WO2021121127A1 - 样本类别识别方法、装置、计算机设备及存储介质 - Google Patents

样本类别识别方法、装置、计算机设备及存储介质 Download PDF

Info

Publication number
WO2021121127A1
WO2021121127A1 PCT/CN2020/135337 CN2020135337W WO2021121127A1 WO 2021121127 A1 WO2021121127 A1 WO 2021121127A1 CN 2020135337 W CN2020135337 W CN 2020135337W WO 2021121127 A1 WO2021121127 A1 WO 2021121127A1
Authority
WO
WIPO (PCT)
Prior art keywords
sample
result
training
category
loss value
Prior art date
Application number
PCT/CN2020/135337
Other languages
English (en)
French (fr)
Inventor
吴海萍
陶蓉
徐尚良
张芮溟
周鑫
Original Assignee
平安科技（深圳）有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 平安科技（深圳）有限公司 filed Critical 平安科技（深圳）有限公司
Publication of WO2021121127A1 publication Critical patent/WO2021121127A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G06F18/24147—Distances to closest patterns, e.g. nearest neighbour classification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/30—Semantic analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/088—Non-supervised learning, e.g. competitive learning

Definitions

  • This application relates to the field of artificial intelligence image classification, and in particular to a sample category recognition method, device, computer equipment, and storage medium.
  • the image is used as a sample for training, because these abnormal types or new types are manually identified, and the probability of occurrence is extremely low (due to the need to rely on destructive experiments or costly to obtain, etc.), As a result, it is impossible to train a model that can accurately identify the above types, and it also requires a high labor cost to identify and classify. Therefore, the zero-sample identification problem is extremely important, and this problem has attracted more and more attention from the industry.
  • the images of normal parts and broken parts are easy to collect, while the images of parts with abnormal conditions such as slight cracks, slight depressions, and internal fractures are difficult to collect, and the recognition results are not good; in the identification of vehicle damage, Scratched, scratched, torn, and dented vehicle damage photos are easy to collect, while vehicle damage photos with serious damage such as wrinkles, dead-folds, missing, etc. are difficult to collect, and damage identification is not good; in medical imaging inspections, lungs Common X-rays such as normal, pneumonia, and pulmonary hydrops are easy to collect, but rare X-rays such as early tuberculosis and early lung cancer are difficult to collect, and lung medical tests are not good.
  • This application provides a sample category recognition method, device, computer equipment and storage medium, which realize the automatic recognition of the sample to be identified through the first sample detection model based on end-to-end and the second sample detection model based on zero-sample learning.
  • Sample category this application is suitable for areas such as smart transportation or smart medical care, which can further promote the construction of smart cities, thereby reducing the cost of manual identification and improving the accuracy of identification.
  • a method for identifying sample categories including:
  • the to-be-identified sample into the first sample detection model to perform sample feature extraction and semantic feature recognition to obtain the to-be-processed sample recognition result;
  • the to-be-processed sample recognition result includes the first sample detection result, the sample feature space result and the sample Semantic space result;
  • the sample category of the sample to be identified is determined and output.
  • a sample type recognition device including:
  • the acquisition module is used to acquire the sample to be identified
  • the identification module is used to input the sample to be identified into the first sample detection model for sample feature extraction and semantic feature recognition to obtain the identification result of the sample to be processed;
  • the identification result of the sample to be processed includes the first sample detection result, the sample Feature space results and sample semantic space results;
  • a matching module configured to input the sample feature space result and the sample semantic space result into the second sample detection model when the first sample detection result does not match all the first sample categories;
  • An abnormality module configured to perform clustering processing on the sample feature space results through the second sample detection model to obtain a first abnormal result, and simultaneously perform similarity matching on the sample semantic space results to obtain a second abnormal result;
  • the output module is used to determine and output the sample category of the sample to be identified according to the first abnormal result and the second abnormal result.
  • a computer device includes a memory, a processor, and computer-readable instructions that are stored in the memory and can run on the processor, and the processor implements the following steps when the processor executes the computer-readable instructions:
  • the to-be-identified sample into the first sample detection model to perform sample feature extraction and semantic feature recognition to obtain the to-be-processed sample recognition result;
  • the to-be-processed sample recognition result includes the first sample detection result, the sample feature space result, and the sample Semantic space result;
  • the sample category of the sample to be identified is determined and output.
  • One or more readable storage media storing computer readable instructions, when the computer readable instructions are executed by one or more processors, the one or more processors execute the following steps:
  • the to-be-identified sample into the first sample detection model to perform sample feature extraction and semantic feature recognition to obtain the to-be-processed sample recognition result;
  • the to-be-processed sample recognition result includes the first sample detection result, the sample feature space result, and the sample Semantic space result;
  • the sample category of the sample to be identified is determined and output.
  • the sample category identification method, device, computer equipment and storage medium provided in this application obtain the sample to be identified; input the sample to be identified into the first sample detection model for sample feature extraction and semantic feature identification, and obtain the sample identification to be processed Result; the sample recognition result to be processed includes the first sample detection result, the sample feature space result, and the sample semantic space result; when the first sample detection result does not match all the first sample categories, the The sample feature space results and the sample semantic space results are input into a second sample detection model; the sample feature space results are clustered through the second sample detection model to obtain the first abnormal result, and at the same time The sample semantic space result is similarly matched to obtain a second abnormal result; according to the first abnormal result and the second abnormal result, the sample category of the sample to be identified is determined and output, so that the first pass is achieved
  • the end-to-end first sample detection model recognizes the sample to be identified.
  • the sample category of the sample to be identified can be identified through the second sample detection model, so that it can automatically perform accurate sample category recognition on samples that are difficult to collect or new sample categories, reducing manual labor
  • the cost of recognition is automatically marked with sample categories for samples that are difficult to collect or new sample categories, which saves labor costs and improves recognition accuracy.
  • FIG. 1 is a schematic diagram of an application environment of a sample category identification method in an embodiment of the present application
  • Fig. 2 is a flowchart of a sample category identification method in an embodiment of the present application
  • Fig. 3 is a flowchart of a sample category identification method in another embodiment of the present application.
  • FIG. 4 is a flowchart of step S20 of a method for identifying a sample category in an embodiment of the present application
  • FIG. 5 is a flowchart of step S206 of the sample category identification method in an embodiment of the present application.
  • FIG. 6 is a flowchart of step S2063 of the sample category identification method in an embodiment of the present application.
  • FIG. 7 is a flowchart of step S40 of a method for identifying a sample category in an embodiment of the present application.
  • Fig. 8 is a functional block diagram of a sample category identification device in an embodiment of the present application.
  • Fig. 9 is a schematic diagram of a computer device in an embodiment of the present application.
  • the sample category identification method provided by this application can be applied in the application environment as shown in Fig. 1, in which the client (computer equipment) communicates with the server through the network.
  • the client includes, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers, cameras, and portable wearable devices.
  • the server can be implemented as an independent server or a server cluster composed of multiple servers.
  • a sample category identification method is provided, and the technical solution mainly includes the following steps S10-S50:
  • the sample recognition request is triggered in the application scenario of recognizing the sample category of the collected samples, and the sample to be recognized in the sample recognition request is obtained.
  • the application scenario can be set according to requirements, for example, in an automobile part Recognize the collected part sample images in the quality inspection, or recognize the samples used to train the damage recognition model on the vehicle damage assessment, or recognize the samples used to train the lung recognition model on the medical image inspection And so on, the sample to be identified is a sample image or photo file that needs to be identified.
  • S20 Input the to-be-identified sample into the first sample detection model to perform sample feature extraction and semantic feature recognition to obtain the to-be-processed sample recognition result; the to-be-processed sample recognition result includes the first sample detection result and the sample feature space result And sample semantic space results.
  • the sample to be identified is input into the first sample detection model, and the first sample detection model is identified by a plurality of branches trained by training samples of a plurality of first sample categories.
  • the first sample category is a conventional, easy-to-collect, and large number of sample categories.
  • the first sample category is vehicle damage, scratches, scratches, tears, etc.
  • the network structure of the first sample detection model can be set according to requirements.
  • the network structure can be the network structure of Resnet50, the network structure of CNN, the network structure of VGG, etc., preferably, the first is the same
  • the network structure of this detection model is the network structure of Resnet50
  • the sample features are features related to the recognition of the first sample category in multiple branch dimensions
  • the sample features include high-dimensional dominant features and low-dimensional recessive features.
  • the sample features can be set according to requirements.
  • high-dimensional explicit features include deformation features, color difference features, etc.
  • low-latitude recessive features include smoothness features, etc.
  • the first sample detection model performs multi-branch task feature convolution and regularization processing on the sample to be identified, and the sample feature vector set is obtained, and the sample feature vector is a feature that reflects one of the sample features and An array measured by feature vectors.
  • the dimension of the sample feature vector can be set according to requirements. For example, the dimension of the sample feature vector can be 256 dimensions.
  • a branch task corresponds to one of the sample features and also corresponds to A sample feature vector
  • the sample feature space result can reflect each feature of the sample feature in multiple dimensions
  • the semantic feature recognition further includes the final mapping function embedded in the space model after input converged
  • the final mapping function outputs the semantic feature vector corresponding to its input.
  • the final mapping function is the converged mapping function
  • the converged embedding space model is output based on the sample feature space result. All the semantic feature vectors are recorded as the result of the sample semantic space.
  • the sample to be identified is identified through the first sample detection model, and the identified probability value corresponding to each of the first sample categories is obtained, and the probabilities corresponding to each first sample category are obtained.
  • the largest value among all the probability values is selected from the values, and if the largest value among all the probability values is greater than or equal to the preset probability threshold, then the first sample corresponding to the largest value among all the probability values
  • the category is determined as the detection result of the first sample corresponding to the sample to be identified; if the largest value among all the probability values is less than the probability threshold, the first sample corresponding to the sample to be identified
  • the test result of the sample is confirmed to be empty or an abnormal type.
  • Also includes:
  • the matching method can be set according to requirements, for example, the matching method is the same as the first sample category.
  • the content of is completely consistent, or when the content similarity with the first sample category reaches a preset category probability, it is determined as a match, etc., and the matched first sample category is recorded as the sample category of the sample to be identified.
  • the sample category of the sample to be identified is determined, so .
  • step S20 before the step S20, that is, before the input of the sample to be identified into the first sample detection model for sample feature extraction and semantic feature recognition, it includes:
  • S201 Obtain a training sample set; the training sample set includes a plurality of training samples, and one training sample is associated with a first sample category and a first sample description.
  • the training sample set is a collection of the training samples
  • the training samples are images of various first sample categories that are easy to collect
  • one training sample is associated with one first sample category
  • it is associated with a description of the first sample, which is a text description of the first sample category in the training sample, such as a photo of a vehicle damage with a scratched car body as Training sample
  • the associated first sample category is "scratch”
  • the associated first sample is described as "the middle part of the car body is scratched out of a damaged area with a length of 7 cm and a width of 2 cm without obvious deformation and depression.”
  • S202 Input the training samples into a multi-branch convolutional neural network model containing initial parameters.
  • the training samples are input to the multi-branch convolutional neural network model
  • the multi-branch convolutional neural network model includes the initial parameters
  • the initial parameters include the parameters of the network structure and the dimensions of the feature vector Parameters and so on.
  • step S202 that is, before inputting the training samples into a multi-branch convolutional neural network model containing initial parameters, the method includes:
