation, it cannot be guaranteed that the sum of the determined first reference probability and the second reference probability is 1. Therefore, in order to improve the accuracy of the finally determined target probability, the target probability can be calculated by the following formula (1):

[0134]

[0135] Where P0 represents the target probability, P1 represents the first reference probability, P2 represents the second reference probability, a is a first set parameter greater than or equal to 0 and less than or equal to 1, and n is a second set parameter greater than a.

[0136] It is understood that the above formula (1) is a harmonic formula, where 0≤a≤1. Specifically, the value can be set by the user, and this application does not limit it here. In addition, the second setting parameter can be 10, 20, etc., and the specific value can be set by the user, and this application does not limit it here.

[0137] As can be seen from the above formula (1), by harmonizing the target probability based on the first reference probability and the second reference probability, the accuracy of the determined target probability can be improved, thereby improving the recognition accuracy of the information to be identified.

[0138] See also Figure 1 S150, determine the recognition result of the information to be recognized based on the target probability of each of the multiple binary classifiers.

[0139] It is understandable that through the processing of steps S110 to S140 above, the target probability of the information to be identified belonging to each of the multiple categories can be determined. For example, if there are 10 categories, then through the target recognition model calculation and processing, and through the corresponding 10 first probability density distribution curves and 10 second probability density distribution curves, the target probability of the information to be identified in each of the 10 categories can be obtained.

[0140] In this embodiment, the specific implementation method is related to the type of target recognition model. If the target recognition model is an acoustic model for speech recognition, then the probabilities of each target can be input into the constructed language model, and then the text information corresponding to the information to be recognized can be obtained through the data processing of the language model.

[0141] If the target recognition model is used for image recognition, after obtaining the target probabilities corresponding to each binary classifier, the binary classifier corresponding to the maximum value among the target probabilities can be determined, and the category corresponding to the binary classifier can be taken as the true category of the information to be recognized.

[0142] In some embodiments of this application, when identifying information to be identified, the target probability of the information to be identified belonging to each of the multiple categories is determined by a target recognition model including multiple binary classifiers and the probability density distribution curve corresponding to each binary classifier. This abandons the ∑P in the traditional recognition model. i =1 constraint, where P i The posterior probability of the information to be identified belonging to the i-th category among multiple categories is used to predict the probability of the information to be identified belonging to each category among multiple categories by using the probability density distribution curve corresponding to the binary classifier. In this way, the independent conditional probability is transformed into a non-independent distribution, and the hard boundary problem is transformed into a soft boundary problem. For categories with similarity, there is a problem of overlapping features. Since the identification method of this application performs independent calculations for each category, the calculation results will not conflict due to similarity, thereby improving the identification accuracy of the information to be identified.

[0143] Furthermore, in traditional recognition models, output modeling requires modeling the entire set, which makes them susceptible to uneven data distribution. Once they encounter out-of-set issues, they will inevitably output incorrect results. However, in the technical solution of this application, the problem of the entire set is transformed into an independent problem of whether the information to be identified belongs to each category. In the technical solution of this application, even for untrained data, the probability of each target can be calculated based on the similarity with the features of each dimension of the existing category, making it more robust to out-of-set data.

[0144] Based on the same inventive concept, embodiments of this application provide an identification device that can be used to execute the identification method in the above embodiments of this application. For details not disclosed in the embodiments of this application, please refer to the embodiments of the identification method described above.

[0145] See Figure 8 The diagram shows a block diagram of an identification device according to an embodiment of this application.

[0146] like Figure 8 As shown, an identification device 800 according to an embodiment of this application includes: an acquisition unit 801, a first determination unit 802, a second determination unit 803, and a third determination unit 804.

