WO2019114523A1 - 分类训练方法、服务器及存储介质 - Google Patents

分类训练方法、服务器及存储介质 Download PDF

Info

Publication number
WO2019114523A1
WO2019114523A1 PCT/CN2018/117158 CN2018117158W WO2019114523A1 WO 2019114523 A1 WO2019114523 A1 WO 2019114523A1 CN 2018117158 W CN2018117158 W CN 2018117158W WO 2019114523 A1 WO2019114523 A1 WO 2019114523A1
Authority
WO
WIPO (PCT)
Prior art keywords
sketch
classification
model
feature
real
Prior art date
Application number
PCT/CN2018/117158
Other languages
English (en)
French (fr)
Inventor
黄飞
马林
刘威
Original Assignee
腾讯科技（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 腾讯科技（深圳）有限公司 filed Critical 腾讯科技（深圳）有限公司
Priority to EP18887500.9A priority Critical patent/EP3726426A4/en
Publication of WO2019114523A1 publication Critical patent/WO2019114523A1/zh
Priority to US16/696,361 priority patent/US11017220B2/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • 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
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/254—Fusion techniques of classification results, e.g. of results related to same input data
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G06N20/10—Machine learning using kernel methods, e.g. support vector machines [SVM]
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • G06V10/809—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of classification results, e.g. where the classifiers operate on the same input data
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
    • G06V30/10—Character recognition
    • G06V30/19—Recognition using electronic means
    • G06V30/191—Design or setup of recognition systems or techniques; Extraction of features in feature space; Clustering techniques; Blind source separation
    • G06V30/19147—Obtaining sets of training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
    • G06V30/10—Character recognition
    • G06V30/19—Recognition using electronic means
    • G06V30/191—Design or setup of recognition systems or techniques; Extraction of features in feature space; Clustering techniques; Blind source separation
    • G06V30/19167—Active pattern learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
    • G06V30/10—Character recognition
    • G06V30/19—Recognition using electronic means
    • G06V30/191—Design or setup of recognition systems or techniques; Extraction of features in feature space; Clustering techniques; Blind source separation
    • G06V30/19173—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
    • G06V30/10—Character recognition
    • G06V30/32—Digital ink
    • G06V30/333—Preprocessing; Feature extraction
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
    • G06V30/10—Character recognition
    • G06V30/32—Digital ink
    • G06V30/36—Matching; Classification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
    • G06V30/40—Document-oriented image-based pattern recognition
    • G06V30/41—Analysis of document content
    • G06V30/413—Classification of content, e.g. text, photographs or tables

Definitions

  • the present application relates to the field of information processing technologies, and in particular, to a classification training method, a server, and a storage medium.
  • Sketch recognition can be applied to many fields, such as early childhood education. Type recognition and the same category of search based on hand-drawn sketches are very important for the growth of children's divergent thinking and graphic understanding. It can also be used for other graphics. Retrieve in the system.
  • the user can input the hand-drawn sketch through the terminal device and send it to the background server, and the background server receives the received message according to a pre-trained classifier such as a Support Vector Machine (SVM) or a pre-trained classification network.
  • a pre-trained classifier such as a Support Vector Machine (SVM) or a pre-trained classification network.
  • SVM Support Vector Machine
  • Hand-drawn sketches for type recognition. In order to ensure the accuracy of hand-drawn sketch recognition, it is necessary to ensure the accuracy of the pre-trained classifier or classification network. Therefore, the process of training the classifier or classifying the network is more important.
  • the training is mainly based on the feature information of a large number of pre-marked sketches.
  • the trained classifier or classification network may have more or less over-fitting or under-fitting problems, which in turn affects the accuracy of the type recognition of the hand-drawn sketches.
  • a classification training method a server, and a storage medium are provided.
  • the embodiment of the present application provides a classification training method, including:
  • the server determines a sketch classification model, where the sketch classification model includes a first feature extraction module and a first classification module; and a second feature analysis model that determines an output result of the second feature extraction module, the second feature extraction
  • the module belongs to the real graph classification model
  • the server selects a training set that includes sketches of multiple categories
  • the server calculates a function value of the first loss function of the sketch classification model according to the first category processing result and the second sketch feature analysis result;
  • the server adjusts the first fixed parameter value in the sketch classification model according to the function value of the first loss function.
  • the embodiment of the present application provides one or more storage media, the storage medium storing computer readable instructions adapted to be loaded by a processor and to perform a classification training method as described in the embodiments of the present application.
  • An embodiment of the present application provides a server, including one or more processors and a memory, where the computer stores computer readable instructions, and the one or more processors are configured to implement respective computer readable instructions.
  • Computer readable instructions are for loading and executing a classification training method as described in embodiments of the present application by one or more processors.
  • FIG. 2 is a schematic diagram of calculating function values of respective loss functions in one embodiment of the present application.
  • FIG. 3 is a schematic diagram of a sketch classification model and a real graph classification model in one embodiment of the present application
  • FIG. 5 is a schematic diagram of classification of a sketch classification model and a real image classification model in an application embodiment of the present application
  • FIG. 6 is a flowchart of a classification training method provided by an application embodiment of the present application.
  • FIG. 7 is a block diagram of a classification training apparatus according to an embodiment of the present application.
  • FIG. 8 is a block diagram of another classification training apparatus according to an embodiment of the present application.
  • FIG. 9 is a block diagram of a server provided by an embodiment of the present application.
  • the embodiment of the present application provides a classification training method, which mainly trains a sketch classification model and a real image classification model according to a sketch of a marked category and a real map of the marked category.
  • the classification training device passes Train as follows:
  • the sketch classification model includes a first feature extraction module and a first classification module, and a second feature analysis model for determining an output result of the second feature extraction module, wherein the second feature extraction module belongs to the real image classification a model; a training set is selected, the training set includes a plurality of categories of sketches; the category of the training set is determined according to the sketch classification model to obtain a first category processing result, and the second feature analysis model is used to extract the sketch extracted by the first feature extraction module The feature is analyzed to obtain the second sketch feature analysis result; according to the first category processing result and the second sketch feature analysis result, the function value of the first loss function of the sketch classification model is calculated; and the sketch classification model is adjusted according to the function value of the first loss function The first fixed parameter value in .
  • the sketch classification model and the real graph classification model obtained by the training may be applied to, but not limited to, the following scenarios: the user may input the hand-drawn sketch through the terminal device and send it to the background server; the background server may pass the pre-trained sketch classification model. Determine the category to which the hand-drawn sketch received by the background server belongs; and retrieve the real image corresponding to the hand-drawn sketch according to the pre-trained sketch classification model and the real image classification model.
  • a classification model such as the sketch classification model
  • another classification model such as the real graph classification model
  • the useful information in the classification process that is, the feature analysis model (such as the second feature analysis model) that analyzes the features extracted by the feature extraction module in another classification model, so that the adjusted sketch classification model and the real image classification model are Classification calculations are more accurate.
  • the embodiment of the present application provides a classification training method, which is a method performed by the classification training device.
  • the flowchart is shown in FIG. 1
  • the schematic diagram is as shown in FIG. 2 , and includes:
  • Step 101 Determine a sketch classification model, where the sketch classification model includes a first feature extraction module and a first classification module; and determine a second feature analysis model for analyzing an output result of the second feature extraction module, where the second feature extraction module Belongs to the real graph classification model.
  • the sketch classification model is used to identify the category of the sketch, and the sketch refers to the image that the user draws through the electronic device, or the image that the user manually draws.
  • the sketch classification model generally includes a first feature extraction module that performs feature extraction on the sketch, and a first classification module that classifies according to the extracted sketch features, such as a classification model based on a convolutional neural network, or a classification model such as SVM.
  • the classification training device determines that the sketch classification model specifically includes determining a structure of the sketch classification model and a corresponding fixed parameter value, wherein the structure of the sketch classification model specifically includes the structure of the first feature extraction module and the first classification module, and The corresponding fixed parameter value specifically includes the parameter value of the fixed parameter used by the first feature extraction module and the first classification module in the calculation process.
  • the real map classification model is used to identify the category of the real map, and the real map refers to the physical image acquired by the camera or the camera.
  • the real graph classification model generally includes a second feature extraction module that performs feature extraction on the real graph, and a second classification module that classifies according to the extracted real graph features.
  • the second classification module is a classification model based on a convolutional neural network, or a classification model such as SVM.