  • the transfer learning is to use the parameters of existing training models in other fields to be applied to tasks in this field, that is, the multi-branch convolutional neural network model obtains data through transfer learning. All model parameters of an unsupervised pre-training model.
  • the unsupervised pre-training model can be a multi-branch convolutional neural network model of unsupervised learning in the natural field or the vehicle recognition field, and then all the model parameters Determined as the initial parameters of the multi-branch convolutional neural network model.
  • the present application obtains initial parameters from the trained unsupervised domain training model through migration learning, which can shorten the number of iterations of the model, simplify the training process, and improve training efficiency.
  • S203 Perform sample feature extraction on the training sample through the multi-branch convolutional neural network model to obtain an image feature space result.
  • the sample features are features related to sample identification in multiple branch dimensions
  • the sample features include high-dimensional explicit features and low-dimensional recessive features
  • the sample features can be set according to requirements, such as In car damage recognition, high-dimensional explicit features include deformation features, chromatic aberration features, etc., low-latitude recessive features include smoothness features, etc.
  • the result of the image feature space is the training of the multi-branch convolutional neural network model.
  • An image feature vector obtained after the sample is subjected to feature convolution and regularization processing of a multi-branch task.
  • the image feature vector is an array that reflects one of the features of the sample and is measured by the feature vector.
  • the image feature vector The dimensions of can be set according to requirements.
  • the dimension of the image feature vector can be 256 dimensions, where a branch task corresponds to one of the sample features and also corresponds to one of the image feature vectors.
  • the image feature space results Each of the characteristics of the sample can be embodied in multiple dimensions.
  • S203 that is, performing sample feature extraction on the training sample through the multi-branch convolutional neural network model to obtain an image feature space result, includes:
  • S2031 Perform image preprocessing on the training samples to obtain preprocessed sample images
  • the preprocessing is an operation process of recognizing feature regions of the training samples, extracting images of a preset size, and performing enhancement processing on the extracted images, and identifying regions with sample features in the training samples , It can be recognized by the YOLO (You Only Look Once) algorithm, and the feature area containing the sample characteristics is identified and then the image of the preset size is extracted.
  • the preset size can be set according to the requirements. As a preference, the preset size is 224 ⁇ 224
  • the size of the extracted image is enhanced, and the enhanced processing can be set according to requirements, for example, the enhanced processing is denoising and sharpening processing, so as to obtain the preprocessed sample image.
  • S2032 Perform feature extraction on the preprocessed sample image to obtain at least one feature vector image
  • feature extraction is performed on each channel of the preprocessed sample image, and the feature extraction is performed by convolving each channel through different convolution kernels corresponding to different features to output a feature vector image.
  • the convolution process is based on The network structure of the multi-branch convolutional neural network model is determined, for example, the feature extraction of high-dimensional features is performed on each channel of the preprocessed sample image to obtain a feature vector map corresponding to the high-dimensional feature, and the feature vector map is composed of multiple The eigenvector value composition.
  • S2033 Perform regularization processing on each of the feature vector graphs to obtain image feature vectors corresponding to each of the feature vector graphs;
  • the regularization processing is to perform N times root regularization nonlinear processing on the eigenvector values in each eigenvector diagram, and the eigenvector diagram can be fine-tuned through the regularization processing, which can better reflect each Feature, thereby outputting the image feature vector.
  • S2034 Determine all the image feature vectors as the result of the image feature space.
  • This application realizes that by performing image preprocessing on the training sample, a preprocessed sample image is obtained; feature extraction is performed on the preprocessed sample image to obtain at least one feature vector diagram; each feature vector diagram is regularized , Obtain the image feature vector corresponding to each of the feature vector diagrams; determine all the image feature vectors as the result of the image feature space, so that through image preprocessing, feature extraction and regularization processing, the accuracy of recognition can be improved Rate and reliability.
  • the embedded space model is constructed by constructing image feature vectors and semantic feature vectors The relationship between the obtained.
  • the embedded space model is obtained by learning the mapping relationship between the image feature vector corresponding to the training sample and the first sample description corresponding to the training sample converted into the semantic feature vector, that is, the first sample of the training sample is the same.
  • This description input is based on the semantic recognition model of Word2vec, and the semantic vector conversion is performed on the first sample description to obtain the semantic feature vector corresponding to the first sample description.
  • the semantic recognition model based on Word2vec is a trained deep neural network model
  • the semantic recognition model can transform the input text into semantic vector to generate the vector value of the semantic feature corresponding to each feature of the sample feature, that is, generate the semantic feature vector, for example, the description of the first sample is described as "the middle of the car body Part of the scratched area with a length of 7 cm and a width of 2 cm, without obvious deformation and dents, will be transformed into the vector value corresponding to the semantic feature related to the deformation and the vector corresponding to the semantic feature related to the color difference after the semantic vector conversion.
  • the multi-branch convolutional neural network model is continuously iterated to continuously Learning until the convergence of the multi-branch convolutional neural network model is reached.
  • the learning of the mapping function between the image feature vector and the semantic feature vector corresponding to the training sample is completed, that is, the value of the image feature vector input to the mapping function and the output value and the image
  • the mean square error between the semantic feature vectors corresponding to the feature vectors is minimized, that is, the embedding space model is converged, and the semantic feature recognition is converted by inputting the mapping function in the embedding space model to obtain the mapping
  • the semantic feature vector output by the function corresponding to its input, and all the semantic feature vectors output by the embedded space model according to the image feature space result are recorded as the semantic feature space result.
  • S205 Recognizing the image feature space result through the K-means clustering algorithm to obtain a training category result, and determining the training category result, the image feature space result, and the semantic feature space result as sample training result.
  • the K-means clustering algorithm is also called K-means clustering algorithm. It uses distance as an evaluation index of similarity, and determines its corresponding training category according to the distance of clusters close to each training category.
  • -means clustering algorithm uses distance as an evaluation index of similarity, and determines its corresponding training category according to the distance of clusters close to each training category.
  • -means clustering algorithm to calculate the Euclidean distance between each image feature vector in the image feature space result and the center of each training category result, and determine the corresponding training category result according to each Euclidean distance, the training category result contains The category of is exactly the same as the category contained in the first sample category, thereby achieving end-to-end recognition, achieving centralized extraction of features corresponding to the same category, and improving the recognition accuracy.
  • the training category result, the image feature space result, and the semantic feature space result corresponding to the training sample are marked as the sample training result corresponding to the training sample.
  • the first loss value is obtained by calculating the difference between the result of the image feature space and the center vector corresponding to the first sample category; the first sample description is input into Word2vec-based
  • the semantic vector conversion is performed on the first sample description to obtain the semantic feature vector corresponding to the first sample description, and the semantic feature vector is calculated by calculating the semantic feature space result and the semantic feature vector.
  • Semantic similarity value to obtain the second loss value, and the second loss value represents the difference between the semantic feature space result and the semantic feature vector.
  • step S206 that is, performing sample feature extraction on the training sample through the multi-branch convolutional neural network model to obtain an image feature space result includes:
  • S2061 Perform vector conversion on the first sample category corresponding to the training sample through the multi-branch convolutional neural network model to obtain a center vector corresponding to the first sample category; the center vector includes Euclidean domain center vector and angle domain center vector.
  • the first sample category is subjected to a vector conversion, and the vector conversion is to convert the first sample category of the text type according to a preset mapping relationship to obtain the first sample category
  • the matched center vector, the center vector includes the Euclidean domain center vector and the angle domain center vector.
  • S2062 Obtain a Euclidean loss value according to the result of the image feature space and the center vector of the Euclidean domain through a cross-entropy loss algorithm.
  • the loss value between each of the image feature vectors in the image feature space result and the center vector of the Euclidean domain in the image feature space result is calculated through the cross entropy loss function, and the loss value is recorded as the Euclidean loss value .
  • the loss value between each of the image feature vectors and the angle domain center vector in the image feature space result is calculated through the ArcFace loss function, and the loss value is recorded as the angle loss value.
  • the ArcFace loss algorithm is used to obtain an angle loss value according to the image feature space result and the angle domain center vector, including:
  • S20631 Perform regularization processing on the image feature space result through the regularization model in the multi-branch convolutional neural network model to obtain regularized feature vectors.
  • the multi-branch convolutional neural network model includes the regularization model, the regularization model includes a regularization function, each of the image feature vectors is input into the regularization function, and the regularization feature is output Vector, the regularized feature vector is:
  • B 1 is the first feature vector value in the image feature vector
  • b m is the m-th feature vector value in the image feature vector
  • b i is the i-th feature vector value in the image feature vector
  • N is the pre- The value of the nth root of suppose.
  • S20632 Input the regularized feature vector and the angle domain center vector into an angle domain loss model, and obtain an angle loss value through the ArcFace loss algorithm in the angle domain loss model.
  • the ArcFace loss algorithm is an algorithm that calculates its loss value through the ArcFace loss function.
  • the regularized feature vector and the angle domain center vector are compared through the cosine angle comparison method in the ArcFace loss function through the angle domain.
  • the direction of the measurement of the gap so as to obtain the value of the angle loss.
  • this application realizes that the angle loss value can be obtained through the regularization model and the angle domain loss model in the multi-branch convolutional neural network model. Therefore, the regularization can limit the divergence of the model and achieve fine-tuning in the recognition and classification process.
  • the role of the angle domain loss model can increase the compact distance between categories.
  • S2064 Perform weighting processing on the Euclidean loss value and the angle loss value to obtain the first loss value.
  • the first loss value output by the first loss function can not only consider the correct category classification, but also consider the compact distance between categories, so that The first loss function is very expressive in the classification results.
  • the present application realizes the vector conversion of the first sample category corresponding to the training sample through the multi-branch convolutional neural network model to obtain the center vector corresponding to the first sample category; Cross entropy loss algorithm to obtain the Euclidean loss value; through the ArcFace loss algorithm to obtain the angle loss value; weighting the Euclidean loss value and the angle loss value to obtain the first loss value, therefore, by introducing the angle loss The value makes the recognition move closer to a more compact recognition direction, making the recognition accuracy more accurate.
  • the first loss value and the second loss value are input to a total loss model containing a total loss function in the multi-branch convolutional neural network model, and the total loss function in the total loss model can be According to the demand setting, the loss model is a model for generating the total loss value, and the total loss value is calculated by the total loss function.
  • step S207 that is, obtaining a total loss value according to the first loss value and the second loss value, includes:
  • X1 is the first loss value
  • X2 is the second loss value
  • W 1 is the weight of the first loss value
  • W 2 is the weight of the second loss value.
  • the convergence condition may be a condition that the value of the total loss value is small and will not drop after 10,000 calculations, that is, the value of the total loss value is small and will not decrease after 10,000 calculations.
  • the convergence condition may also be that the total loss value is less than
  • the first sample detection model that has been trained can be stored in the blockchain.
  • the initial parameters of the iterative multi-branch convolutional neural network model are constantly updated, and the multi-branch convolutional neural network model is triggered to perform the
  • the step of extracting the sample feature of the training sample to obtain the result of the image feature space can continuously move closer to the accurate result, so that the accuracy of the recognition is getting higher and higher.
  • the first sample detection model may also be stored in a node of the blockchain.
  • Blockchain essentially a decentralized database
  • Each data block contains a batch of network transaction information for verification. The validity of the information (anti-counterfeiting) and the generation of the next block.