[0147] The system includes an acquisition unit 801 for acquiring information to be identified; a processing unit 802 for processing the information to be identified using a trained target recognition model to obtain multiple target feature values ​​of the information to be identified, wherein the target recognition model includes multiple binary classifiers corresponding one-to-one with multiple categories, each binary classifier outputting a target feature value of the information to be identified, and the target feature value characterizing the degree of consistency between the true category of the information to be identified and the category corresponding to the binary classifier; and a first determination unit 803 for determining, for each binary classifier, the target feature value corresponding to the output of the binary classifier on a preset first probability density distribution curve corresponding to the binary classifier. The probability corresponding to the target feature value is used as a first reference probability that the information to be identified belongs to the category corresponding to the binary classifier, and the probability corresponding to the target feature value output by the binary classifier is determined on the preset second probability density distribution curve corresponding to the binary classifier, which is used as a second reference probability that the information to be identified does not belong to the category corresponding to the binary classifier; the second determining unit 804 is used to determine the target probability that the information to be identified belongs to the category corresponding to the binary classifier based on the first reference probability and the second reference probability; the third determining unit 804 is used to determine the identification result of the information to be identified based on the target probability of each of the multiple binary classifiers.

[0148] In some embodiments of this application, based on the foregoing scheme, the processing unit 802 is further configured to: acquire training data, the training data including various sample information; and train a pre-constructed first recognition model based on the training data to obtain a target recognition model, wherein the first recognition model includes a first output layer, the first output layer including multiple initial binary classifiers corresponding one-to-one with various categories, and each initial binary classifier is used to output the feature value of the sample information.

[0149] In some embodiments of this application, based on the foregoing scheme, before training the pre-built first recognition model, the processing unit 802 is further configured to: set a matching first reference feature value and second reference feature value for each initial binary classifier, so that the feature value output by the initial binary classifier is located within the interval formed by the first reference feature value and the second reference feature value, wherein the first reference feature value is the feature value that the initial binary classifier should output when the true category of the sample information is consistent with the category corresponding to the initial binary classifier, and the second reference feature value is the feature value that the initial binary classifier should output when the true category of the sample information is inconsistent with the category corresponding to the initial binary classifier.

[0150] In some embodiments of this application, based on the foregoing scheme, the first recognition model further includes a second output layer parallel to the first output layer. The second output layer includes a multi-classifier, which outputs the posterior probability of sample information belonging to each of the multiple categories. The processing unit 802 is further configured to: input training data into the first recognition model to obtain a first output result output by the first output layer and a second output result output by the second output layer; determine the first loss data of the first recognition model using a probabilistic loss function based on the first output result, and determine the second loss data of the first recognition model using a distributed loss function based on the second output result; determine the target loss data of the first recognition model based on the first loss data and the second loss data; update the model parameters of the first recognition model in reverse based on the target loss data to obtain a new first recognition model; return to the step of obtaining training data until the first recognition model converges to obtain the second recognition model; and remove or freeze the first output layer in the second recognition model to obtain the target recognition model.

[0151] In some embodiments of this application, based on the foregoing scheme, the first determining unit 803 is further configured to: acquire test data, the test data including various test information; input the test data into a target recognition model to obtain multiple sets of test results corresponding one-to-one with multiple binary classifiers, the test results including feature values ​​of various test information output by the binary classifiers; for each binary classifier, extract first result data and second result data from the test results corresponding to the binary classifier, the first result data including the feature value output by the binary classifier when the true category of the test information is consistent with the category corresponding to the binary classifier, the second result data including the feature value output by the binary classifier when the true category of the test information is inconsistent with the category corresponding to the binary classifier; determine a first probability density distribution curve corresponding to the binary classifier based on the first result data, and determine a second probability density distribution curve corresponding to the binary classifier based on the second result data.