  • the second feature analysis model is configured to analyze the feature of the real image extracted by the second feature extraction module (ie, the output result) in the process of identifying the category of the real image by the real image classification model, and specifically, the feature of the real image may be determined to belong to Real map.
  • the determining the second feature analysis model by the classification training device specifically includes determining a structure of the second feature analysis model and a parameter value of the fixed parameter used by the second feature analysis model in the calculation process.
  • the structure of the sketch classification model and the real image classification model may be the same or different. If the sketch classification model and the real graph classification model have the same structure, the parameter values of the fixed calculation parameters of the respective calculation sub-modules constituting the sketch classification model, and the parameter values of the fixed calculation parameters of the respective calculation sub-modules constituting the real graph classification model ( Abbreviated as fixed parameter values).
  • the fixed calculation parameter refers to the parameters used in the calculation process that do not need to be assigned at any time, such as weights, angles and the like.
  • the sketch classification model and the real graph classification model are both Dense Net models, and include the same number of Dense Blocks and prediction modules.
  • three depth blocks are taken as an example. These depth blocks are connected in series by transition layers.
  • the depth block and the transition layer belong to the feature extraction module, each depth block performs convolution calculation, and outputs a certain number of feature maps, and the transition layer connected with the depth block can output the feature map of the corresponding depth block output.
  • the dimension is dimension-reduced; the prediction module mainly maps the extracted features to the class probability distribution of the fixed dimension after passing through the plurality of depth blocks and the plurality of transition layers, and predicts the category of an image, belonging to the classification module.
  • the fixed calculation parameters of the depth block such as the number of output feature maps, the convolution kernel size, etc.
  • the fixed calculation parameters of the transition layer such as a dimensional reduction multiple, It can also be different.
  • the classification training device can acquire the sketch classification model and the real image classification model that have been trained in some systems, and initiate steps 102 to 105 of the embodiment.
  • the classification training device can obtain a sketch classification model according to the sketch of the marked category, and obtain a real graph classification model according to the real map of the marked category, and classify the currently trained sketch.
  • the model and the real graph classification model initiate the execution of steps 102 to 105 of the present embodiment.
  • the classification training device can train the sketch classification model and the real image classification model through the loss function associated with the initial model of each classification. specifically:
  • the initial model for determining the classification of the sketch is specifically for determining the initial model of the sketch classification and the initial value of the fixed parameter
  • determining the initial model of the real graph classification is specifically determining the initial model of the real graph classification and the initial value of the fixed parameter.
  • the category of the sketch of the marked category is determined, and the category of the real map of the marked category is determined according to the initial model of the real graph classification, and the initial classification result is obtained.
  • the initial category of the sketch of the marked category and the initial category of the real map of the marked category may be included in the initial classification result.
  • the initial model of the classification adjusts the fixed parameter values in the initial model of the real graph classification according to the function value of the fourth loss function to obtain a sketch classification model and a real graph classification model.
  • Step 102 Select a training set, and not only the sketches of the multiple categories may be included in the training set, but also the real map of each category may be included. That is to say, the training set can include sketches and real maps of multiple categories, and each category can correspond to multiple sketches and multiple real images. Each image in the training set is marked as follows: the mark of the sketch or the mark of the real picture, and the mark of the category corresponding to the image.
  • the classification training device may preprocess each image in the training set before performing step 103.
  • the pre-processing process may include: performing scaling processing on each image, or cropping processing, so that the processed images are the same size, so that the calculation when determining the category of each image in step 103 can be simplified.
  • the pre-processing process may further include: performing enhancement of the main image on each image, so that the main image in each image is clearer and not blurred, so that when step 103 is performed, the unclearness of the main image is eliminated and the image category is determined.
  • the influence of the process where the subject image refers to the main image in an image, not the background image, such as a character, object, etc. included in an image.
  • Step 103 Determine a category of the training set according to the sketch classification model determined in step 101 above to obtain a first category processing result. And in the process of determining the category of the sketch by the sketch classification model, according to the second feature analysis model, the feature of the sketch extracted by the first feature extraction module in the sketch classification model is analyzed to obtain the second sketch feature analysis result.
  • the analysis of the features of the sketch may include determining whether the feature of the sketch belongs to the sketch. In other embodiments, other analysis processing may also be included, and no one example is given here.
  • the obtained first category processing result may specifically include a category of each sketch in the training set determined by the sketch classification model, for example, the training set selected in the above step 102 includes n sketches, which may be “airplane”, “trees”
  • n sketches which may be “airplane”, “trees”
  • the categories of n sketches are determined according to the sketch classification model, and the categories C1, C2, ..., Cn of n sketches are obtained.
  • the second sketch feature analysis result may specifically include the second feature analysis model determining whether the feature of the sketch extracted by the first feature extraction module belongs to the sketch.
  • Step 104 Calculate a function value of the first loss function of the sketch classification model according to the first category processing result and the second sketch feature analysis result obtained in the above step 103.
  • the first loss function includes: a loss function associated with the first classification module, and a loss function of the second feature analysis model for sketch feature analysis.
  • the loss function associated with the first classification module may be obtained according to the first category processing result, and may specifically be a cross entropy loss function, which is used to represent the difference between the category determined by the first classification module and the actual category of the sketch, that is, the error. .
  • the loss function associated with the first classification module may also be a sorting loss function for representing the similar feature sorting process.
  • the loss function in .
  • the first classification module determines the feature similarity between the sketch 1 and the sketch 2, which is greater than the feature similarity between the sketch 1 and the sketch 2, thereby determining the categories of the sketch 1 and the sketch 2
  • the sorting loss function can represent the difference between the sorting determined in the feature similarity sorting process and the sorting of the actual feature similarity.
  • the loss function of the second feature analysis model for the feature analysis of the sketch is obtained according to the second sketch feature analysis result, and is used to represent the analysis result obtained by analyzing the feature of the sketch extracted by the first feature extraction module by the second feature analysis model, and The difference between the actual characteristics of the sketch.
  • Step 105 Adjust a first fixed parameter value in the sketch classification model according to a function value of the first loss function.
  • the first fixed parameter value is a fixed parameter that is used by the first feature extraction module and the first classification module included in the sketch classification model, and does not need to be assigned at any time, such as weights, angles, and the like. Parameter value. If the calculated function value of the first loss function is large, such as greater than the preset value, then the first fixed parameter value needs to be changed, such as increasing the weight value of a certain weight or decreasing the angle value of an angle. And so that the function value of the first loss function calculated according to the adjusted first fixed parameter value is decreased.
  • FIG. 1 shows one specific implementation manner.
  • the classification training device can also adjust the fixed parameter values in the second feature analysis model.
  • the classification training device determines the real graph classification model, specifically including determining the structure of the sketch classification model and the corresponding fixed parameter values, and the real graph classification model includes the second feature extraction module and the second classification module; determining the real graph classification model
  • the feature of the real graph extracted by the second feature extraction module is analyzed to obtain the first real graph feature analysis result; then the result is analyzed according to the first real graph feature.
  • the first real graph feature analysis result may specifically include the second feature analysis model determining whether the feature of the real graph extracted by the second feature extraction module belongs to the real graph.
  • the second anti-loss function may include: a loss function of the second feature analysis model on the feature analysis of the real graph, which may be obtained according to the first real graph feature analysis result; and a loss function of the second feature analysis model for the sketch feature analysis may be based on The second sketch feature analysis results were obtained.
  • the loss function of the second feature analysis model on the feature analysis of the real graph is used to represent the analysis result obtained by analyzing the feature of the real graph extracted by the second feature extraction module by the second feature analysis model, and the actual feature of the real graph. The difference between them.
  • the above steps 103 to 105 are the process of adjusting the first fixed parameter value by the classification training device after processing the respective sketches in the training set for the sketch classification model determined in the above step 101. In practical applications, it is necessary to continuously perform the above steps 103 to 105 until the adjustment of the first fixed parameter value satisfies a certain stop condition.
  • the classification training device further needs to determine whether the current adjustment of the first fixed parameter value satisfies the preset stop condition, and if so, ends the process; if not, the target is The sketch classification model after adjusting the first fixed parameter value is returned to the steps of performing steps 103 to 105 above. That is, the step of obtaining the first category processing result and the second sketch feature analysis result, calculating the function value of the first loss function of the sketch classification model, and adjusting the first fixed parameter value are performed.
  • the preset stop condition includes, but is not limited to, any one of the following conditions: the first difference between the currently adjusted first fixed parameter value and the last adjusted first fixed parameter value is less than the first threshold, that is, the adjusted A fixed parameter value reaches convergence; and the number of adjustments to the first fixed parameter value reaches a preset number of times. It can be understood that the stop condition here refers to the condition for stopping the adjustment of the first fixed parameter value.