  • the blockchain can include the underlying platform of the blockchain, the platform product service layer, and the application service layer.
  • the decentralized and fully distributed DNS service provided by the blockchain can realize the query and resolution of domain names through the point-to-point data transmission service between various nodes in the network, which can be used to ensure that the operating system and firmware of an important infrastructure are not available.
  • step S207 that is, after obtaining a total loss value according to the first loss value and the second loss value, the method further includes:
  • the multi-branch convolutional neural network model Having converged, the multi-branch convolutional neural network model after convergence is recorded as the first sample detection model that has been trained.
  • the sample to be identified is recorded as the abnormal sample, and the abnormal sample Indicates that all results other than the first sample category are abnormal, and the sample feature space result and the sample semantic space result are input into the second sample detection model, and the second sample detection model is based on
  • the final mapping function and the K-nearest neighbor algorithm are an adaptive model that is learned by artificially marking the features of the known second sample category in each of the corresponding sample features, that is, establishing the The mapping relationship between the clustering range of each feature in the sample feature and the range of the semantic feature vector and the known second sample category, the second sample category is difficult to collect (a small amount) and known description (already known
  • the description related to the characteristics of the sample is also the second sample description in the full text) of the sample category and the newly discovered and known description of the sample category, such as vehicle damage recognition, based on the recognition of "missing" features (with The common features of "tear” and "sag"), within the preset
  • S40 Perform clustering processing on the sample feature space results through the second sample detection model to obtain a first abnormal result, and simultaneously perform similarity matching on the sample semantic space results to obtain a second abnormal result.
  • the method of clustering processing is to calculate the result of the sample feature space by the K-means clustering algorithm, and calculate the distance between the result of the sample feature space and the cluster class corresponding to each first sample category.
  • the Euclidean distance value of the center (centroid) after calculating each Euclidean distance value through the second sample detection model, the second sample category corresponding to the area range is determined according to the range of the area where each Euclidean distance value falls, and the It is determined as a clustering abnormal category, and each Euclidean distance value is converted into a probability value with each of the second sample categories, and the clustering abnormal category and the probability value with each of the second sample categories are recorded as all.
  • the first abnormal result indicates the result identified by the dimension of the sample feature space
  • the second sample detection model calculates the sample semantic space result and each first sample
  • the similarity value between the semantic feature vectors corresponding to the category is obtained, and the positive similarity value is obtained.
  • the algorithm for calculating the positive similarity value can be set according to requirements, such as the cosine similarity algorithm, the Jaccard similarity algorithm, etc., according to the positive
  • the similarity value determines the adjacent first sample category, and then maps the semantic similarity category according to the adjacent first sample category, and records the mapped semantic similarity category and each positive similarity value as the second An abnormal result, where the second abnormal result indicates the result identified by the dimension of the semantic space result.
  • step S40 that is, performing similarity matching on the sample semantic space results to obtain a second abnormal result includes:
  • S401 Calculate the similarity value between the sample semantic space result and the second sample description corresponding to each preset second sample category through the Word2vec model in the second sample detection model.
  • the second sample detection model can also use the Word2vec model to perform word vector conversion on the second sample description corresponding to each second sample category to obtain the semantics corresponding to each of the second sample categories.
  • Word vector the second sample category is a sample category that is difficult to collect (a small amount) and has a known description (the description related to the sample feature that is already known is also the second sample description in the full text) and the newly discovered sample category Knowing the described sample category, the similarity value between the sample semantic space result and each semantic word vector is calculated through the Word2vec algorithm in the Word2vec model.
  • the Word2vec model is a neural network model constructed based on the Word2vec algorithm, and a second sample description corresponding to each second sample category and related to each feature of the sample feature is preset in the Word2vec model.
  • the largest similarity value is obtained from the similarity values between all the sample semantic space results and each semantic word vector, and the corresponding abnormality type is determined as the second An abnormal result, where the second abnormal result includes the largest obtained similarity value.
  • this application realizes the calculation of the similarity value between the sample semantic space result and the second sample description corresponding to each preset second sample category through the Word2vec model in the second sample detection model; and then through The second sample category corresponding to the largest similarity value among all the similarity values is acquired, and the acquired second sample category is determined as the second abnormal result. Therefore, the algorithm based on Word2vec is implemented The second abnormal result is automatically recognized, which provides a semantic recognition method to make the recognition more accurate.
  • S50 According to the first abnormal result and the second abnormal result, determine and output the sample category of the sample to be identified.
  • the final second sample category is obtained.
  • the weight parameter is used to widen the gap between the probability values corresponding to each second sample category to make the classification more accurate. For example, the "missing" second sample category weighted by the weight parameter is in the first abnormal result and the second abnormal result. If the sum of the probability values is the largest, the "missing" second sample category is determined as the sample category of the sample to be identified.
  • the sample category includes the first sample category and the second sample category.
  • samples of the first sample category can be identified from the batch of samples.
  • samples of the second sample category based on the zero training samples of the second sample category. For example, in the collected car damage photo sample set, the first samples of scratches, scratches, tears, and dents can be identified.
  • the damaged photos of the sample category, and the damaged photos of the second sample category such as wrinkles, dead folds, and missing pieces can be identified (the photos of the second sample category such as folds, dead folds, and missing pieces have not been collected before this).
  • This application realizes that by obtaining the sample to be identified; inputting the sample to be identified into the first sample detection model to perform sample feature extraction and semantic feature recognition, to obtain the identification result of the sample to be processed; the identification result of the sample to be processed includes the first This detection result, the sample feature space result, and the sample semantic space result; when the first sample detection result does not match all the first sample categories, the sample feature space result and the sample semantic space result are combined Input to the second sample detection model; clustering the sample feature space results through the second sample detection model to obtain the first abnormal result, and at the same time, perform similarity matching on the sample semantic space results to obtain the first 2.
  • Abnormal result according to the first abnormal result and the second abnormal result, the sample category of the sample to be identified is determined and output.
  • the sample to be identified is first performed through the end-to-end first sample detection model Recognition, when the first sample detection result output by the first sample detection model does not match the first sample category, according to the correlation between the sample feature space result and the sample semantic space result, the first sample based on zero-sample learning Two-sample detection model, which identifies the sample category of the sample to be identified, so that it can automatically identify the sample category that is difficult to collect or new sample category, reduces the cost of manual identification, and automatically assigns difficult-to-collect or new sample category
  • the samples are labeled with sample categories, which saves labor costs and improves recognition accuracy.
  • a sample type identification device is provided, and the sample type identification device corresponds to the sample type identification method in the above-mentioned embodiment in a one-to-one correspondence.
  • the sample category identification device includes an acquisition module 11, an identification module 12, a matching module 13, an abnormality module 14 and an output module 15.
  • the detailed description of each functional module is as follows:
  • the obtaining module 11 is used to obtain a sample to be identified
  • the recognition module 12 is configured to input the sample to be identified into the first sample detection model for sample feature extraction and semantic feature recognition, to obtain the identification result of the sample to be processed; the identification result of the sample to be processed includes the first sample detection result, Sample feature space result and sample semantic space result;
  • the matching module 13 is configured to input the sample feature space result and the sample semantic space result into the second sample detection model when the first sample detection result does not match all the first sample categories ;
  • the abnormality module 14 is configured to perform clustering processing on the sample feature space results through the second sample detection model to obtain a first abnormal result, and at the same time perform similarity matching on the sample semantic space results to obtain a second abnormal result ;
  • the output module 15 is configured to determine and output the sample category of the sample to be identified according to the first abnormal result and the second abnormal result.
  • each module in the above-mentioned sample type identification device can be implemented in whole or in part by software, hardware, and a combination thereof.
  • the above-mentioned modules may be embedded in the form of hardware or independent of the processor in the computer equipment, or may be stored in the memory of the computer equipment in the form of software, so that the processor can call and execute the operations corresponding to the above-mentioned modules.
  • a computer device including a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, wherein the processor executes the computer
  • the following steps are implemented when the instruction is readable:
  • the to-be-identified sample into the first sample detection model to perform sample feature extraction and semantic feature recognition to obtain the to-be-processed sample recognition result;
  • the to-be-processed sample recognition result includes the first sample detection result, the sample feature space result, and the sample Semantic space result;
  • the sample category of the sample to be identified is determined and output.
  • one or more readable storage media storing computer readable instructions are provided, wherein, when the computer readable instructions are executed by one or more processors, the one or more processing The device performs the following steps:
  • the to-be-identified sample into the first sample detection model to perform sample feature extraction and semantic feature recognition to obtain the to-be-processed sample recognition result;
  • the to-be-processed sample recognition result includes the first sample detection result, the sample feature space result, and the sample Semantic space result;
  • the sample category of the sample to be identified is determined and output.
  • a computer device is provided.
  • the computer device may be a server, and its internal structure diagram may be as shown in FIG. 9.
  • the computer equipment includes a processor, a memory, a network interface, and a database connected through a system bus.
  • the processor of the computer device is used to provide calculation and control capabilities.
  • the memory of the computer device includes a readable storage medium and an internal memory.
  • the readable storage medium stores an operating system, computer readable instructions, and a database.
  • the internal memory provides an environment for the operation of the operating system and computer readable instructions in the readable storage medium.
  • the network interface of the computer device is used to communicate with an external terminal through a network connection.
  • the computer-readable instructions are executed by the processor to realize a method for identifying sample categories.
  • the readable storage medium provided in this embodiment includes a non-volatile readable storage medium and a volatile readable storage medium.
  • one or more readable storage media storing computer readable instructions are provided.
  • the readable storage media provided in this embodiment include non-volatile readable storage media and volatile readable storage. Medium; the readable storage medium stores computer readable instructions, and when the computer readable instructions are executed by one or more processors, the one or more processors implement the sample category identification method in the foregoing embodiment.
  • a computer-readable storage medium on which computer-readable instructions are stored, and when the computer-readable instructions are executed by a processor, the sample category identification method in the foregoing embodiment is implemented.
  • Non-volatile memory may include read only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory.
  • Volatile memory may include random access memory (RAM) or external cache memory.
  • RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous chain Channel (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
  • the blockchain referred to in this application is a new application mode of computer technology such as distributed data storage, point-to-point transmission, consensus mechanism, and encryption algorithm.
  • Blockchain essentially a decentralized database, is a series of data blocks associated with cryptographic methods. Each data block contains a batch of network transaction information for verification. The validity of the information (anti-counterfeiting) and the generation of the next block.
  • the blockchain can include the underlying platform of the blockchain, the platform product service layer, and the application service layer.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Computational Linguistics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Probability & Statistics with Applications (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Image Analysis (AREA)