[0152] In some embodiments of this application, based on the foregoing scheme, the first determining unit 803 is further configured to: obtain an interval consisting of the first reference feature value and the second reference feature value corresponding to the binary classifier, as a feature value interval; divide the feature value interval into multiple feature value sub-intervals; determine the number of feature values ​​in the first result data located in each feature value sub-interval, and obtain multiple first quantities corresponding one-to-one with the multiple feature value sub-intervals; perform curve fitting on the relationship between the multiple feature value sub-intervals and the multiple first quantities to obtain a first reference curve; and perform normalization processing on the first reference curve to obtain a first probability density distribution curve corresponding to the binary classifier.

[0153] In some embodiments of this application, based on the aforementioned scheme, the first determining unit 803 is further configured to: determine the number of feature values ​​located in each feature value sub-interval in the second result data, and obtain a plurality of second quantities corresponding one-to-one with the plurality of feature value sub-intervals; perform curve fitting on the relationship between the plurality of feature value sub-intervals and the plurality of second quantities to obtain a second reference curve; and perform normalization processing on the second reference curve to obtain a second probability density distribution curve corresponding to the binary classifier.

[0154] In some embodiments of this application, based on the foregoing scheme, the second determining unit 804 is further configured to: calculate the target probability using the following formula:

[0155]

[0156] Where P0 represents the target probability, P1 represents the first reference probability, P2 represents the second reference probability, a is a first set parameter greater than or equal to 0 and less than or equal to 1, and n is a second set parameter greater than a.

[0157] The above description is merely an embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A recognition method, characterized in that, The method includes: Obtain the information to be identified; The target recognition model is used to process the information to be recognized to obtain multiple target feature values ​​of the information to be recognized. The target recognition model includes an acoustic model for speech recognition and multiple binary classifiers that correspond one-to-one with multiple categories. Each binary classifier is used to output the target feature value of the information to be recognized. The target feature value is used to characterize the degree of consistency between the true category of the information to be recognized and the category corresponding to the binary classifier. For each binary classifier, on a pre-defined first probability density distribution curve corresponding to that binary classifier, the probability corresponding to the target feature value output by that binary classifier is determined as the first reference probability that the information to be identified belongs to the category corresponding to that binary classifier. On the preset second probability density distribution curve corresponding to the binary classifier, the probability corresponding to the target feature value output by the binary classifier is determined as the second reference probability that the information to be identified does not belong to the category corresponding to the binary classifier. Based on the first reference probability and the second reference probability, determine the target probability that the information to be identified belongs to the category corresponding to the binary classifier; The recognition result of the information to be identified is determined based on the target probability of each of the plurality of binary classifiers.

2. The method according to claim 1, characterized in that, The method further includes: Acquire training data, which includes various sample information; Based on the training data, a pre-constructed first recognition model is trained to obtain the target recognition model. The first recognition model includes a first output layer, which includes multiple initial binary classifiers that correspond one-to-one with the multiple categories. Each initial binary classifier is used to output the feature value of the sample information.

3. The method according to claim 2, characterized in that, Before training the pre-built first recognition model, the method further includes: For each initial binary classifier, a matching first reference feature value and a second reference feature value are set so that the feature value output by the initial binary classifier lies within the interval formed by the first reference feature value and the second reference feature value, wherein, The first reference feature value is the feature value that the initial binary classifier should output when the true category of the sample information is consistent with the category corresponding to the initial binary classifier. The second reference feature value is the feature value that the initial binary classifier should output when the true category of the sample information is inconsistent with the category corresponding to the initial binary classifier.

4. The method according to claim 2, characterized in that, The first recognition model further includes a second output layer parallel to the first output layer. The second output layer includes a multi-classifier, which outputs the posterior probability of a sample belonging to each of the multiple categories. The pre-built first recognition model is trained to obtain the target recognition model, which includes: The training data is input into the first recognition model to obtain a first output result output by the first output layer and a second output result output by the second output layer. Based on the first output result, the first loss data of the first recognition model is determined by a probabilistic loss function, and based on the second output result, the second loss data of the first recognition model is determined by a distributed loss function. Based on the first loss data and the second loss data, the target loss data of the first recognition model is determined; Based on the target loss data, the model parameters of the first recognition model are updated in reverse to obtain a new first recognition model; Return to the step of obtaining training data until the first recognition model converges, and obtain the second recognition model; Remove or freeze the first output layer in the second recognition model to obtain the target recognition model.