  • classification training device may further adjust the second fixed parameter value of the real graph classification model by the following steps.
  • the flowchart is shown in FIG. 4, and the schematic diagram is as shown in FIG. 2, including:
  • Step 201 Determine a first feature analysis model that analyzes an output result of the first feature extraction module.
  • the first feature analysis model is used to analyze the feature (ie, the output result) of the sketch extracted by the first feature extraction module in the process of identifying the sketch class by the sketch classification model, and specifically, the feature of the sketch belongs to the sketch.
  • Step 202 Determine a category of the real graph in the training set according to the real graph classification model to obtain a second category processing result; and analyze, according to the first feature analyzing model, a feature of the real graph extracted by the second feature extracting module to obtain a second real graph. Feature analysis results.
  • the obtained second category processing result may specifically include categories of respective real maps in the training set determined by the real graph classification model.
  • the second real graph feature analysis result may specifically include the first feature analysis model determining whether the feature of the real graph extracted by the second feature extraction module belongs to the real graph.
  • Step 203 Calculate a function value of a second loss function of the real graph classification model according to the second category processing result and the second real graph feature analysis result, where the second loss function includes a loss function associated with the second classification module, And the loss function of the first feature analysis model for the feature analysis of the real graph.
  • the loss function associated with the second classification module can be obtained according to the second category processing result. Specifically, it may be a cross entropy loss function for indicating a difference between a category determined by the second classification module and an actual category of the real graph; if the second classification module determines a method for sorting similar features in the process of determining the real graph category, the method
  • the loss function associated with the second classification module can also be a ranking loss function for representing the loss function in the similar feature ranking process.
  • the loss function of the first feature analysis model for real feature analysis can be obtained according to the second real image feature analysis result, and is used to represent the analysis result obtained by analyzing the feature of the real image extracted by the second feature extraction module by the first feature analysis model. , the difference between the actual features of the real map.
  • Step 204 Adjust a second fixed parameter value in the real graph classification model according to a function value of the second loss function.
  • the second fixed parameter value is a fixed parameter that is used by the second feature extraction module and the second classification module included in the real graph classification model, and does not need to be assigned at any time, such as weights, angles, and the like.
  • the parameter value If the calculated function value of the second loss function is large, such as greater than the preset value, then the second fixed parameter value needs to be changed, such as increasing the weight value of a certain weight, or decreasing the angle value of an angle, etc. So that the function value of the second loss function calculated according to the adjusted second fixed parameter value is decreased.
  • the classification training device can also adjust the fixed parameter values in the first feature analysis model. Specifically, the classification training device analyzes the feature of the sketch extracted by the first feature extraction module according to the first feature analysis model to obtain a first sketch feature analysis result; and then analyzes the first sketch feature analysis result and the second real image feature feature according to the first sketch feature As a result, the function value of the first anti-loss function of the first feature analysis model is calculated, and the fixed parameter value of the first feature analysis model is adjusted according to the function value of the first anti-loss function.
  • the first sketch feature analysis result may specifically include the first feature analysis model determining whether the feature of the sketch extracted by the first feature extraction module belongs to a sketch.
  • the first anti-loss function may include: a loss function of the first feature analysis model on the feature analysis of the real graph, which may be obtained according to the second real graph feature analysis result; and the loss function of the first feature analysis model for the sketch feature analysis may be based on The first sketch feature analysis results were obtained.
  • the loss function of the first feature analysis model for the feature analysis of the sketch is used to represent the difference between the analysis result obtained by analyzing the feature of the sketch extracted by the first feature extraction module by the first feature analysis model, and the actual feature of the sketch. .
  • steps 202 to 204 are processes for adjusting the second fixed parameter value by the classification training device after processing the real image in the training set for the real image classification model. In practical applications, it is necessary to continuously perform the above steps 202 to 204 until the adjustment of the second fixed parameter value satisfies a certain stop condition.
  • the classification training device needs to determine whether the current adjustment of the second fixed parameter value satisfies the preset stop condition, and if so, ends the flow; if not, adjusts the second
  • the preset stop condition includes, but is not limited to, any one of the following conditions: the first difference between the currently adjusted second fixed parameter value and the last adjusted second fixed parameter value is less than the second threshold, that is, the adjusted The two fixed parameter values reach convergence; and the number of adjustments to the second fixed parameter value reaches a preset number of times.
  • steps 202 to 204 and steps 103 to 105 may be alternately performed.
  • the first fixed parameter value of the sketch classification model may be adjusted and the second feature analysis model may be adjusted. Parameter values, that is, steps 103 to 105 are performed; in another adjustment process, the second fixed parameter value of the real graph classification model is adjusted and the fixed parameter value of the first feature analysis model is adjusted, that is, steps 202 to 204 are performed; In the process, the first fixed parameter value of the sketch classification model is adjusted and the fixed parameter value of the second feature analysis model is adjusted, that is, steps 103 to 105 are performed, and so on.
  • the classification training device obtains the adjusted sketch classification model and the real graph classification model by the method of the above embodiment, and in the actual application of the adjusted sketch classification model and the real graph classification model: in one case, classification The training device can first obtain the sketch to be classified (such as the sketch to be classified by the user through the terminal device), and then classify the classified sketch according to the adjusted sketch classification model, and obtain the category of the sketch to be classified, thereby realizing the classification of the sketch.
  • classification The training device can first obtain the sketch to be classified (such as the sketch to be classified by the user through the terminal device), and then classify the classified sketch according to the adjusted sketch classification model, and obtain the category of the sketch to be classified, thereby realizing the classification of the sketch.
  • the classification training device may first obtain a sketch to be classified (such as a sketch to be classified input by the user through the terminal device), and obtain each real image stored in the classification training device; and then classify according to the adjusted sketch classification model.
  • the sketch is classified to obtain the category of the sketch to be classified, and the stored real maps are classified according to the adjusted real graph classification model to obtain the categories of the real maps; finally, the real maps with the same categories as the sketches to be classified are selected.
  • a sketch to be classified such as a sketch to be classified input by the user through the terminal device
  • the stored real maps are classified according to the adjusted real graph classification model to obtain the categories of the real maps
  • the real maps with the same categories as the sketches to be classified are selected.
  • the classification training device first selects the training set, and determines the category of the sketch in the training set according to the sketch classification model to obtain the first category processing result, and can analyze the model according to the second feature.
  • the sketch feature extracted by the feature extraction model is analyzed, and the second sketch analyzes the result; then the function value of the first loss function is obtained according to the first category processing result and the second sketch analysis result; finally, the sketch is classified according to the function value of the first loss function
  • the first fixed parameter value of the model is adjusted.
  • the classification training device further determines the category of the real map in the training set according to the real graph classification model, obtains the second category processing result, and can analyze the real graph feature extracted by the second feature extraction module according to the first feature analysis model, and obtain The second real graph analyzes the result; then obtains the function value of the second loss function according to the second category processing result and the second real graph analysis result; finally, the second fixed parameter value of the real graph classification model according to the function value of the second loss function Make adjustments.
  • a classification model such as the sketch classification model
  • another classification model such as the real graph classification model
  • the useful information in the classification process that is, the feature analysis model (such as the second feature analysis model) that analyzes the features extracted by the feature extraction module in another classification model, so that the adjusted sketch classification model and the real image classification model are Classification calculations are more accurate.
  • the sketch classification model and the real graph classification model can use the convolutional neural network of the same structure, respectively, as CNN_1 and CNN_2; the first feature extraction module and the first classification module included in the sketch classification model are respectively recorded as CNN11 and CNN12, the first feature analysis model is specifically a sketch discriminator D_1; the second feature extraction module included in the real map classification model and The second classification module is respectively recorded as CNN21 and CNN22, and the second feature analysis model is specifically the real map discriminator D_2.
  • the classification training method in this embodiment can be implemented by the following steps.
  • the flowchart is as shown in FIG. 6, and includes:
  • Step 301 determining a sketch classification model CNN_1 and a real map classification model CNN_2, and a sketch discriminator D_1 and a real graph discriminator D_2, specifically including determining an initial value of a structure of each model and a fixed parameter.
  • Step 302 Select a training set, and include image pairs of multiple categories in the training set, and the image pair of each category includes a sketch and a real image of the same category, specifically Where Si is a sketch image of a certain category i, and Ii is a real image of the category i.