Abstract

本申请涉及人工智能领域，本申请公开了一种样本类别识别方法、装置、计算机设备及存储介质，所述方法包括：获取待识别样本；输入第一样本检测模型进行样本特征提取及语义特征识别，得到待处理样本识别结果；在与所有第一样本类别均不匹配时，将样本特征空间结果和样本语义空间结果输入至第二样本检测模型中；对样本特征空间结果进行聚类处理得到第一异常结果，并对样本语义空间结果进行相似度匹配得到第二异常结果；输出待识别样本的样本类别。本申请实现了自动识别样本类别，节省人工成本，提高识别准确率。本申请适用于智慧交通或智慧医疗等领域，可进一步推动智慧城市的建设，本申请还涉及区块链技术，第一样本检测模型可存储于区块链中。

Description

样本类别识别方法、装置、计算机设备及存储介质
本申请要求于2020年7月28日提交中国专利局、申请号为202010738650.1，发明名称为“样本类别识别方法、装置、计算机设备及存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及人工智能的图像分类领域，尤其涉及一种样本类别识别方法、装置、计算机设备及存储介质。
背景技术
目前，发明人发现随着科技的发展，在生活中图形识别的应用已经越来越普遍。在图像识别中要想得到比较好的识别效果就必须需要每一种类的大量样本图像来进行训练。但是，对于大量收集每一种类的训练样本在实际应用中往往很难实现，特别是对一种含有已知描述且难以收集的异常种类或者新的种类进行识别时，很难收集到这些种类的图像作为样本来进行训练，因为都是靠人工识别出这些异常种类或者新的种类，并且出现的概率极低(是由于需要靠破坏性实验得出或者需耗费代价成本高得出等等)，导致无法训练出可以准确识别出上述种类的模型，还需要花费较高的人工成本进行识别分类，所以在零样本的识别问题上就显得极为重要，而且这个问题越来越受到工业界的关注。比如，在零件质检上，正常零件、断裂零件的图像容易收集，出现轻微裂痕、轻微凹陷、内部断裂等异常情况的零件的图像就难于收集，识别结果不佳；在车辆定损识别上，划痕、刮擦、撕裂、凹陷的车辆损伤照片容易收集，出现褶皱、死折、缺失等严重损伤的车辆损伤照片就难于收集，定损识别就不佳；在医学影像检验上，肺部正常、肺炎、肺积水等常见的X光片容易收集，但是肺结核早期、肺癌早期等罕见的X光片就难于收集，肺部医学检测不佳等等。
发明内容
本申请提供一种样本类别识别方法、装置、计算机设备及存储介质，实现了通过基于端到端的第一样本检测模型和基于零样本学习的第二样本检测模型，自动识别出待识别样本的样本类别，本申请适用于智慧交通或智慧医疗等领域，可进一步推动智慧城市的建设，从而减少了人工识别的成本，并且提高了识别准确率。
一种样本类别识别方法，包括：
获取待识别样本；
将所述待识别样本输入第一样本检测模型进行样本特征提取及语义特征识别，得到待处理样本识别结果；所述待处理样本识别结果包括第一样本检测结果、样本特征空间结果和样本语义空间结果；
在所述第一样本检测结果与所有第一样本类别均不匹配时，则将所述样本特征空间结果和所述样本语义空间结果输入至第二样本检测模型中；
通过所述第二样本检测模型对所述样本特征空间结果进行聚类处理，得到第一异常结果，同时对所述样本语义空间结果进行相似度匹配，得到第二异常结果；
根据所述第一异常结果和所述第二异常结果，确定所述待识别样本的样本类别并输出。
一种样本类别识别装置，包括：
获取模块，用于获取待识别样本；
识别模块，用于将所述待识别样本输入第一样本检测模型进行样本特征提取及语义特征识别，得到待处理样本识别结果；所述待处理样本识别结果包括第一样本检测结果、样本特征空间结果和样本语义空间结果；
匹配模块，用于在所述第一样本检测结果与所有第一样本类别均不匹配时，则将所述样本特征空间结果和所述样本语义空间结果输入至第二样本检测模型中；
异常模块，用于通过所述第二样本检测模型对所述样本特征空间结果进行聚类处理，得到第一异常结果，同时对所述样本语义空间结果进行相似度匹配，得到第二异常结果；
输出模块，用于根据所述第一异常结果和所述第二异常结果，确定所述待识别样本的样本类别并输出。
一种计算机设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令，所述处理器执行所述计算机可读指令时实现如下步骤：
获取待识别样本；
将所述待识别样本输入第一样本检测模型进行样本特征提取及语义特征识别，得到待处理样本识别结果；所述待处理样本识别结果包括第一样本检测结果、样本特征空间结果和样本语义空间结果；
在所述第一样本检测结果与所有第一样本类别均不匹配时，则将所述样本特征空间结果和所述样本语义空间结果输入至第二样本检测模型中；
通过所述第二样本检测模型对所述样本特征空间结果进行聚类处理，得到第一异常结果，同时对所述样本语义空间结果进行相似度匹配，得到第二异常结果；
根据所述第一异常结果和所述第二异常结果，确定所述待识别样本的样本类别并输出。
一个或多个存储有计算机可读指令的可读存储介质，所述计算机可读指令被一个或多个处理器执行时，使得所述一个或多个处理器执行如下步骤：
获取待识别样本；
将所述待识别样本输入第一样本检测模型进行样本特征提取及语义特征识别，得到待处理样本识别结果；所述待处理样本识别结果包括第一样本检测结果、样本特征空间结果和样本语义空间结果；
在所述第一样本检测结果与所有第一样本类别均不匹配时，则将所述样本特征空间结果和所述样本语义空间结果输入至第二样本检测模型中；
通过所述第二样本检测模型对所述样本特征空间结果进行聚类处理，得到第一异常结果，同时对所述样本语义空间结果进行相似度匹配，得到第二异常结果；
根据所述第一异常结果和所述第二异常结果，确定所述待识别样本的样本类别并输出。
本申请提供的样本类别识别方法、装置、计算机设备及存储介质，通过获取待识别样本；将所述待识别样本输入第一样本检测模型进行样本特征提取及语义特征识别，得到待处理样本识别结果；所述待处理样本识别结果包括第一样本检测结果、样本特征空间结果和样本语义空间结果；在所述第一样本检测结果与所有第一样本类别均不匹配时，则将所述样本特征空间结果和所述样本语义空间结果输入至第二样本检测模型中；通过所述第二样本检测模型对所述样本特征空间结果进行聚类处理，得到第一异常结果，同时对所述样本语义空间结果进行相似度匹配，得到第二异常结果；根据所述第一异常结果和所述第二异常结果，确定所述待识别样本的样本类别并输出，如此，实现了首先通过端到端的第一样本检测模型对待识别样本进行识别，在第一样本检测模型输出的第一样本检测结果与第一样本类别不匹配时，根据样本特征空间结果与样本语义空间结果之间的关联关系，通过基于零样本学习的第二样本检测模型，识别出待识别样本的样本类别，从而能够自动对难以收集或者新的样本类别的样本进行准确的样本类别识别，减少了人工识别的成本，自动 给难以收集或者新的样本类别的样本标注样本类别，节省了人工成本，并且提高了识别准确率。
本申请的一个或多个实施例的细节在下面的附图和描述中提出，本申请的其他特征和优点将从说明书、附图以及权利要求变得明显。
附图说明
为了更清楚地说明本申请实施例的技术方案，下面将对本申请实施例的描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动性的前提下，还可以根据这些附图获得其他的附图。
图1是本申请一实施例中样本类别识别方法的应用环境示意图；
图2是本申请一实施例中样本类别识别方法的流程图；
图3是本申请另一实施例中样本类别识别方法的流程图；
图4是本申请一实施例中样本类别识别方法的步骤S20的流程图；
图5是本申请一实施例中样本类别识别方法的步骤S206的流程图；
图6是本申请一实施例中样本类别识别方法的步骤S2063的流程图；
图7是本申请一实施例中样本类别识别方法的步骤S40的流程图；
图8是本申请一实施例中样本类别识别装置的原理框图；
图9是本申请一实施例中计算机设备的示意图。
具体实施方式
下面将结合本申请实施例中的附图，对本申请实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例是本申请一部分实施例，而不是全部的实施例。基于本申请中的实施例，本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其他实施例，都属于本申请保护的范围。
本申请提供的样本类别识别方法，可应用在如图1的应用环境中，其中，客户端(计算机设备)通过网络与服务器进行通信。其中，客户端(计算机设备)包括但不限于为各种个人计算机、笔记本电脑、智能手机、平板电脑、摄像头和便携式可穿戴设备。服务器可以用独立的服务器或者是多个服务器组成的服务器集群来实现。
在一实施例中，如图2所示，提供一种样本类别识别方法，其技术方案主要包括以下步骤S10-S50：
S10，获取待识别样本。
可理解地，在对收集的样本进行识别样本类别的应用场景下触发样本识别请求，获取所述样本识别请求中的所述待识别样本，所述应用场景可以根据需求设定，比如在汽车零件的质检上对收集的零件样本图像进行识别，或者在车辆定损上对用于训练定损识别模型的样本进行识别，或者在医学影像检验上对用于训练肺部识别模型的样本进行识别等等，所述待识别样本为需要进行识别的样本图像或者照片文件。
S20，将所述待识别样本输入第一样本检测模型进行样本特征提取及语义特征识别，得到待处理样本识别结果；所述待处理样本识别结果包括第一样本检测结果、样本特征空间结果和样本语义空间结果。
可理解地，将所述待识别样本输入所述第一样本检测模型中，所述第一样本检测模型为通过多个第一样本类别的训练样本进行训练的多个分支识别的且训练完成的多分支卷积神经网络模型，所述第一样本类别为常规的、容易收集的、大量的样本类别，比如第一样本类别为车辆无损、划痕、刮擦、撕裂、凹陷等，所述第一样本检测模型的网络结构可以根据需求设定，比如网络结构可以为Resnet50的网络结构、CNN的网络结构、VGG的 网络结构等等，作为优选，所述第一样本检测模型的网络结构为Resnet50的网络结构，所述样本特征为多个分支维度的与第一样本类别识别相关的特征，所述样本特征包括高维度显性特征和低维度隐性特征，所述样本特征可以根据需求设定，比如在车损识别中，高维度显性特征包括变形特征、色差特征等，低纬度隐性特征包括光滑度特征等，所述样本特征空间结果为通过所述第一样本检测模型对所述待识别样本进行多分支任务的特征卷积及正则化处理后得到的样本特征向量集合，所述样本特征向量为体现所述样本特征中的一种特征且通过特征向量进行衡量的数组，所述样本特征向量的维度可以根据需求设定，比如样本特征向量的维度可以为256维度，其中，一个分支任务对应所述样本特征中的一种特征，也对应一个所述样本特征向量，所述样本特征空间结果能够通过多个维度体现所述样本特征中的各特征，所述语义特征识别还包括通过输入收敛后的所述嵌入空间模型中的最终映射函数中进行转换，获取所述最终映射函数输出的与其输入对应的语义特征向量，所述最终映射函数为收敛之后的映射函数，将收敛后的所述嵌入空间模型根据所述样本特征空间结果输出的所有所述语义特征向量记录为所述样本语义空间结果。
其中，通过所述第一样本检测模型对所述待识别样本进行识别，获取识别出的与各所述第一样本类别对应的概率值，在所有与各第一样本类别对应的概率值中筛选出所有所述概率值中最大的值，若所有所述概率值中最大的值大于或等于预设的概率阈值，则将所有所述概率值中最大的值对应的第一样本类别确定为与所述待识别样本对应的所述第一样本检测结果；若所有所述概率值中最大的值小于所述概率阈值，则将与所述待识别样本对应的所述第一样本检测结果确认为空或者异常种类。
在一实施例中，如图3所示，所述步骤S20之后，即所述将所述待识别样本输入第一样本检测模型进行样本特征提取及语义特征识别，得到待处理样本识别结果之后，还包括：
S60，在所述第一样本检测结果与任一所述第一样本类别匹配时，将匹配的所述第一样本类别确定为所述待识别样本的样本类别。
可理解地，如果所述第一样本检测结果与所有所述第一样本类别中的其中一个匹配时，所述匹配的方式可以根据需求设定，比如匹配方式为与第一样本类别的内容完全一致，或者与第一样本类别的内容相似度达到预设的类别概率时确定为匹配等等，将匹配的所述第一样本类别记录为所述待识别样本的样本类别。
如此，通过将所述第一样本检测结果与任一个所述第一样本类别进行匹配，若存在与其中一个第一样本类别匹配，则确定出所述待识别样本的样本类别，因此，通过端到端的第一样本检测模型，能够准确识别出所述待识别样本是否为第一样本类别中的一种，提高了识别准确率。
在一实施例中，如图4所示，所述步骤S20之前，即所述将所述待识别样本输入第一样本检测模型进行样本特征提取及语义特征识别之前，包括：
S201，获取训练样本集；所述训练样本集包含多个训练样本，一个所述训练样本与一个第一样本类别以及一个第一样本描述关联。
可理解地，所述训练样本集为所述训练样本的集合，所述训练样本为容易收集的各种第一样本类别的图像，一个所述训练样本都与一个第一样本类别关联，而且与一个所述第一样本描述关联，所述第一样本描述为对所述训练样本中的所述第一样本类别的文本描述，比如一张车身被刮擦的车辆损伤照片作为训练样本，关联的第一样本类别为“刮擦”，关联的第一样本描述为“车身中间部分被刮出长度为7厘米和宽度为2厘米的无明显变形及凹陷的损伤面积”。
S202，将所述训练样本输入含有初始参数的多分支卷积神经网络模型。
可理解地，将所述训练样本输入至所述多分支卷积神经网络模型，所述多分支卷积神经网络模型包括所述初始参数，所述初始参数包括网络结构的参数和特征向量的维度参数等等。