5. The method according to claim 1, characterized in that, The method further includes: Acquire test data, which includes various types of test information; The test data is input into the target recognition model to obtain multiple sets of test results that correspond one-to-one with the multiple binary classifiers. The test results include feature values ​​of various test information output by the binary classifiers. For each binary classifier, first result data and second result data are extracted from the test results corresponding to the binary classifier. The first result data includes the feature value output by the binary classifier when the true category of the test information is consistent with the category corresponding to the binary classifier. The second result data includes the feature value output by the binary classifier when the true category of the test information is inconsistent with the category corresponding to the binary classifier. Based on the first result data, a first probability density distribution curve corresponding to the binary classifier is determined, and based on the second result data, a second probability density distribution curve corresponding to the binary classifier is determined.

6. The method according to claim 5, characterized in that, The step of determining the first probability density distribution curve corresponding to the binary classifier based on the first result data includes: Obtain the interval formed by the first reference feature value and the second reference feature value corresponding to the binary classifier, and use it as the feature value interval; The eigenvalue interval is divided into multiple eigenvalue sub-intervals; Determine the number of feature values ​​in the first result data that are located in each of the feature value sub-intervals, and obtain a plurality of first quantities that correspond one-to-one with the plurality of feature value sub-intervals; A first reference curve is obtained by performing curve fitting on the relationship between the plurality of feature value sub-intervals and the plurality of first quantities; The first reference curve is normalized to obtain the first probability density distribution curve corresponding to the binary classifier.

7. The method according to claim 6, characterized in that, The step of determining the second probability density distribution curve corresponding to the binary classifier based on the second result data includes: Determine the number of feature values ​​located in each feature value sub-interval in the second result data to obtain a plurality of second quantities that correspond one-to-one with the plurality of feature value sub-intervals; A second reference curve is obtained by performing curve fitting on the relationship between the plurality of feature value sub-intervals and the plurality of second quantities; The second reference curve is normalized to obtain the second probability density distribution curve corresponding to the binary classifier.

8. The method according to claim 1, characterized in that, The step of determining the target probability that the information to be identified belongs to the category corresponding to the binary classifier based on the first reference probability and the second reference probability includes: The target probability is calculated using the following formula: in, This represents the target probability. This represents the first reference probability. This represents the second reference probability. The first set parameter is greater than or equal to 0 and less than or equal to 1. greater than The second setting parameter.

9. An identification device, characterized in that, The device includes: The acquisition unit is used to acquire the information to be identified; The processing unit is used to process the information to be identified using a trained target recognition model to obtain multiple target feature values ​​of the information to be identified. The target recognition model includes an acoustic model for speech recognition and multiple binary classifiers that correspond one-to-one with multiple categories. Each binary classifier is used to output the target feature value of the information to be identified. The target feature value is used to characterize the degree of consistency between the true category of the information to be identified and the category corresponding to the binary classifier. The first determining unit is configured to, for each binary classifier, determine, on a preset first probability density distribution curve corresponding to that binary classifier, the probability corresponding to the target feature value output by that binary classifier, as a first reference probability that the information to be identified belongs to the category corresponding to that binary classifier, and On the preset second probability density distribution curve corresponding to the binary classifier, the probability corresponding to the target feature value output by the binary classifier is determined as the second reference probability that the information to be identified does not belong to the category corresponding to the binary classifier. The second determining unit is used to determine the target probability that the information to be identified belongs to the category corresponding to the binary classifier based on the first reference probability and the second reference probability. The third determining unit is used to determine the recognition result of the information to be identified based on the target probability of each of the plurality of binary classifiers.

Images