  • Step 303 Determine the category of the sketch in the training set according to the sketch classification model CNN_1, and determine the category of the real map according to the real map classification model CNN_2.
  • the sketch feature extracted by the sketch feature extraction module CNN11 in the sketch classification model CNN_1 is identified by the sketch discriminator D_1 to obtain the sketch feature discrimination result 11; the real image classification model CNN_2 is true by the sketch discriminator D_1 The real map features extracted by the graph feature extraction module CNN21 are discriminated to obtain a real graph feature discriminating result 12.
  • the real graph feature extracted by the real graph feature extraction module CNN21 in the real graph classification model CNN_2 is discriminated by the real graph discriminator D_2 to obtain the real graph feature discriminating result 21; the sketch in the sketch scoring model CNN_1 is obtained by the real graph discriminator D_2 The sketch features extracted by the feature extraction module CNN11 are identified to obtain a sketch feature identification result 22.
  • Step 304 first fix the sketch classification model CNN_1 and the real map discriminator D_2, and adjust the fixed parameter value of the real graph classification model CNN_2 and the fixed parameter value of the sketch discriminator D_1, so that the adjustment of the real graph classification model CNN_2 draws a sketch Useful information of the classification model CNN_1 in the classification process.
  • the classification training device calculates the confrontation function of the sketch discriminator D_1 according to the following formula 1, and the sketch feature identification result 11 and the real map feature discrimination result 12 obtained in the above step 303.
  • the function value adjusts the fixed parameter value of the sketch discriminator D_1 according to the function value.
  • the category of the real graph determined by the real graph classification model CNN_2 in the above step 303 and the real graph feature discriminating result 12 the function value of the loss function of the real graph classification model CNN_2 is calculated, and the real graph classification model is adjusted according to the function value.
  • the method may include: a loss function associated with the real graph classification module CNN22 in the real graph classification model CNN_2, such as a cross entropy loss function.
  • the loss function of the sketch discriminator D_1 for discriminating the real graph feature can be obtained according to the real graph feature discriminating result 12.
  • the related loss function It can be obtained by the following formula 3 and the category of the real map determined by the real graph classification model CNN_2 in the above step 303.
  • log(D_1(CNN_2(I i ))) can represent the loss function of the sketch discriminator D_1 for distinguishing the real graph features.
  • Step 305 the real map classification model CNN_2 and the sketch discriminator D_1 are fixed, and the fixed parameter value of the sketch classification model CNN_1 and the fixed parameter value of the real map discriminator D_2 are adjusted, so that the adjustment of the sketch classification model CNN_1 draws on the real map.
  • the classification training device calculates the confrontation function of the real graph discriminator D_2 according to the following formula 4, and the real graph feature discrimination result 21 and the sketch feature discriminating result 22 obtained in the above step 303.
  • the function value adjusts the fixed parameter value of the real image discriminator D_2 according to the function value.
  • the category classification and sketch feature identification result 22 determined by the sketch classification model CNN_1 in the above step 303 the function value of the loss function of the sketch classification model CNN_1 is calculated, and the fixed parameter value of the sketch classification model CNN_1 is adjusted according to the function value.
  • the loss function related to the sketch classification module CNN12 in the sketch classification model CNN_1 such as the cross entropy loss function, may be included.
  • the loss function of the real graph discriminator D_2 for discriminating the sketch feature can be obtained according to the sketch feature discriminating result 22.
  • the related loss function It can be obtained by the following formula 6 and the category of the sketch.
  • log(D_2(CNN_1(S i )))) may represent a loss function of the real graph discriminator D_2 for discriminating the sketch features.
  • Step 306 after performing the above steps 301 to 305, determine whether the adjustment of the fixed parameter values of the sketch classification model CNN_1 and the real image classification model CNN_2 satisfies the preset condition, and if so, ends the process; if not, adjusts The subsequent sketch classification model CNN_1 and the real map classification model CNN_2, and the adjusted sketch discriminator D_1 and the real map discriminator D_2, return to perform the above steps 303 to 305.
  • the embodiment of the present application further provides a classification training device, and a schematic structural diagram thereof is shown in FIG. 7 , which may specifically include:
  • a model determining unit 10 configured to determine a sketch classification model, where the sketch classification model includes a first feature extraction module and a first classification module; and a second feature analysis model that determines an output result of the second feature extraction module,
  • the second feature extraction module belongs to a real image classification model.
  • the model determining unit 10 is specifically configured to determine an initial model of the sketch classification of the same structure and an initial model of the real graph classification, and determine a real map of the marked category and the marked category; the initial classification according to the sketch
  • the model determines a category of the sketch of the marked category, determines a category of the real map of the marked category according to an initial model of the real graph classification, and obtains an initial classification result; and calculates an initial model related to the sketch classification according to the initial classification result a function value of the third loss function, and a function value of the fourth loss function associated with the initial model of the real graph classification, and adjusting the initial model of the sketch classification according to the function value of the third loss function
  • the fixed parameter value is adjusted according to the function value of the fourth loss function to adjust the fixed parameter value in the initial model of the real graph classification to obtain the sketch classification model and the real graph classification model.
  • the training set unit 11 is for selecting a training set, the training set including a plurality of categories of sketches.
  • the training set also includes real maps of the corresponding categories.
  • the processing unit 12 is configured to determine, according to the sketch classification model determined by the model determining unit 10, the training set unit 11 to select a category of the sketch in the training set to obtain a first category processing result; according to the second feature analyzing model, A feature extracted by the feature extraction module is analyzed to obtain a second sketch feature analysis result;
  • the function value calculation unit 13 is configured to calculate a function value of the first loss function of the sketch classification model according to the first category processing result and the second sketch feature analysis result obtained by the processing unit 12, wherein the first The loss function includes a loss function associated with the first classification module and a loss function of the feature analysis of the sketch by the second feature analysis model.
  • the adjusting unit 14 is configured to adjust the first fixed parameter value in the sketch classification model according to the function value of the first loss function calculated by the function value calculating unit 13.
  • the training set selected by the training set unit 11 may further include a real map of a corresponding category
  • the model determining unit 10 may further determine a real graph classification model, where the real graph classification model includes a second feature extraction module and a second classification module
  • the processing unit 12 is further configured to: when the real map classification model determines the category of the real map, according to the second feature analysis model, the reality extracted by the second feature extraction module
  • the feature of the graph is analyzed to obtain a first real graph feature analysis result
  • the function value calculating unit 13 is further configured to calculate the second feature analysis model according to the first real graph feature analysis result and the second sketch feature analysis result.
  • the adjusting unit 14 is further configured to adjust the fixed parameter value of the second feature analysis model according to the function value of the second anti-loss function.
  • the model determining unit 10 is further configured to determine a first feature analysis model that analyzes an output result of the first feature extraction module
  • the processing unit 12 is further configured to determine the training according to the real image classification model.
  • the category of the centralized real graph obtains the second category processing result; according to the first feature analysis model, the feature of the real graph extracted by the second feature extraction module is analyzed to obtain a second real graph feature analysis result;
  • the function value calculation The unit 13 is further configured to calculate a function value of the second loss function of the real graph classification model according to the second category processing result and the second real graph feature analysis result, where the second loss function includes a loss function associated with the second classification module, and a loss function of the feature analysis of the real feature by the first feature analysis model;
  • the adjusting unit 14 is further configured to adjust the function value according to the function of the second loss function The second fixed parameter value in the real graph classification model.
  • the processing unit 12 is further configured to analyze, according to the first feature analysis model, the feature of the sketch extracted by the first feature extraction module to obtain a first sketch feature analysis result; the function value calculation unit 13, And is further configured to calculate a function value of the first anti-loss function of the first feature analysis model according to the first sketch feature analysis result and the second real graph feature analysis result; the adjusting unit 14 is further configured to be used according to the first A function value against the loss function adjusts the fixed parameter value of the first feature analysis model.
  • the training set unit 11 selects a training set
  • the processing unit 12 determines the category of the training set sketch according to the sketch classification model, obtains the first category processing result, and can analyze the model pair according to the second feature.
  • the sketch feature extracted by the first feature extraction model is analyzed to obtain a second sketch analysis result; then the function value calculation unit 13 obtains the function value of the first loss function according to the first category processing result and the second sketch analysis result; and finally the adjusting unit 14
  • the first fixed parameter value of the sketch classification model is adjusted according to the function value of the first loss function.