在一实施例中，所述步骤S202之前，即所述将所述训练样本输入含有初始参数的多分支卷积神经网络模型之前，包括：
S2021，通过迁移学习，获取训练完成的无监督域训练模型的所有迁移参数，将所有所述迁移参数确定为所述多分支卷积神经网络模型中的所述初始参数。
可理解地，所述迁移学习(Transfer Learning，TL)为利用其他领域已有的训练模型的参数应用在本领域的任务中，即所述多分支卷积神经网络模型通过迁移学习的方式获取无监督域训练模型(unsupervised pre-training)的所有模型参数，所述无监督域训练模型可以为自然领域或者车辆识别领域的无监督学习的多分支卷积神经网络模型，然后将所述所有模型参数确定为所述多分支卷积神经网络模型的初始参数。如此，本申请通过迁移学习，从训练完成的无监督域训练模型中获取初始参数，能够缩短模型的迭代次数，简化训练过程，及提高训练效率。
S203，通过所述多分支卷积神经网络模型对所述训练样本进行样本特征提取，得到图像特征空间结果。
可理解地，所述样本特征为多个分支维度的与样本识别相关的特征，所述样本特征包括高维度显性特征和低维度隐性特征，所述样本特征可以根据需求设定，比如在车损识别中，高维度显性特征包括变形特征、色差特征等，低纬度隐性特征包括光滑度特征等，所述图像特征空间结果为通过所述多分支卷积神经网络模型对所述训练样本进行多分支任务的特征卷积及正则化处理后得到的图像特征向量，所述图像特征向量为体现所述样本特征中的一种特征且通过特征向量进行衡量的数组，所述图像特征向量的维度可以根据需求设定，比如图像特征向量的维度可以为256维度，其中，一个分支任务对应所述样本特征中的一种特征，也对应一个所述图像特征向量，所述图像特征空间结果能够通过多个维度体现所述样本特征中的各特征。
在一实施例中，所述S203中，即所述通过所述多分支卷积神经网络模型对所述训练样本进行样本特征提取，得到图像特征空间结果，包括：
S2031，对所述训练样本进行图像预处理，得到预处理样本图像；
可理解地，所述预处理为对所述训练样本进行特征区域识别、提取预设尺寸图像并对提取后的图像进行增强处理的操作过程，在所述训练样本中识别出存在样本特征的区域，可以通过YOLO(You Only Look Once)算法进行识别，识别出含有样本特征的特征区域后提取出预设尺寸的图像，预设尺寸可以根据需求设定，作为优选，预设尺寸为224×224的尺寸，对提取后的图像进行增强处理，所述增强处理可以根据需求设定，比如增强处理为去噪及锐化的处理，从而获得所述预处理样本图像。
S2032，对所述预处理样本图像进行特征提取，得到至少一个特征向量图；
可理解地，对所述预处理样本图像的各通道进行特征提取，所述特征提取为通过不同特征对应的各个不同的卷积核对各通道进行卷积输出特征向量图，卷积的过程根据所述多分支卷积神经网络模型的网络结构确定，比如对所述预处理样本图像的各个通道进行高维度特征的特征提取，得到高维度特征对应的特征向量图，所述特征向量图由多个特征向量值组成。
S2033，对各所述特征向量图进行正则化处理，得到与各所述特征向量图对应的图像特征向量；
可理解地，所述正则化处理为对各特征向量图中的特征向量值进行N次根正则化非线性处理，通过所述正则化处理可以对所述特征向量图进行微调，更能体现各特征，从而输出图像特征向量。
S2034，将所有所述图像特征向量确定为所述图像特征空间结果。
可理解地，将所有所述图像特征向量标记为所述图像特征空间结果。
本申请实现了通过对所述训练样本进行图像预处理，得到预处理样本图像；对所述预 处理样本图像进行特征提取，得到至少一个特征向量图；对各所述特征向量图进行正则化处理，得到与各所述特征向量图对应的图像特征向量；将所有所述图像特征向量确定为所述图像特征空间结果，如此，通过图像预处理、特征提取和正则化处理，能够提高识别的准确率和可靠性。
S204，通过所述多分支卷积神经网络模型中的嵌入空间模型对所述图像特征空间结果进行语义特征识别，得到语义特征空间结果；所述嵌入空间模型为通过构建图像特征向量与语义特征向量之间的关联关系获得。
可理解地，所述嵌入空间模型为通过训练样本对应的图像特征向量和与训练样本对应的第一样本描述转换成语义特征向量之间的映射关系进行学习获得，即将训练样本的第一样本描述输入基于Word2vec的语义识别模型，对第一样本描述进行语义向量转换，得到与第一样本描述对应的语义特征向量，所述基于Word2vec的语义识别模型为训练完成的深度神经网络模型，所述语义识别模型能够将输入的文本进行语义向量转换生成与所述样本特征的各特征对应的语义特征的向量值，即生成语义特征向量，例如：第一样本描述描述为“车身中间部分被刮出长度为7厘米和宽度为2厘米的无明显变形及凹陷的损伤面积”经过语义向量转换之后会得到在变形相关的语义特征对应的向量值，在色差相关的语义特征对应的向量值，在光滑面相关的语义特征对应的向量值等等，建立训练样本对应的图像特征向量和语义特征向量之间的映射函数，通过所述多分支卷积神经网络模型的不断迭代，进而不断学习，直至达到所述多分支卷积神经网络模型收敛，此时训练样本对应的图像特征向量和语义特征向量之间的映射函数学习完成，即图像特征向量输入映射函数后输出的值和与图像特征向量对应的语义特征向量之间的均方差达到最小，也即所述嵌入空间模型收敛完成，所述语义特征识别为通过输入所述嵌入空间模型中的映射函数中进行转换，获取所述映射函数输出的与其输入对应的语义特征向量，将嵌入空间模型根据所述图像特征空间结果输出的所有所述语义特征向量记录为所述语义特征空间结果。
S205，通过K-means聚类算法，对所述图像特征空间结果进行识别，得到训练类别结果，并将所述训练类别结果、所述图像特征空间结果和所述语义特征空间结果确定为样本训练结果。
可理解地，所述K-means聚类算法也称K均值聚类算法，采用距离作为相似性的评价指标，通过根据靠近各训练类别的簇的距离确定其对应的训练类别，通过所述K-means聚类算法，计算出所述图像特征空间结果中的各图像特征向量与各训练类别结果的中心的欧式距离，根据各欧式距离确定出与其对应训练类别结果，所述训练类别结果中包含的类别与所述第一样本类别中包含的类别完全相同，从而做到端到端的识别，达到集中提取相同类别对应的特征，提高了识别准确率。
其中，将所述训练样本对应的所述训练类别结果、所述图像特征空间结果和所述语义特征空间结果标记为所述训练样本对应的所述样本训练结果。
S206，根据所述图像特征空间结果和与所述训练样本对应的所述第一样本类别，得到第一损失值；根据所述语义特征空间结果和与所述训练样本对应的所述第一样本描述，得到第二损失值。
可理解地，通过计算所述图像特征空间结果和与所述第一样本类别对应的中心向量之间的差距，得到所述第一损失值；将所述第一样本描述输入基于Word2vec的语义识别模型中，对所述第一样本描述进行语义向量转换，得到与所述第一样本描述对应的语义特征向量，通过计算所述语义特征空间结果和与该语义特征向量之间的语义相似度值，得到所述第二损失值，所述第二损失值表征了所述语义特征空间结果与该语义特征向量之间的差距。
在一实施例中，如图5所示，所述步骤S206中，即所述通过所述多分支卷积神经网络模型对所述训练样本进行样本特征提取，得到图像特征空间结果，包括：
S2061，通过所述多分支卷积神经网络模型将与所述训练样本对应的所述第一样本类别进行向量转换，得到与所述第一样本类别对应的中心向量；所述中心向量包括欧式域中心向量和角度域中心向量。
可理解地，将所述第一样本类别进行向量转换，所述向量转换为将文本类型的所述第一样本类别按照预设的映射关系进行转换，得到与所述第一样本类别匹配的所述中心向量，所述中心向量包括所述欧式域中心向量和所述角度域中心向量。
S2062，通过交叉熵损失算法，根据所述图像特征空间结果和所述欧式域中心向量，得到欧式损失值。
可理解地，通过交叉熵损失函数，计算出所述图像特征空间结果中的各所述图像特征向量与所述欧式域中心向量之间的损失值，将该损失值记录为所述欧式损失值。
S2063，通过ArcFace损失算法，根据所述图像特征空间结果和所述角度域中心向量，得到角度损失值。
可理解地，通过ArcFace损失函数，计算出所述图像特征空间结果中的各所述图像特征向量与所述角度域中心向量之间的损失值，将该损失值记录为所述角度损失值。
在一实例中，如图6所示，所述步骤S2063中，即所述通过ArcFace损失算法，根据所述图像特征空间结果和所述角度域中心向量，得到角度损失值，包括：
S20631，通过所述多分支卷积神经网络模型中的正则化模型，对所述图像特征空间结果进行正则化处理，得到正则化特征向量。
可理解地，所述多分支卷积神经网络模型包括所述正则化模型，所述正则化模型包含有正则化函数，将各所述图像特征向量输入正则化函数中，输出所述正则化特征向量，所述正则化特征向量为：
Figure PCTCN2020135337-appb-000001
，其中，数学符号函数：
Figure PCTCN2020135337-appb-000002
，b 1为图像特征向量中的第一个特征向量值，,b m为图像特征向量中的第m个特征向量值，b i为图像特征向量中的第i个特征向量值，N为预设的n次方根的值。
S20632，将所述正则化特征向量与所述角度域中心向量输入角度域损失模型，通过所述角度域损失模型中的所述ArcFace损失算法，得到角度损失值。
可理解地，所述ArcFace损失算法为通过ArcFace损失函数计算其损失值的算法，将所述正则化特征向量与所述角度域中心向量通过ArcFace损失函数中的余弦角度比对方法，通过角度域的方向衡量其差距，从而得到所述角度损失值。
如此，本申请实现了通过所述多分支卷积神经网络模型中的正则化模型、角度域损失模型，获得角度损失值，因此，通过正则化能够限制模型发散，起到在识别分类过程中微调的作用，在通过角度域损失模型能够增加类别之间距离紧凑的考虑。
S2064，对所述欧式损失值和所述角度损失值进行加权处理，得到所述第一损失值。
可理解地，通过将所述欧式损失值和所述角度损失值输入至第一损失函数中，将所述欧式损失值和所述角度损失值转换成同一维度的指标并进行加权乘积处理，得到所述第一损失函数输出的所述第一损失值，通过在欧式损失值的基础上增加角度损失值的考虑，能够不仅对正确类别分类的考虑，还对类别之间距离紧凑的考虑，使得第一损失函数在分类结果上有很好的表现力。如此，本申请实现了通过所述多分支卷积神经网络模型将与所述训练样本对应的所述第一样本类别进行向量转换，得到与所述第一样本类别对应的中心向量；通过交叉熵损失算法，获取欧式损失值；通过ArcFace损失算法，获取角度损失值； 对所述欧式损失值和所述角度损失值进行加权处理，得到所述第一损失值，因此，通过引入角度损失值让识别向更加紧凑的识别方向靠拢，让识别精度更加准确。
S207，根据所述第一损失值和所述第二损失值，得到总损失值。
可理解地，将所述第一损失值和所述第二损失值输入所述多分支卷积神经网络模型中的含有总损失函数的总损失模型，所述总损失模型中的总损失函数可以根据需求设定，所述损失模型为生成所述总损失值的模型，通过所述总损失函数计算出所述总损失值。
在一实施例中，所述步骤S207中，即所述根据所述第一损失值和所述第二损失值，得到总损失值，包括：
S2071，将所述第一损失值和所述第二损失值输入预设的总损失模型，通过所述总损失模型中的总损失函数计算出所述总损失值；所述总损失函数为：
L＝W 1×X1+W 2×X2
其中，
X1为第一损失值；
X2为第二损失值；
W 1为第一损失值的权重；
W 2为第二损失值的权重。
S208，在所述总损失值未达到预设的收敛条件时，迭代更新所述多分支卷积神经网络模型的初始参数，并触发所述通过所述多分支卷积神经网络模型对所述训练样本进行样本特征提取，得到图像特征空间结果的步骤，直至所述总损失值达到所述预设的收敛条件时，将收敛之后的所述多分支卷积神经网络模型记录为训练完成的所述第一样本检测模型。
可理解地，所述收敛条件可以为所述总损失值经过了10000次计算后值为很小且不会再下降的条件，即在所述总损失值经过10000次计算后值为很小且不会再下降时，停止训练，并将收敛之后的所述多分支卷积神经网络模型记录为训练完成的所述第一样本检测模型；所述收敛条件也可以为所述总损失值小于设定阈值的条件，即在所述总损失值小于设定阈值时，停止训练，并将收敛之后的所述多分支卷积神经网络模型记录为训练完成的所述第一样本检测模型，可将训练完成的所述第一样本检测模型存储在区块链中。如此，在所述总损失值未达到预设的收敛条件时，不断更新迭代所述多分支卷积神经网络模型的初始参数，并触发所述通过所述多分支卷积神经网络模型对所述训练样本进行样本特征提取，得到图像特征空间结果的步骤，可以不断向准确的结果靠拢，让识别的准确率越来越高。
需要强调的是，为进一步保证上述第一样本检测模型的私密和安全性，上述第一样本检测模型还可以存储于区块链的节点中。
其中，本申请所指区块链是分布式数据存储、点对点传输、共识机制、加密算法等计算机技术的新型应用模式。区块链(Blockchain)，本质上是一个去中心化的数据库，是一串使用密码学方法相关联产生的数据块，每一个数据块中包含了一批次网络交易的信息，用于验证其信息的有效性(防伪)和生成下一个区块。区块链可以包括区块链底层平台、平台产品服务层以及应用服务层等。区块链提供的去中心化的完全分布式DNS服务通过网络中各个节点之间的点对点数据传输服务就能实现域名的查询和解析，可用于确保某个重要的基础设施的操作系统和固件没有被篡改，可以监控软件的状态和完整性，发现不良的篡改，并确保所传输的数据没用经过篡改，将所述第一样本检测模型存储在区块链中，能够确保第一样本检测模型的私密和安全性。