  • the processing unit 12 may further determine a category of the real map in the training set according to the real map classification model, and obtain a second category processing result, and may analyze the real graph feature extracted by the second feature extraction module according to the first feature analysis model, and obtain the first The second real graph analysis result; then the function value calculating unit 13 obtains the function value of the second loss function according to the second category processing result and the second real graph analysis result; finally, the adjusting unit 14 classifies the real graph according to the function value of the second loss function.
  • the second fixed parameter value of the model is adjusted.
  • a classification model such as the sketch classification model
  • another classification model such as the real graph classification model
  • the useful information in the classification process that is, the feature analysis model (such as the second feature analysis model) that analyzes the features extracted by the feature extraction module in another classification model, so that the adjusted sketch classification model and the real image classification model are Classification calculations are more accurate.
  • the classification training device may include, in addition to the structure shown in FIG. 7, a judging unit 15 and a classifying unit 16, wherein:
  • the determining unit 15 is configured to determine whether the adjustment of the first fixed parameter value by the adjusting unit 14 satisfies a preset stopping condition, and if not, notify the processing unit 12 of adjusting the first fixed parameter value.
  • the sketch classification model obtains the first category processing result and the second sketch feature analysis result.
  • the preset stop condition may include but is not limited to any one of the following conditions:
  • the first difference between the currently adjusted first fixed parameter value and the last adjusted first fixed parameter value is less than the first threshold; and the number of adjustments to the first fixed parameter reaches a preset number of times.
  • the determining unit 15 is configured to determine whether the adjustment of the second fixed parameter value by the adjusting unit 14 satisfies a preset stopping condition, and if not, notify the processing unit 12 to adjust the second fixed parameter. After the value of the real graph classification model, the second category processing result and the second real graph feature analysis result are obtained.
  • the preset stop condition herein may include, but is not limited to, any one of the following conditions: the first difference between the currently adjusted second fixed parameter value and the last adjusted second fixed parameter value is less than the first threshold; The number of adjustments of the second fixed parameter reaches the preset number of times and the like.
  • the classification unit 16 is configured to obtain a sketch to be classified, and classify the to-be-classified sketch according to the sketch classification model adjusted by the adjusting unit 14 to obtain a category of the sketch to be classified.
  • the classification unit 16 can also obtain a sketch to be classified, and obtain each real map stored in the classification training device; and then classify the classified sketch according to the adjusted sketch classification model adjusted by the adjusting unit 14 to obtain the category of the sketch to be classified, according to the adjusted
  • the real map classification model separately classifies the stored real maps to obtain the categories of the real maps; finally, selects the real maps with the same categories as the sketches to be classified.
  • the embodiment of the present application further provides a server.
  • the schematic diagram of the server is shown in FIG. 9.
  • the server may have a large difference due to different configurations or performances, and may include one or more central processing units (CPUs). 20 (e.g., one or more processors) and memory 21, one or more storage media 22 storing application 221 or data 222 (e.g., one or one storage device in Shanghai).
  • the memory 21 and the storage medium 22 may be short-term storage or persistent storage.
  • the program stored on storage medium 22 may include one or more modules (not shown), each of which may include a series of computer readable instructions in a server.
  • central processor 20 may be arranged to communicate with storage medium 22 to execute a series of computer readable instructions in storage medium 22.
  • the computer readable instructions when executed, may cause one or more processors included in central processor 222 to perform a classification training method.
  • the storage medium 730 can be a non-volatile storage medium.
  • the non-volatile storage medium can be a non-volatile readable storage medium.
  • the application 221 stored in the storage medium 22 includes a classification training application
  • the program may include the model determination unit 10, the training set unit 11, the processing unit 12, and the function value calculation unit 13, in the classification training device described above,
  • the adjustment unit 14, the determination unit 15 and the classification unit 16 are not described herein.
  • the central processor 20 can be configured to communicate with the storage medium 22 to perform a series of operations corresponding to the classification training application stored in the storage medium 22 on the server.
  • the server may also include one or more power sources 23, one or more wired or wireless network interfaces 24, one or more input and output interfaces 25, and/or one or more operating systems 223, such as Windows ServerTM, Mac OS XTM , UnixTM, LinuxTM, FreeBSDTM and more.
  • Each unit module included in the classification training device may be implemented in whole or in part by software, hardware, or a combination thereof.
  • the embodiment of the present application further provides a storage medium storing computer readable instructions adapted to be loaded by a processor and to perform a classification training method as described in various embodiments of the present application.
  • the embodiment of the present application further provides a server, including one or more processors and a memory, where the computer stores computer readable instructions, and the one or more processors are configured to implement the respective computer readable instructions.
  • Computer readable instructions are for loading and executing a classification training method as described in various embodiments of the present application by one or more processors.
  • the storage medium may include a read only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and the like.

Abstract

一种分类训练方法，包括：选定训练集，并根据草图分类模型确定训练集中的草图的类别，得到第一类别处理结果，且可以根据第二特征分析模型对第一特征提取模型提取的草图特征进行分析，得到第二草图分析结果；然后根据第一类别处理结果和第二草图分析结果得到第一损失函数的函数值；最后根据第一损失函数的函数值对草图分类模型的第一固定参数值进行调整。

Description

分类训练方法、服务器及存储介质
本申请要求于2017年12月12日提交中国专利局，申请号为2017113226122，申请名称为“一种分类训练方法、装置及存储介质”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
本申请涉及信息处理技术领域，特别涉及一种分类训练方法、服务器及存储介质。
背景技术
草图识别可以应用于许多领域，比如应用于儿童早期教育中，可以基于手绘草图进行类型识别和同一类别的检索，对儿童的发散思维和图形理解能力的成长非常重要；也可以用于其它的图形检索系统中。
具体地，用户可以通过终端设备输入手绘草图并发送到后台服务器，由后台服务器根据预先训练好的分类器比如支持向量机(Support Vector Machine，SVM)，或者预先训练好的分类网络，对接收的手绘草图进行类型识别。为了保证对手绘草图识别的准确性，就需要保证预先训练好的分类器或分类网络的精确性，因此，训练分类器或分类网络的过程就比较重要。
传统方法中，在训练分类器或分类网络时，主要是根据大量的预先标记好类别的草图的特征信息进行训练得到。但是，会由于草图的训练样本稀缺导致训练的分类器或分类网络或多或少地出现过拟合或者欠拟合的问题，进而影响到对手绘草图的类型识别的准确性。
发明内容
根据本申请提供的的各种实施例，提供一种分类训练方法、服务器及存储介质。
本申请实施例提供一种分类训练方法，包括：
服务器确定草图分类模型，所述草图分类模型中包括第一特征提取模块和第一分类模块；及确定对第二特征提取模块的输出结果进行分析的第二特征分析模型，所述第二特征提取模块属于真实图分类模型；
服务器选定训练集，所述训练集包括多个类别的草图；
服务器根据所述草图分类模型确定所述训练集中草图的类别得到第一类别处理结果；根据所述第二特征分析模型，对第一特征提取模块提取的所述草图的特征进行分析得到第二草图特征分析结果；
服务器根据所述第一类别处理结果及第二草图特征分析结果，计算所述草图分类模型的第一损失函数的函数值；
服务器根据所述第一损失函数的函数值调整所述草图分类模型中的第一固定参数值。
本申请实施例提供一个或多个存储介质，所述存储介质储存有计算机可读指令，所述计算机可读指令适于由处理器加载并执行如本申请实施例所述的分类训练方法。
本申请实施例提供一种服务器，包括一个或多个处理器和存储器，所述存储器中存储有计算机可读指令，所述一个或多个处理器，用于实现各个计算机可读指令，所述计算机可读指令用于由一个或多个处理器加载并执行如本申请实施例所述的分类训练方法。
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。基于本申请的说明书、附图以及权利要求书，本申请的其它特征、目的和优点将变得更加明显。
附图说明
为了更清楚地说明本申请实施例或现有技术中的技术方案，下面将对实施例或现有技术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是本申请一个实施例提供的一种分类训练方法的流程图；
图2是本申请一个实施例中计算各个损失函数的函数值的示意图；
图3是本申请一个实施例中草图分类模型和真实图分类模型的示意图；
图4是本申请一个实施例提供的另一种分类训练方法的流程图；
图5是本申请应用实施例中草图分类模型和真实图分类模型进行分类的示意图；
图6是本申请应用实施例提供的一种分类训练方法的流程图；
图7是本申请实施例提供的一种分类训练装置的框图；
图8是本申请实施例提供的另一种分类训练装置的框图；及
图9是本申请实施例提供的一种服务器的框图。
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，所描述的实施例仅仅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。
本申请的说明书和权利要求书及上述附图中的术语“第一”、“第二”、“第三”“第四”等(如果存在)是用于区别类似的对象，而不必用于描述特定的顺序或先后次序。应该理解这样使用的数据在适当情况下可以互换，以便这里描述的本申请的实施例例如能够以除了在这里图示或描述的那些以外的顺序实施。此外，术语“包括”和“具有”以及他们的任何变形，意图在于覆盖不排它的包含，例如，包含了一系列步骤或单元的过程、方法、系统、产品或设备不必限于清楚地列出的那些步骤或单元，而是可包括没有清楚地列出的或对于这些过程、方法、产品或设备固有的其它步骤或单元。
本申请实施例提供一种分类训练方法，主要是根据已标记类别的草图和已标记类别的真实图，训练得到草图分类模型和真实图分类模型，具体地，本实施例中由分类训练装置通过如下方法进行训练：
确定草图分类模型，草图分类模型中包括第一特征提取模块和第一分类模块，及确定对第二特征提取模块的输出结果进行分析的第二特征分析模型，第二特征提取模块属于真实图分类模型；选定训练集，训练集包括多个类别的草图；根据草图分类模型确定训练集中草图的类别得到第一类别处理结果，根据第二特征分析模型，对第一特征提取模块提取的草图的特征进行分析得到第二草图特征分析结果；根据第一类别处理结果及第二草图特征分析结果，计算草图分类模型的第一损失函数的函数值；根据第一损失函数的函数值调整草图分类模型中的第一固定参数值。
进一步地，训练得到的草图分类模型和真实图分类模型可以应用于但不限于如下场景中：用户可以通过终端设备输入手绘草图并发送到后台服务器；由后台服务器通过预先训练好的草图分类模型，确定后台服务器接收的手绘 草图所属的类别；且可以根据预先训练好的草图分类模型和真实图分类模型，检索到该手绘草图对应的真实图像。