在一实施例中，所述步骤S207之后，即所述根据所述第一损失值和所述第二损失值，得到总损失值之后，还包括：
S209，在所述总损失值达到所述收敛条件时，将收敛之后的所述多分支卷积神经网络模型记录为训练完成的所述第一样本检测模型。
可理解地，在第一次出现总损失值时，若所述总损失值达到所述收敛条件，说明所述总损失值已经达到最优的结果，此时所述多分支卷积神经网络模型已经收敛，将收敛之后的所述多分支卷积神经网络模型记录为训练完成的所述第一样本检测模型。
S30，在所述第一样本检测结果与所有第一样本类别均不匹配时，则将所述样本特征空间结果和所述样本语义空间结果输入至第二样本检测模型中。
可理解地，如果所述第一样本检测结果与所有所述第一样本类别中的任何一种类别都不匹配，则将所述待识别样本记录为所述异常样本，所述异常样本表明了与所有第一样本类别以外的结果都为异常，将所述样本特征空间结果和所述样本语义空间结果输入至所述第二样本检测模型中，所述第二样本检测模型为根据所述最终映射函数和K-近邻算法，对已知的第二样本类别在其对应的所述样本特征中的各特征中所体现的特征进行人工标记而学习的自适应模型，即建立所述样本特征中的各特征的聚类范围和语义特征向量的范围与已知的第二样本类别的映射关系，所述第二样本类别为难以收集的(少量的)且已知描述(已经知道的与样本特征相关的描述，也为全文中的第二样本描述)的样本类别和新发现的且已知描述的样本类别，比如在车辆定损识别上，基于对“缺失”特征认知(具有“撕裂”和“凹陷”的共性特征)，将邻近“撕裂”和“凹陷”的聚类簇的中心的预设区域范围内，或/和将在变形相关的语义特征对应的向量值，在色差相关的语义特征对应的向量值和在光滑面相关的语义特征对应的向量值进行加权相加后获得的值落在“撕裂”和“凹陷”的近邻范围值内，映射为“缺失”的第二样本类别，所述第二样本检测模型根据人工标记各第二样本类别与第一样本类别之间的共性特征自动调整映射关系中的映射函数，最终达到拟合程度最高的参数。
S40，通过所述第二样本检测模型对所述样本特征空间结果进行聚类处理，得到第一异常结果，同时对所述样本语义空间结果进行相似度匹配，得到第二异常结果。
可理解地，所述聚类处理的方式为通过K-means聚类算法对所述样本特征空间结果进行计算，计算出所述样本特征空间结果距离与各第一样本类别对应的簇类的中心(质心)的欧式距离值，通过所述第二样本检测模型计算出各欧式距离值后，根据各欧式距离值落在的区域范围内确定出与该区域范围对应的第二样本类别，将其确定为聚类异常类别，并将各欧式距离值转换成与各所述第二样本类别的概率值，将所述聚类异常类别和与各所述第二样本类别的概率值记录为所述第一异常结果，所述第一异常结果表明了通过样本特征空间的维度识别出的结果，同时通过所述第二样本检测模型计算出所述样本语义空间结果与所述各第一样本类别对应的语义特征向量之间的相似度值，得到正相似度值，计算正相似度值的算法可以根据需求设定，比如余弦相似度算法、Jaccard相似度算法等等，根据各所述正相似度值确定出邻近的第一样本类别，再根据其邻近的第一样本类别映射出语义相似类别，将映射出的所述语义相似类别和各正相似度值记录为所述第二异常结果，所述第二异常结果表明了通过语义空间结果的维度识别出的结果。
在一实施例中，如图7所示，所述步骤S40中，即所述对所述样本语义空间结果进行相似度匹配，得到第二异常结果，包括：
S401，通过所述第二样本检测模型中的Word2vec模型，计算所述样本语义空间结果与预设的各第二样本类别对应的第二样本描述之间的相似度值。
可理解地，所述第二样本检测模型还可以通过所述Word2vec模型将预设的各第二样本类别对应的第二样本描述进行词向量转换，得到与各所述第二样本类别对应的语义词向量，所述第二样本类别为难以收集的(少量的)且已知描述(已经知道的与样本特征相关的描述，也为全文中的第二样本描述)的样本类别和新发现的且已知描述的样本类别，再通过所述Word2vec模型中的Word2vec算法计算出所述样本语义空间结果与各所述语义词向量之间的相似度值。
其中，所述Word2vec模型为基于Word2vec算法构建的神经网络模型，在所述Word2vec 模型中预先设置了各第二样本类别对应的与样本特征中的各特征相关的第二样本描述。
S402，通过所述第二样本检测模型获取与所有所述相似度值中最大的所述相似度值对应的所述第二样本类别，将获取的所述第二样本类别确定为所述第二异常结果。
可理解地，从所有所述样本语义空间结果与各所述语义词向量之间的相似度值中获取最大的所述相似度值，并将其对应的所述异常类型确定为所述第二异常结果，所述第二异常结果中包括获取的最大的所述相似度值。
如此，本申请实现了通过所述第二样本检测模型中的Word2vec模型，计算所述样本语义空间结果与预设的各第二样本类别对应的第二样本描述之间的相似度值；再通过获取与所有所述相似度值中最大的所述相似度值对应的所述第二样本类别，将获取的所述第二样本类别确定为所述第二异常结果，因此，实现了基于Word2vec算法自动识别出第二异常结果，提供了一种语义识别的方法，让识别更加准确。
S50，根据所述第一异常结果和所述第二异常结果，确定所述待识别样本的样本类别并输出。
可理解地，通过将所述第一异常结果中的各概率值和所述第二异常结果中的各概率值进行加权处理，根据设定的权重参数进行计算得出最终的第二样本类别，通过权重参数拉开各第二样本类别对应的概率值的差距，让分类更加准确，比如经过权重参数进行加权后的“缺失”的第二样本类别在第一异常结果和第二异常结果中的概率值之和最大，则将“缺失”的第二样本类别确定为所述待识别样本的样本类别。
其中，所述样本类别包括所述第一样本类别和所述第二样本类别，在一应用场景中，从一批样本中，可以从该批样本中识别出第一样本类别的样本，也可以基于零个第二样本类别的训练样本基础上识别出第二样本类别的样本，例如：在收集的车损照片样本集中，可以识别出划痕、刮擦、撕裂、凹陷的第一样本类别的车损照片，以及可以识别出褶皱、死折、缺失等第二样本类别的车损照片(在此之前未收集到褶皱、死折、缺失等第二样本类别的照片)。
本申请实现了通过获取待识别样本；将所述待识别样本输入第一样本检测模型进行样本特征提取及语义特征识别，得到待处理样本识别结果；所述待处理样本识别结果包括第一样本检测结果、样本特征空间结果和样本语义空间结果；在所述第一样本检测结果与所有第一样本类别均不匹配时，则将所述样本特征空间结果和所述样本语义空间结果输入至第二样本检测模型中；通过所述第二样本检测模型对所述样本特征空间结果进行聚类处理，得到第一异常结果，同时对所述样本语义空间结果进行相似度匹配，得到第二异常结果；根据所述第一异常结果和所述第二异常结果，确定所述待识别样本的样本类别并输出，如此，实现了首先通过端到端的第一样本检测模型对待识别样本进行识别，在第一样本检测模型输出的第一样本检测结果与第一样本类别不匹配时，根据样本特征空间结果与样本语义空间结果之间的关联关系，通过基于零样本学习的第二样本检测模型，识别出待识别样本的样本类别，从而能够自动对难以收集或者新的样本类别的样本进行准确的样本类别识别，减少了人工识别的成本，自动给难以收集或者新的样本类别的样本标注样本类别，节省了人工成本，并且提高了识别准确率。
在一实施例中，提供一种样本类别识别装置，该样本类别识别装置与上述实施例中样本类别识别方法一一对应。如图8所示，该样本类别识别装置包括获取模块11、识别模块12、匹配模块13、异常模块14和输出模块15。各功能模块详细说明如下：
获取模块11，用于获取待识别样本；
识别模块12，用于将所述待识别样本输入第一样本检测模型进行样本特征提取及语义特征识别，得到待处理样本识别结果；所述待处理样本识别结果包括第一样本检测结果、样本特征空间结果和样本语义空间结果；
匹配模块13，用于在所述第一样本检测结果与所有第一样本类别均不匹配时，则将所 述样本特征空间结果和所述样本语义空间结果输入至第二样本检测模型中；
异常模块14，用于通过所述第二样本检测模型对所述样本特征空间结果进行聚类处理，得到第一异常结果，同时对所述样本语义空间结果进行相似度匹配，得到第二异常结果；
输出模块15，用于根据所述第一异常结果和所述第二异常结果，确定所述待识别样本的样本类别并输出。
关于样本类别识别装置的具体限定可以参见上文中对于样本类别识别方法的限定，在此不再赘述。上述样本类别识别装置中的各个模块可全部或部分通过软件、硬件及其组合来实现。上述各模块可以硬件形式内嵌于或独立于计算机设备中的处理器中，也可以以软件形式存储于计算机设备中的存储器中，以便于处理器调用执行以上各个模块对应的操作。
在一实施例中，提供了一种计算机设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令，其中，所述处理器执行所述计算机可读指令时实现如下步骤：
获取待识别样本；
将所述待识别样本输入第一样本检测模型进行样本特征提取及语义特征识别，得到待处理样本识别结果；所述待处理样本识别结果包括第一样本检测结果、样本特征空间结果和样本语义空间结果；
在所述第一样本检测结果与所有第一样本类别均不匹配时，则将所述样本特征空间结果和所述样本语义空间结果输入至第二样本检测模型中；
通过所述第二样本检测模型对所述样本特征空间结果进行聚类处理，得到第一异常结果，同时对所述样本语义空间结果进行相似度匹配，得到第二异常结果；
根据所述第一异常结果和所述第二异常结果，确定所述待识别样本的样本类别并输出。
在一实施例中，提供了一个或多个存储有计算机可读指令的可读存储介质，其中，所述计算机可读指令被一个或多个处理器执行时，使得所述一个或多个处理器执行如下步骤：
获取待识别样本；
将所述待识别样本输入第一样本检测模型进行样本特征提取及语义特征识别，得到待处理样本识别结果；所述待处理样本识别结果包括第一样本检测结果、样本特征空间结果和样本语义空间结果；
在所述第一样本检测结果与所有第一样本类别均不匹配时，则将所述样本特征空间结果和所述样本语义空间结果输入至第二样本检测模型中；
通过所述第二样本检测模型对所述样本特征空间结果进行聚类处理，得到第一异常结果，同时对所述样本语义空间结果进行相似度匹配，得到第二异常结果；
根据所述第一异常结果和所述第二异常结果，确定所述待识别样本的样本类别并输出。
在一个实施例中，提供了一种计算机设备，该计算机设备可以是服务器，其内部结构图可以如图9所示。该计算机设备包括通过系统总线连接的处理器、存储器、网络接口和数据库。其中，该计算机设备的处理器用于提供计算和控制能力。该计算机设备的存储器包括可读存储介质、内存储器。该可读存储介质存储有操作系统、计算机可读指令和数据库。该内存储器为可读存储介质中的操作系统和计算机可读指令的运行提供环境。该计算机设备的网络接口用于与外部的终端通过网络连接通信。该计算机可读指令被处理器执行时以实现一种样本类别识别方法。本实施例所提供的可读存储介质包括非易失性可读存储介质和易失性可读存储介质。
在一个实施例中，提供了一个或多个存储有计算机可读指令的可读存储介质，本实施例所提供的可读存储介质包括非易失性可读存储介质和易失性可读存储介质；该可读存储介质上存储有计算机可读指令，该计算机可读指令被一个或多个处理器执行时，使得一个或多个处理器实现上述实施例中样本类别识别方法。
在一个实施例中，提供了一种计算机可读存储介质，其上存储有计算机可读指令，计算机可读指令被处理器执行时实现上述实施例中样本类别识别方法。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程，是可以通过计算机可读指令来指令相关的硬件来完成，所述的计算机可读指令可存储于一非易失性计算机可读取存储介质或易失性可读存储介质中，该计算机可读指令在执行时，可包括如上述各方法的实施例的流程。其中，本申请所提供的各实施例中所使用的对存储器、存储、数据库或其它介质的任何引用，均可包括非易失性和/或易失性存储器。非易失性存储器可包括只读存储器(ROM)、可编程ROM(PROM)、电可编程ROM(EPROM)、电可擦除可编程ROM(EEPROM)或闪存。易失性存储器可包括随机存取存储器(RAM)或者外部高速缓冲存储器。作为说明而非局限，RAM以多种形式可得，诸如静态RAM(SRAM)、动态RAM(DRAM)、同步DRAM(SDRAM)、双数据率SDRAM(DDRSDRAM)、增强型SDRAM(ESDRAM)、同步链路(Synchlink)DRAM(SLDRAM)、存储器总线(Rambus)直接RAM(RDRAM)、直接存储器总线动态RAM(DRDRAM)、以及存储器总线动态RAM(RDRAM)等。
本申请所指区块链是分布式数据存储、点对点传输、共识机制、加密算法等计算机技术的新型应用模式。区块链(Blockchain)，本质上是一个去中心化的数据库，是一串使用密码学方法相关联产生的数据块，每一个数据块中包含了一批次网络交易的信息，用于验证其信息的有效性(防伪)和生成下一个区块。区块链可以包括区块链底层平台、平台产品服务层以及应用服务层等。
所属领域的技术人员可以清楚地了解到，为了描述的方便和简洁，仅以上述各功能单元、模块的划分进行举例说明，实际应用中，可以根据需要而将上述功能分配由不同的功能单元、模块完成，即将所述装置的内部结构划分成不同的功能单元或模块，以完成以上描述的全部或者部分功能。
以上所述实施例仅用以说明本申请的技术方案，而非对其限制；尽管参照前述实施例对本申请进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施例所记载的技术方案进行修改，或者对其中部分技术特征进行等同替换；而这些修改或者替换，并不使相应技术方案的本质脱离本申请各实施例技术方案的精神和范围，均应包含在本申请的保护范围之内。