这样，在对某一分类模型(比如草图分类模型)的固定参数值进行调整时，不仅会参考该分类模型对相应图像进行分类的误差，还会借鉴另一分类模型(比如真实图分类模型)在分类过程中的有用信息，即对另一分类模型中特征提取模块提取的特征进行分析的特征分析模型(比如第二特征分析模型)，从而使得调整后的草图分类模型和真实图分类模型的分类计算更准确。
本申请实施例提供一种分类训练方法，为分类训练装置所执行的方法，流程图如图1所示，示意图如图2所示，包括：
步骤101，确定草图分类模型，草图分类模型中包括第一特征提取模块和第一分类模块；及确定对第二特征提取模块的输出结果进行分析的第二特征分析模型，这里第二特征提取模块属于真实图分类模型。
草图分类模型用于识别草图的类别，而草图是指用户通过电子设备绘制的图像，或者用户手动绘制的图像等。具体地，草图分类模型一般包括对草图进行特征提取的第一特征提取模块，和根据提取的草图特征进行分类的第一分类模块，比如基于卷积神经网络的分类模型，或SVM等分类模型。本实施例中，分类训练装置确定草图分类模型具体包括确定草图分类模型的结构和相应的固定参数值，其中，草图分类模型的结构具体包括第一特征提取模块和第一分类模块的结构，而相应的固定参数值具体包括第一特征提取模块和第一分类模块在计算过程中用到的固定参数的参数值。
真实图分类模型用于识别真实图的类别，真实图是指通过照相机或摄像头获取的实物图像。具体地，真实图分类模型一般包括对真实图进行特征提取的第二特征提取模块，和根据提取的真实图特征进行分类的第二分类模块。该第二分类模块如基于卷积神经网络的分类模型，或SVM等分类模型。第二特征分析模型用于在真实图分类模型识别真实图的类别的过程中，对第二特征提取模块提取的真实图的特征(即输出结果)进行分析，具体可以判别该真实图的特征属于真实图。分类训练装置确定第二特征分析模型具体包括确定第二特征分析模型的结构，及第二特征分析模型在计算过程中用到的固定参数的参数值。
本实施例中，草图分类模型和真实图分类模型的结构可以相同，也可以 不同。如果草图分类模型和真实图分类模型的结构相同，则组成草图分类模型的各个计算子模块的固定计算参数的参数值，与组成真实图分类模型的各个计算子模块的固定计算参数的参数值(简称固定参数值)不同。这里固定计算参数是指在计算过程中所用到的不需要随时赋值的参数，比如权重，角度等参数。
例如图3所示，草图分类模型和真实图分类模型都是深度网络(Dense Net)模型，且包括相同数量的深度块(Dense Block)及预测模块，图2中以3个深度块为例说明，这些深度块之间通过过渡层(Transition Layers)串联。其中，深度块和过渡层属于特征提取模块，每个深度块进行卷积计算，并输出一定数量的特征地图(feature map)，而与深度块连接的过渡层可以将对应深度块输出的特征地图的维度进行降维；预测模块主要是将通过多个深度块和多个过渡层之后，提取的特征映射到固定维度的类别概率分布，并预测出某一图像的类别，属于分类模块。但是，草图分类模型和真实图分类模型之间，深度块的固定计算参数，比如输出的特征地图的个数，卷积核大小等可以不同，且过渡层的固定计算参数比如降维倍数等，也可以不同。
可以理解，在一种情况下，分类训练装置在执行本步骤时，可以获取某些系统中已训练好的草图分类模型和真实图分类模型，并发起执行本实施例的步骤102到105。
另一种情况下，分类训练装置在执行本步骤时，可以根据已标记类别的草图训练得到草图分类模型，及根据已标记类别的真实图训练得到真实图分类模型，针对当前训练好的草图分类模型和真实图分类模型，发起执行本实施例的步骤102到105。
在这种情况下，分类训练装置在执行本步骤时，可以通过与各个分类的初始模型相关的损失函数，训练得到草图分类模型和真实图分类模型。具体地：
先确定草图分类的初始模型和真实图分类的初始模型，及确定已标记类别的草图和已标记类别的真实图。其中，确定草图分类的初始模型具体为确定草图分类的初始模型的结构和固定参数的初始值，确定真实图分类的初始模型具体为确定真实图分类的初始模型的结构和固定参数的初始值。
然后根据草图分类的初始模型确定已标记类别的草图的类别，根据真实 图分类的初始模型确定已标记类别的真实图的类别，得到初始分类结果。在初始分类结果中可以包括已标记类别的草图的初始类别，及已标记类别的真实图的初始类别。
根据初始分类结果计算与草图分类的初始模型相关的第三损失函数的函数值，及与真实图分类的初始模型相关的第四损失函数的函数值；最后根据第三损失函数的函数值调整草图分类的初始模型，根据第四损失函数的函数值调整真实图分类的初始模型中的固定参数值，以得到草图分类模型和真实图分类模型。
步骤102，选定训练集，在训练集中不仅可以包括多个类别的草图，还可以包括每个类别的真实图。也就是说，训练集中可以包括多个类别的草图和真实图，而每个类别可以对应多个草图和多个真实图。在训练集中的每个图像都具体如下标记：草图的标记或真实图的标记，以及图像所对应类别的标记。
在选定训练集后，分类训练装置可以先对训练集中的各个图像进行预处理，然后再执行步骤103。其中，预处理过程可以包括：将各个图像进行缩放处理，或者裁剪处理，使得处理后的各个图像的尺寸大小相同，这样，可以使得在执行步骤103中确定各个图像的类别时的计算得到简化。
该预处理过程还可以包括：将各个图像进行主体图像的增强，使得各个图像中的主体图像较为清晰，不会模糊，这样，在执行步骤103时，消除了主体图像的不清晰对图像类别确定过程的影响，这里主体图像是指一个图像中的主要图像，而非背景图像，比如一个图像中包括的人物，实物等。还可以有其它预处理，只要是能消除对分类训练装置执行步骤103中图像类别确定过程的影响的预处理方法都属于本申请实施例的范围，在此不一一举例说明。
步骤103，根据上述步骤101确定的草图分类模型确定训练集中草图的类别得到第一类别处理结果。并在草图分类模型确定草图的类别的过程中，根据第二特征分析模型，对草图分类模型中第一特征提取模块提取的草图的特征进行分析得到第二草图特征分析结果。其中对草图的特征的分析可以包括判断该草图的特征是否属于草图，在其它实施例中，还可以包括其它分析处理，在此不进行一一举例。
这里，得到的第一类别处理结果具体可以包括草图分类模型确定的训练 集中各个草图的类别，比如在上述步骤102中选定的训练集中包括n个草图，这些草图可以是“飞机”，“树木”等类别的草图，根据草图分类模型分别确定n个草图的类别，得到n个草图的类别C1，C2，……，Cn。
在第二草图特征分析结果具体可以包括第二特征分析模型判断第一特征提取模块提取的草图的特征是否属于草图的结果。
步骤104，根据上述步骤103得到的第一类别处理结果和第二草图特征分析结果，计算草图分类模型的第一损失函数的函数值。
其中，第一损失函数包括：与第一分类模块相关的损失函数，及第二特征分析模型对草图特征分析的损失函数。
其中，与第一分类模块相关的损失函数可以根据第一类别处理结果得到，具体可以为交叉熵损失函数，用于表示第一分类模块确定的类别与草图的实际类别之间的差别，即误差。
另一种情况下，如果第一分类模块确定草图类别的过程中使用相似特征排序的方法，则该与第一分类模块相关的损失函数也可以为排序损失函数，用于表示在相似特征排序过程中的损失函数。例如，第一分类模块在确定草图1的类别过程中，确定草图1与草图2之间的特征相似度，大于草图1与草图2之间的特征相似度，从而确定草图1与草图2的类别相同，则排序损失函数就可以表示在特征相似度排序过程中确定的排序，与实际特征相似度的排序之间的差别
第二特征分析模型对草图特征分析的损失函数是根据第二草图特征分析结果得到的，用于表示第二特征分析模型对第一特征提取模块提取的草图的特征进行分析得到的分析结果，与草图的实际特征之间的差别。
步骤105，根据第一损失函数的函数值调整草图分类模型中的第一固定参数值。
其中，第一固定参数值为草图分类模型所包括的第一特征提取模块和第一分类模块分别在计算过程中所用到的固定的，且不需要随时赋值的参数，比如权重，角度等参数的参数值。如果计算的第一损失函数的函数值较大，比如大于预置的值，则需要改变第一固定参数值，比如将某个权重的权重值增大，或将某个角度的角度值减小等，使得按照调整后的第一固定参数值计算的第一损失函数的函数值减小。
需要说明的是，上述步骤101和102之间并没有绝对的顺序关系，可以同时进行，也可以顺序进行，图1所示的是其中一种具体实现方式。
进一步地，参考图2所示，分类训练装置还可以调整第二特征分析模型中的固定参数值。具体地，分类训练装置会确定真实图分类模型，具体包括确定草图分类模型的结构和相应的固定参数值，真实图分类模型包括第二特征提取模块和第二分类模块；在真实图分类模型确定训练集中真实图的类别的过程中，根据第二特征分析模型，对第二特征提取模块提取的真实图的特征进行分析得到第一真实图特征分析结果；然后根据第一真实图特征分析结果和第二草图特征分析结果，计算第二特征分析模型的第二对抗损失函数的函数值，并根据第二对抗损失函数的函数值调整第二特征分析模型的固定参数值，使得按照调整后的第二特征分析模型的固定参数值计算的第二对抗损失函数的函数值减小。
其中，第一真实图特征分析结果具体可以包括第二特征分析模型判断第二特征提取模块提取的真实图的特征是否属于真实图的结果。而第二对抗损失函数可以包括：第二特征分析模型对真实图特征分析的损失函数，可以根据第一真实图特征分析结果得到；及第二特征分析模型对草图特征分析的损失函数，可以根据第二草图特征分析结果得到。其中，第二特征分析模型对真实图特征分析的损失函数，用于表示第二特征分析模型对第二特征提取模块提取的真实图的特征进行分析得到的分析结果，与真实图的实际特征之间的差别。
另外，需要说明的是，上述步骤103到105是针对上述步骤101确定的草图分类模型，分别对训练集中的各个草图进行处理后，由分类训练装置调整第一固定参数值的过程。而在实际应用中，需要通过不断地循环执行上述步骤103到105，直到对第一固定参数值的调整满足一定的停止条件为止。
因此，分类训练装置在执行了上述实施例步骤101到105之后，还需要判断当前对第一固定参数值的调整是否满足预置的停止条件，如果满足，则结束流程；如果不满足，则针对调整第一固定参数值后的草图分类模型，返回执行上述步骤103到105的步骤。即执行得到第一类别处理结果和第二草图特征分析结果，计算草图分类模型的第一损失函数的函数值及调整第一固定参数值的步骤。
其中，预置的停止条件包括但不限于如下条件中的任何一个：当前调整的第一固定参数值与上一次调整的第一固定参数值的第一差值小于第一阈值，即调整的第一固定参数值达到收敛；及对第一固定参数值的调整次数达到预置次数等。可以理解，这里停止条件，是指停止对第一固定参数值进行调整的条件。
进一步地，分类训练装置还可以通过如下步骤对真实图分类模型的第二固定参数值进行调整，流程图如图4所示，示意图如图2所示，包括：
步骤201，确定对第一特征提取模块的输出结果进行分析的第一特征分析模型。
第一特征分析模型用于在草图分类模型识别草图类别的过程中，对第一特征提取模块提取的草图的特征(即输出结果)进行分析，具体可以判别该草图的特征属于草图。
步骤202，根据真实图分类模型确定上述训练集中的真实图的类别得到第二类别处理结果；根据第一特征分析模型，对第二特征提取模块提取的真实图的特征进行分析得到第二真实图特征分析结果。