Claims (20)

  1. 一种样本类别识别方法，其中，包括：
    获取待识别样本；
    将所述待识别样本输入第一样本检测模型进行样本特征提取及语义特征识别，得到待处理样本识别结果；所述待处理样本识别结果包括第一样本检测结果、样本特征空间结果和样本语义空间结果；
    在所述第一样本检测结果与所有第一样本类别均不匹配时，则将所述样本特征空间结果和所述样本语义空间结果输入至第二样本检测模型中；
    通过所述第二样本检测模型对所述样本特征空间结果进行聚类处理，得到第一异常结果，同时对所述样本语义空间结果进行相似度匹配，得到第二异常结果；
    根据所述第一异常结果和所述第二异常结果，确定所述待识别样本的样本类别并输出。
  2. 如权利要求1所述的样本类别识别方法，其中，所述将所述待识别样本输入第一样本检测模型进行样本特征提取及语义特征识别，得到待处理样本识别结果之后，还包括：
    在所述第一样本检测结果与任一所述第一样本类别匹配时，将匹配的所述第一样本类别确定为所述待识别样本的样本类别。
  3. 如权利要求1所述的样本类别识别方法，其中，所述将所述待识别样本输入第一样本检测模型进行样本特征提取及语义特征识别之前，包括：
    获取训练样本集；所述训练样本集包含多个训练样本，一个所述训练样本与一个第一样本类别以及一个第一样本描述关联；
    将所述训练样本输入含有初始参数的多分支卷积神经网络模型；
    通过所述多分支卷积神经网络模型对所述训练样本进行样本特征提取，得到图像特征空间结果；
    通过所述多分支卷积神经网络模型中的嵌入空间模型对所述图像特征空间结果进行语义特征识别，得到语义特征空间结果；所述嵌入空间模型为通过构建图像特征向量与语义特征向量之间的关联关系获得；
    通过K-means聚类算法，对所述图像特征空间结果进行识别，得到训练类别结果，并将所述训练类别结果、所述图像特征空间结果和所述语义特征空间结果确定为样本训练结果；
    根据所述图像特征空间结果和与所述训练样本对应的所述第一样本类别，得到第一损失值；根据所述语义特征空间结果和与所述训练样本对应的所述第一样本描述，得到第二损失值；
    根据所述第一损失值和所述第二损失值，得到总损失值；
    在所述总损失值未达到预设的收敛条件时，迭代更新所述多分支卷积神经网络模型的初始参数，并触发所述通过所述多分支卷积神经网络模型对所述训练样本进行样本特征提取，得到图像特征空间结果的步骤，直至所述总损失值达到所述预设的收敛条件时，将收敛之后的所述多分支卷积神经网络模型记录为训练完成的所述第一样本检测模型。
  4. 如权利要求3所述的样本类别识别方法，其中，所述将所述训练样本输入含有初始参数的多分支卷积神经网络模型之前，包括：
    通过迁移学习，获取训练完成的无监督域训练模型的所有迁移参数，将所有所述迁移参数确定为所述多分支卷积神经网络模型中的所述初始参数。
  5. 如权利要求3所述的样本类别识别方法，其中，所述根据所述图像特征空间结果和与所述训练样本对应的所述第一样本类别，得到第一损失值，包括：
    通过所述多分支卷积神经网络模型将与所述训练样本对应的所述第一样本类别进行 向量转换，得到与所述第一样本类别对应的中心向量；所述中心向量包括欧式域中心向量和角度域中心向量；
    通过交叉熵损失算法，根据所述图像特征空间结果和所述欧式域中心向量，得到欧式损失值；
    通过ArcFace损失算法，根据所述图像特征空间结果和所述角度域中心向量，得到角度损失值；
    对所述欧式损失值和所述角度损失值进行加权处理，得到所述第一损失值。
  6. 如权利要求5所述的样本类别识别方法，其中，所述通过ArcFace损失算法，根据所述图像特征空间结果和所述角度域中心向量，得到角度损失值，包括：
    通过所述多分支卷积神经网络模型中的正则化模型，对所述图像特征空间结果进行正则化处理，得到正则化特征向量；
    将所述正则化特征向量与所述角度域中心向量输入角度域损失模型，通过所述角度域损失模型中的所述ArcFace损失算法，得到角度损失值。
  7. 如权利要求1所述的样本类别识别方法，其中，所述对所述样本语义空间结果进行相似度匹配，得到第二异常结果，包括：
    通过所述第二样本检测模型中的Word2vec模型，计算所述样本语义空间结果与预设的各第二样本类别对应的第二样本描述之间的相似度值；
    通过所述第二样本检测模型获取与所有所述相似度值中最大的所述相似度值对应的所述第二样本类别，将获取的所述第二样本类别确定为所述第二异常结果。
  8. 一种样本类别识别装置，其中，包括：
    获取模块，用于获取待识别样本；
    识别模块，用于将所述待识别样本输入第一样本检测模型进行样本特征提取及语义特征识别，得到待处理样本识别结果；所述待处理样本识别结果包括第一样本检测结果、样本特征空间结果和样本语义空间结果；
    匹配模块，用于在所述第一样本检测结果与所有第一样本类别均不匹配时，则将所述样本特征空间结果和所述样本语义空间结果输入至第二样本检测模型中；
    异常模块，用于通过所述第二样本检测模型对所述样本特征空间结果进行聚类处理，得到第一异常结果，同时对所述样本语义空间结果进行相似度匹配，得到第二异常结果；
    输出模块，用于根据所述第一异常结果和所述第二异常结果，确定所述待识别样本的样本类别并输出。
  9. 一种计算机设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令，其中，所述处理器执行所述计算机可读指令时实现如下步骤：
    获取待识别样本；
    将所述待识别样本输入第一样本检测模型进行样本特征提取及语义特征识别，得到待处理样本识别结果；所述待处理样本识别结果包括第一样本检测结果、样本特征空间结果和样本语义空间结果；
    在所述第一样本检测结果与所有第一样本类别均不匹配时，则将所述样本特征空间结果和所述样本语义空间结果输入至第二样本检测模型中；
    通过所述第二样本检测模型对所述样本特征空间结果进行聚类处理，得到第一异常结果，同时对所述样本语义空间结果进行相似度匹配，得到第二异常结果；
    根据所述第一异常结果和所述第二异常结果，确定所述待识别样本的样本类别并输出。
  10. 如权利要求9所述的计算机设备，其中，所述将所述待识别样本输入第一样本检测模型进行样本特征提取及语义特征识别，得到待处理样本识别结果之后，所述处理器执行所述计算机可读指令时还实现如下步骤：
    在所述第一样本检测结果与任一所述第一样本类别匹配时，将匹配的所述第一样本类别确定为所述待识别样本的样本类别。
  11. 如权利要求9所述的计算机设备，其中，所述将所述待识别样本输入第一样本检测模型进行样本特征提取及语义特征识别之前，所述处理器执行所述计算机可读指令时还实现如下步骤：
    获取训练样本集；所述训练样本集包含多个训练样本，一个所述训练样本与一个第一样本类别以及一个第一样本描述关联；
    将所述训练样本输入含有初始参数的多分支卷积神经网络模型；
    通过所述多分支卷积神经网络模型对所述训练样本进行样本特征提取，得到图像特征空间结果；
    通过所述多分支卷积神经网络模型中的嵌入空间模型对所述图像特征空间结果进行语义特征识别，得到语义特征空间结果；所述嵌入空间模型为通过构建图像特征向量与语义特征向量之间的关联关系获得；
    通过K-means聚类算法，对所述图像特征空间结果进行识别，得到训练类别结果，并将所述训练类别结果、所述图像特征空间结果和所述语义特征空间结果确定为样本训练结果；
    根据所述图像特征空间结果和与所述训练样本对应的所述第一样本类别，得到第一损失值；根据所述语义特征空间结果和与所述训练样本对应的所述第一样本描述，得到第二损失值；
    根据所述第一损失值和所述第二损失值，得到总损失值；
    在所述总损失值未达到预设的收敛条件时，迭代更新所述多分支卷积神经网络模型的初始参数，并触发所述通过所述多分支卷积神经网络模型对所述训练样本进行样本特征提取，得到图像特征空间结果的步骤，直至所述总损失值达到所述预设的收敛条件时，将收敛之后的所述多分支卷积神经网络模型记录为训练完成的所述第一样本检测模型。
  12. 如权利要求11所述的计算机设备，其中，所述将所述训练样本输入含有初始参数的多分支卷积神经网络模型之前，所述处理器执行所述计算机可读指令时还实现如下步骤：
    通过迁移学习，获取训练完成的无监督域训练模型的所有迁移参数，将所有所述迁移参数确定为所述多分支卷积神经网络模型中的所述初始参数。
  13. 如权利要求11所述的计算机设备，其中，所述根据所述图像特征空间结果和与所述训练样本对应的所述第一样本类别，得到第一损失值，包括：
    通过所述多分支卷积神经网络模型将与所述训练样本对应的所述第一样本类别进行向量转换，得到与所述第一样本类别对应的中心向量；所述中心向量包括欧式域中心向量和角度域中心向量；
    通过交叉熵损失算法，根据所述图像特征空间结果和所述欧式域中心向量，得到欧式损失值；
    通过ArcFace损失算法，根据所述图像特征空间结果和所述角度域中心向量，得到角度损失值；
    对所述欧式损失值和所述角度损失值进行加权处理，得到所述第一损失值。
  14. 如权利要求13所述的计算机设备，其中，所述通过ArcFace损失算法，根据所述图像特征空间结果和所述角度域中心向量，得到角度损失值，包括：
    通过所述多分支卷积神经网络模型中的正则化模型，对所述图像特征空间结果进行正则化处理，得到正则化特征向量；
    将所述正则化特征向量与所述角度域中心向量输入角度域损失模型，通过所述角度域损失模型中的所述ArcFace损失算法，得到角度损失值。
  15. 一个或多个存储有计算机可读指令的可读存储介质，其中，所述计算机可读指令被一个或多个处理器执行时，使得所述一个或多个处理器执行如下步骤：
    获取待识别样本；
    将所述待识别样本输入第一样本检测模型进行样本特征提取及语义特征识别，得到待处理样本识别结果；所述待处理样本识别结果包括第一样本检测结果、样本特征空间结果和样本语义空间结果；
    在所述第一样本检测结果与所有第一样本类别均不匹配时，则将所述样本特征空间结果和所述样本语义空间结果输入至第二样本检测模型中；
    通过所述第二样本检测模型对所述样本特征空间结果进行聚类处理，得到第一异常结果，同时对所述样本语义空间结果进行相似度匹配，得到第二异常结果；
    根据所述第一异常结果和所述第二异常结果，确定所述待识别样本的样本类别并输出。
  16. 如权利要求15所述的可读存储介质，其中，所述将所述待识别样本输入第一样本检测模型进行样本特征提取及语义特征识别，得到待处理样本识别结果之后，所述计算机可读指令被一个或多个处理器执行时，使得所述一个或多个处理器还执行如下步骤：
    在所述第一样本检测结果与任一所述第一样本类别匹配时，将匹配的所述第一样本类别确定为所述待识别样本的样本类别。
  17. 如权利要求15所述的可读存储介质，其中，所述将所述待识别样本输入第一样本检测模型进行样本特征提取及语义特征识别之前，所述计算机可读指令被一个或多个处理器执行时，使得所述一个或多个处理器还执行如下步骤：
    获取训练样本集；所述训练样本集包含多个训练样本，一个所述训练样本与一个第一样本类别以及一个第一样本描述关联；
    将所述训练样本输入含有初始参数的多分支卷积神经网络模型；
    通过所述多分支卷积神经网络模型对所述训练样本进行样本特征提取，得到图像特征空间结果；
    通过所述多分支卷积神经网络模型中的嵌入空间模型对所述图像特征空间结果进行语义特征识别，得到语义特征空间结果；所述嵌入空间模型为通过构建图像特征向量与语义特征向量之间的关联关系获得；
    通过K-means聚类算法，对所述图像特征空间结果进行识别，得到训练类别结果，并将所述训练类别结果、所述图像特征空间结果和所述语义特征空间结果确定为样本训练结果；
    根据所述图像特征空间结果和与所述训练样本对应的所述第一样本类别，得到第一损失值；根据所述语义特征空间结果和与所述训练样本对应的所述第一样本描述，得到第二损失值；
    根据所述第一损失值和所述第二损失值，得到总损失值；
    在所述总损失值未达到预设的收敛条件时，迭代更新所述多分支卷积神经网络模型的初始参数，并触发所述通过所述多分支卷积神经网络模型对所述训练样本进行样本特征提取，得到图像特征空间结果的步骤，直至所述总损失值达到所述预设的收敛条件时，将收敛之后的所述多分支卷积神经网络模型记录为训练完成的所述第一样本检测模型。
  18. 如权利要求17所述的可读存储介质，其中，所述将所述训练样本输入含有初始参数的多分支卷积神经网络模型之前，所述计算机可读指令被一个或多个处理器执行时，使得所述一个或多个处理器还执行如下步骤：
    通过迁移学习，获取训练完成的无监督域训练模型的所有迁移参数，将所有所述迁移参数确定为所述多分支卷积神经网络模型中的所述初始参数。
  19. 如权利要求17所述的可读存储介质，其中，所述根据所述图像特征空间结果和与 所述训练样本对应的所述第一样本类别，得到第一损失值，包括：
    通过所述多分支卷积神经网络模型将与所述训练样本对应的所述第一样本类别进行向量转换，得到与所述第一样本类别对应的中心向量；所述中心向量包括欧式域中心向量和角度域中心向量；
    通过交叉熵损失算法，根据所述图像特征空间结果和所述欧式域中心向量，得到欧式损失值；
    通过ArcFace损失算法，根据所述图像特征空间结果和所述角度域中心向量，得到角度损失值；
    对所述欧式损失值和所述角度损失值进行加权处理，得到所述第一损失值。
  20. 如权利要求19所述的可读存储介质，其中，所述通过ArcFace损失算法，根据所述图像特征空间结果和所述角度域中心向量，得到角度损失值，包括：
    通过所述多分支卷积神经网络模型中的正则化模型，对所述图像特征空间结果进行正则化处理，得到正则化特征向量；
    将所述正则化特征向量与所述角度域中心向量输入角度域损失模型，通过所述角度域损失模型中的所述ArcFace损失算法，得到角度损失值。
PCT/CN2020/135337 2020-07-28 2020-12-10 样本类别识别方法、装置、计算机设备及存储介质 WO2021121127A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010738650.1A CN111860674B (zh) 2020-07-28 2020-07-28 样本类别识别方法、装置、计算机设备及存储介质
CN202010738650.1 2020-07-28

Publications (1)

Publication Number Publication Date
WO2021121127A1 true WO2021121127A1 (zh) 2021-06-24

Family

ID=72948380

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/135337 WO2021121127A1 (zh) 2020-07-28 2020-12-10 样本类别识别方法、装置、计算机设备及存储介质

Country Status (2)

Country Link
CN (1) CN111860674B (zh)
WO (1) WO2021121127A1 (zh)

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113378833A (zh) * 2021-06-25 2021-09-10 北京百度网讯科技有限公司 图像识别模型训练方法、图像识别方法、装置及电子设备
CN113627470A (zh) * 2021-07-01 2021-11-09 汕头大学 一种基于零次学习的光纤预警系统未知事件分类方法
CN113657545A (zh) * 2021-08-30 2021-11-16 平安医疗健康管理股份有限公司 用户业务数据的处理方法、装置、设备及存储介质
CN114155420A (zh) * 2021-11-12 2022-03-08 青岛信芯微电子科技股份有限公司 一种场景识别模型训练方法、装置、设备及介质
CN114257553A (zh) * 2021-12-21 2022-03-29 科大讯飞股份有限公司 流量检测方法、装置、电子设备和存储介质
CN114298122A (zh) * 2021-10-22 2022-04-08 腾讯科技（深圳）有限公司 数据分类方法、装置、设备、存储介质及计算机程序产品
CN114998613A (zh) * 2022-06-24 2022-09-02 安徽工业大学 一种基于深度互学习的多标记零样本学习方法
CN115331216A (zh) * 2022-08-11 2022-11-11 北京林业大学 核桃果实成熟度聚类识别方法、装置及系统和存储介质
CN116595978A (zh) * 2023-07-14 2023-08-15 腾讯科技（深圳）有限公司 对象类别识别方法、装置、存储介质及计算机设备
CN116994076A (zh) * 2023-09-28 2023-11-03 中国海洋大学 一种基于双分支相互学习特征生成的小样本图像识别方法
CN117992898A (zh) * 2024-04-07 2024-05-07 腾讯科技（深圳）有限公司 异常检测模型的训练方法、物体异常检测方法和装置
CN114257553B (zh) * 2021-12-21 2024-06-07 科大讯飞股份有限公司 流量检测方法、装置、电子设备和存储介质

Families Citing this family (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111860674B (zh) * 2020-07-28 2023-09-19 平安科技（深圳）有限公司 样本类别识别方法、装置、计算机设备及存储介质
CN112364912B (zh) * 2020-11-09 2023-10-13 腾讯科技（深圳）有限公司 信息分类方法、装置、设备及存储介质
CN112418320B (zh) * 2020-11-24 2024-01-19 杭州未名信科科技有限公司 一种企业关联关系识别方法、装置及存储介质
CN112434746B (zh) * 2020-11-27 2023-10-27 平安科技（深圳）有限公司 基于层次化迁移学习的预标注方法及其相关设备
CN113255701B (zh) * 2021-06-24 2021-10-22 军事科学院系统工程研究院网络信息研究所 一种基于绝对-相对学习架构的小样本学习方法和系统
CN113723341B (zh) * 2021-09-08 2023-09-01 北京有竹居网络技术有限公司 视频的识别方法、装置、可读介质和电子设备
US20230186623A1 (en) * 2021-12-14 2023-06-15 Ping An Technology (Shenzhen) Co., Ltd. Systems and methods for crop disease diagnosis
CN115677346B (zh) * 2022-11-07 2023-09-12 北京赛乐米克材料科技有限公司 彩色锆宝石陶瓷鼻托的制备方法
CN116403077B (zh) * 2023-06-07 2023-08-15 中国科学院自动化研究所 异常检测模型训练方法、异常检测方法、装置及电子设备
CN117313899B (zh) * 2023-11-23 2024-02-23 全芯智造技术有限公司 用于数据处理的方法、设备和介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108154222A (zh) * 2016-12-02 2018-06-12 北京市商汤科技开发有限公司 深度神经网络训练方法和系统、电子设备
CN109359725A (zh) * 2018-10-24 2019-02-19 北京周同科技有限公司 卷积神经网络模型的训练方法、装置、设备及计算机可读存储介质
CN109377494A (zh) * 2018-09-14 2019-02-22 阿里巴巴集团控股有限公司 一种用于图像的语义分割方法和装置
CN110059807A (zh) * 2019-04-26 2019-07-26 腾讯科技（深圳）有限公司 图像处理方法、装置和存储介质
CN111860674A (zh) * 2020-07-28 2020-10-30 平安科技（深圳）有限公司 样本类别识别方法、装置、计算机设备及存储介质

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113168567A (zh) * 2018-10-29 2021-07-23 Hrl实验室有限责任公司 用于小样本转移学习的系统和方法
CN110472652B (zh) * 2019-06-30 2023-06-09 天津大学 基于语义引导的少量样本分类方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108154222A (zh) * 2016-12-02 2018-06-12 北京市商汤科技开发有限公司 深度神经网络训练方法和系统、电子设备
CN109377494A (zh) * 2018-09-14 2019-02-22 阿里巴巴集团控股有限公司 一种用于图像的语义分割方法和装置
CN109359725A (zh) * 2018-10-24 2019-02-19 北京周同科技有限公司 卷积神经网络模型的训练方法、装置、设备及计算机可读存储介质
CN110059807A (zh) * 2019-04-26 2019-07-26 腾讯科技（深圳）有限公司 图像处理方法、装置和存储介质
CN111860674A (zh) * 2020-07-28 2020-10-30 平安科技（深圳）有限公司 样本类别识别方法、装置、计算机设备及存储介质

Cited By (20)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113378833B (zh) * 2021-06-25 2023-09-01 北京百度网讯科技有限公司 图像识别模型训练方法、图像识别方法、装置及电子设备
CN113378833A (zh) * 2021-06-25 2021-09-10 北京百度网讯科技有限公司 图像识别模型训练方法、图像识别方法、装置及电子设备
CN113627470A (zh) * 2021-07-01 2021-11-09 汕头大学 一种基于零次学习的光纤预警系统未知事件分类方法
CN113627470B (zh) * 2021-07-01 2023-09-05 汕头大学 一种基于零次学习的光纤预警系统未知事件分类方法
CN113657545A (zh) * 2021-08-30 2021-11-16 平安医疗健康管理股份有限公司 用户业务数据的处理方法、装置、设备及存储介质
CN113657545B (zh) * 2021-08-30 2024-05-21 平安医疗健康管理股份有限公司 用户业务数据的处理方法、装置、设备及存储介质
CN114298122A (zh) * 2021-10-22 2022-04-08 腾讯科技（深圳）有限公司 数据分类方法、装置、设备、存储介质及计算机程序产品
CN114155420A (zh) * 2021-11-12 2022-03-08 青岛信芯微电子科技股份有限公司 一种场景识别模型训练方法、装置、设备及介质
CN114155420B (zh) * 2021-11-12 2024-03-26 青岛信芯微电子科技股份有限公司 一种场景识别模型训练方法、装置、设备及介质
CN114257553A (zh) * 2021-12-21 2022-03-29 科大讯飞股份有限公司 流量检测方法、装置、电子设备和存储介质
CN114257553B (zh) * 2021-12-21 2024-06-07 科大讯飞股份有限公司 流量检测方法、装置、电子设备和存储介质
CN114998613B (zh) * 2022-06-24 2024-04-26 安徽工业大学 一种基于深度互学习的多标记零样本学习方法
CN114998613A (zh) * 2022-06-24 2022-09-02 安徽工业大学 一种基于深度互学习的多标记零样本学习方法
CN115331216A (zh) * 2022-08-11 2022-11-11 北京林业大学 核桃果实成熟度聚类识别方法、装置及系统和存储介质
CN115331216B (zh) * 2022-08-11 2023-05-23 北京林业大学 核桃果实成熟度聚类识别方法、装置及系统和存储介质
CN116595978B (zh) * 2023-07-14 2023-11-14 腾讯科技（深圳）有限公司 对象类别识别方法、装置、存储介质及计算机设备
CN116595978A (zh) * 2023-07-14 2023-08-15 腾讯科技（深圳）有限公司 对象类别识别方法、装置、存储介质及计算机设备
CN116994076A (zh) * 2023-09-28 2023-11-03 中国海洋大学 一种基于双分支相互学习特征生成的小样本图像识别方法
CN116994076B (zh) * 2023-09-28 2024-01-19 中国海洋大学 一种基于双分支相互学习特征生成的小样本图像识别方法
CN117992898A (zh) * 2024-04-07 2024-05-07 腾讯科技（深圳）有限公司 异常检测模型的训练方法、物体异常检测方法和装置

Also Published As

Publication number Publication date
CN111860674B (zh) 2023-09-19
CN111860674A (zh) 2020-10-30

Similar Documents

Publication Publication Date Title
WO2021121127A1 (zh) 样本类别识别方法、装置、计算机设备及存储介质
WO2021120752A1 (zh) 域自适应模型训练、图像检测方法、装置、设备及介质
WO2023087558A1 (zh) 基于嵌入平滑图神经网络的小样本遥感图像场景分类方法
WO2021135499A1 (zh) 损伤检测模型训练、车损检测方法、装置、设备及介质
WO2020000688A1 (zh) 财务风险验证处理方法、装置、计算机设备及存储介质
WO2020077895A1 (zh) 签约意向判断方法、装置、计算机设备和存储介质
WO2021135816A1 (zh) 识别图像中文本的方法、装置和系统
WO2022077646A1 (zh) 一种用于图像处理的学生模型的训练方法及装置
WO2021114809A1 (zh) 车辆损伤特征检测方法、装置、计算机设备及存储介质
US10474872B2 (en) Fingerprint matching using virtual minutiae
US20220076074A1 (en) Multi-source domain adaptation with mutual learning
WO2021189960A1 (zh) 对抗网络训练、医疗数据补充方法、装置、设备及介质
WO2022057309A1 (zh) 肺部特征识别方法、装置、计算机设备及存储介质
CN116597384B (zh) 基于小样本训练的空间目标识别方法、装置及计算机设备
WO2022052459A1 (zh) 一种图像类别检测方法、系统、电子设备及存储介质
WO2022166362A1 (zh) 一种基于隐空间学习和流行约束的无监督特征选择方法
WO2022134580A1 (zh) 证件信息的获取方法及装置、存储介质、计算机设备
CN113837151A (zh) 表格图像处理方法、装置、计算机设备及可读存储介质
CN113065409A (zh) 一种基于摄像分头布差异对齐约束的无监督行人重识别方法
CN116363357A (zh) 基于mim和对比学习的半监督语义分割方法及装置
KR20220116110A (ko) 인공 신경망의 추론 데이터에 대한 신뢰도를 판단하는 방법
CN116910571A (zh) 一种基于原型对比学习的开集域适应方法及系统
CN114255381B (zh) 图像识别模型的训练方法、图像识别方法、装置及介质
CN115439710A (zh) 基于组合迁移学习的遥感样本标注方法
Jun et al. Two-view correspondence learning via complex information extraction

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20900804

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20900804

Country of ref document: EP

Kind code of ref document: A1