这里，得到的第二类别处理结果具体可以包括由真实图分类模型确定的训练集中各个真实图的类别。在第二真实图特征分析结果具体可以包括第一特征分析模型判断第二特征提取模块提取的真实图的特征是否属于真实图的结果。
步骤203，根据第二类别处理结果及第二真实图特征分析结果，计算真实图分类模型的第二损失函数的函数值，其中，第二损失函数中包括与第二分类模块相关的损失函数，及第一特征分析模型对真实图的特征分析的损失函数。
其中，与第二分类模块相关的损失函数可以根据第二类别处理结果得到。具体可以为交叉熵损失函数，用于表示第二分类模块确定的类别与真实图的实际类别之间的差别；如果第二分类模块确定真实图类别的过程中使用相似特征排序的方法，则该与第二分类模块相关的损失函数也可以为排序损失函数，用于表示在相似特征排序过程中的损失函数。
第一特征分析模型对真实图特征分析的损失函数可以根据第二真实图特征分析结果得到，用于表示第一特征分析模型对第二特征提取模块提取的真 实图的特征进行分析得到的分析结果，与真实图的实际特征之间的差别。
步骤204，根据第二损失函数的函数值调整真实图分类模型中的第二固定参数值。
其中，第二固定参数值为真实图分类模型所包括的第二特征提取模块和第二分类模块分别在计算过程中所用到的固定的，且不需要随时赋值的参数，比如权重，角度等参数的参数值。如果计算的第二损失函数的函数值较大，比如大于预置的值，则需要改变第二固定参数值，比如将某个权重的权重值增加，或将某个角度的角度值减小等，使得按照调整后的第二固定参数值计算的第二损失函数的函数值减小。
进一步地，参考图2所示，分类训练装置还可以调整第一特征分析模型中的固定参数值。具体地，分类训练装置会根据第一特征分析模型，对第一特征提取模块提取的草图的特征进行分析得到第一草图特征分析结果；然后根据第一草图特征分析结果和第二真实图特征分析结果，计算第一特征分析模型的第一对抗损失函数的函数值，并根据第一对抗损失函数的函数值调整第一特征分析模型的固定参数值。
其中，第一草图特征分析结果具体可以包括第一特征分析模型判断第一特征提取模块提取的草图的特征是否属于草图的结果。而第一对抗损失函数可以包括：第一特征分析模型对真实图特征分析的损失函数，可以根据第二真实图特征分析结果得到；及第一特征分析模型对草图特征分析的损失函数，可以根据第一草图特征分析结果得到。其中，第一特征分析模型对草图特征分析的损失函数，用于表示第一特征分析模型对第一特征提取模块提取的草图的特征进行分析得到的分析结果，与草图的实际特征之间的差别。
另外，需要说明的是，上述步骤202到204是针对真实图分类模型，对训练集中真实图进行处理后，由分类训练装置调整第二固定参数值的过程。而在实际应用中，需要通过不断地循环执行上述步骤202到204，直到对第二固定参数值的调整满足一定的停止条件为止。
因此，分类训练装置在执行了上述步骤201到204之后，需要判断当前对第二固定参数值的调整是否满足预置的停止条件，如果满足，则结束流程；如果不满足，则针对调整第二固定参数值后的真实图分类模型，返回上述步骤202到204，即执行得到第二类别处理结果和第二真实图特征分析结果，计 算真实图分类模型的第二损失函数的函数值及调整第二固定参数值的步骤。
其中，预置的停止条件包括但不限于如下条件中的任何一个：当前调整的第二固定参数值与上一次调整的第二固定参数值的第一差值小于第二阈值，即调整的第二固定参数值达到收敛；及对第二固定参数值的调整次数达到预置次数等。
且需要说明的是，上述步骤202到204，与步骤103到105可以交替进行，比如，在某一次调整过程中，可以调整草图分类模型的第一固定参数值及调整第二特征分析模型的固定参数值，即执行步骤103到105；在另一调整过程中，调整真实图分类模型的第二固定参数值及调整第一特征分析模型的固定参数值，即执行步骤202到204；而再次调整过程中，再调整草图分类模型的第一固定参数值及调整第二特征分析模型的固定参数值，即执行步骤103到105，这样以此类推。
进一步地，分类训练装置在通过上述实施例的方法，得到调整后的草图分类模型和真实图分类模型，在调整后的草图分类模型和真实图分类模型的实际应用中：一种情况下，分类训练装置可以先获取待分类草图(比如用户通过终端设备输入的待分类草图)，然后根据调整后的草图分类模型对待分类草图进行分类，得到待分类草图的类别，从而实现了对草图的分类。
另一种情况下，分类训练装置可以先获取待分类草图(比如用户通过终端设备输入的待分类草图)，及获取分类训练装置中储存的各个真实图；然后根据调整后的草图分类模型对待分类草图进行分类得到待分类草图的类别，根据调整后的真实图分类模型分别对储存的各个真实图进行分类得到各个真实图的类别；最后再选出类别与待分类草图的类别相同的真实图，以提供给用户的终端设备。从而实现了对草图的检索。
可见，在本实施例的方法中，分类训练装置会先选定训练集，并根据草图分类模型确定训练集中的草图的类别得到第一类别处理结果，且可以根据第二特征分析模型对第一特征提取模型提取的草图特征进行分析，第二草图分析结果；然后根据第一类别处理结果和第二草图分析结果得到第一损失函数的函数值；最后根据第一损失函数的函数值对草图分类模型的第一固定参数值进行调整。且分类训练装置还会根据真实图分类模型确定训练集中的真实图的类别，得到第二类别处理结果，且可以根据第一特征分析模型对第二 特征提取模块提取的真实图特征进行分析，得到第二真实图分析结果；然后根据第二类别处理结果和第二真实图分析结果得到第二损失函数的函数值；最后根据第二损失函数的函数值对真实图分类模型的第二固定参数值进行调整。这样，在对某一分类模型(比如草图分类模型)的固定参数值进行调整时，不仅会参考该分类模型对相应图像进行分类的误差，还会借鉴另一分类模型(比如真实图分类模型)在分类过程中的有用信息，即对另一分类模型中特征提取模块提取的特征进行分析的特征分析模型(比如第二特征分析模型)，从而使得调整后的草图分类模型和真实图分类模型的分类计算更准确。
以下一个具体的应用实例来说明本实施例的方法，参考图5所示，在本实施例中，草图分类模型和真实图分类模型可以使用相同结构的基于卷积神经网络，分别记为CNN_1和CNN_2；草图分类模型中包括的第一特征提取模块和第一分类模块分别记为CNN11和CNN12，第一特征分析模型具体为草图鉴别器D_1；真实图分类模型中包括的第二特征提取模块和第二分类模块分别记为CNN21和CNN22，第二特征分析模型具体为真实图鉴别器D_2。则本实施例的分类训练方法可以通过如下步骤实现，流程图如图6所示，包括：
步骤301，确定草图分类模型CNN_1和真实图分类模型CNN_2，及草图鉴别器D_1和真实图鉴别器D_2，具体包括确定各个模型的结构及固定参数的初始值。
步骤302，选定训练集，在训练集中包括多个类别的图像对，每个类别的图像对中包括同一类别的草图和真实图，具体为
Figure PCTCN2018117158-appb-000001
其中，Si为某一类别i的草图图像，Ii为该类别i的真实图图像。
步骤303，根据草图分类模型CNN_1确定训练集中草图的类别得到草图的类别，根据真实图分类模型CNN_2确定训练集中真实图的类别得到真实图的类别。
在这个过程中，通过草图鉴别器D_1对草图分类模型CNN_1中的草图特征提取模块CNN11提取的草图特征进行鉴别，得到草图特征鉴别结果11；通过草图鉴别器D_1对真实图分类模型CNN_2中的真实图特征提取模块CNN21提取的真实图特征进行鉴别，得到真实图特征鉴别结果12。
通过真实图鉴别器D_2对真实图分类模型CNN_2中的真实图特征提取模块CNN21提取的真实图特征进行鉴别，得到真实图特征鉴别结果21；通过真实图鉴别器D_2对草图分类模型CNN_1中的草图特征提取模块CNN11提取的草图特征进行鉴别，得到草图特征鉴别结果22。
步骤304，先固定草图分类模型CNN_1和真实图鉴别器D_2，对真实图分类模型CNN_2的固定参数值和草图鉴别器D_1的固定参数值进行调整，使得对真实图分类模型CNN_2的调整借鉴了草图分类模型CNN_1在分类过程中的有用信息。
具体地，分类训练装置会根据如下公式1，及上述步骤303中得到的草图特征鉴别结果11和真实图特征鉴别结果12，计算草图鉴别器D_1的对抗函数
Figure PCTCN2018117158-appb-000002
的函数值，根据该函数值调整草图鉴别器D_1的固定参数值。
Figure PCTCN2018117158-appb-000003
根据如下公式2，上述步骤303中真实图分类模型CNN_2确定的真实图的类别及真实图特征鉴别结果12，计算真实图分类模型CNN_2的损失函数的函数值，根据该函数值调整真实图分类模型CNN_2的固定参数值。具体可以包括：与真实图分类模型CNN_2中真实图分类模块CNN22相关的损失函数，比如交叉熵损失函数
Figure PCTCN2018117158-appb-000004
及草图鉴别器D_1对真实图特征判别的损失函数，可以根据真实图特征鉴别结果12得到。其中，相关的损失函数
Figure PCTCN2018117158-appb-000005
可以通过如下公式3及上述步骤303中真实图分类模型CNN_2确定的真实图的类别得到。
Figure PCTCN2018117158-appb-000006
Figure PCTCN2018117158-appb-000007
其中，log(D_1(CNN_2(I i)))可以表示草图鉴别器D_1对真实图特征判别的损失函数。
步骤305，再固定真实图分类模型CNN_2和草图鉴别器D_1，对草图分类模型CNN_1的固定参数值和真实图鉴别器D_2的固定参数值进行调整，使得对草图分类模型CNN_1的调整借鉴了真实图分类模型CNN_2在分类过程中的有用信息。
具体地，分类训练装置会根据如下公式4，及上述步骤303中得到的真实图特征鉴别结果21和草图特征鉴别结果22，计算真实图鉴别器D_2的对抗函数
Figure PCTCN2018117158-appb-000008
的函数值，根据该函数值调整真实图鉴别器D_2的固定参数值。
Figure PCTCN2018117158-appb-000009
根据如下公式5，上述步骤303中草图分类模型CNN_1确定的草图的类别和草图特征鉴别结果22，计算草图分类模型CNN_1的损失函数的函数值，根据该函数值调整草图分类模型CNN_1的固定参数值。具体可以包括：与草图分类模型CNN_1中草图分类模块CNN12相关的损失函数，比如交叉熵损失函数
Figure PCTCN2018117158-appb-000010
及真实图鉴别器D_2对草图特征判别的损失函数，可以根据草图特征鉴别结果22得到。其中，相关的损失函数
Figure PCTCN2018117158-appb-000011
可以通过如下公式6及草图的类别得到。
Figure PCTCN2018117158-appb-000012
Figure PCTCN2018117158-appb-000013
其中，log(D_2(CNN_1(S i)))可以表示真实图鉴别器D_2对草图特征判别的损失函数。
步骤306，执行完上述步骤301到305之后，判断对草图分类模型CNN_1和真实图分类模型CNN_2的固定参数值的调整是否满足预置条件，如果满足，则结束流程；如果不满足，则针对调整后的草图分类模型CNN_1和真实图分类模型CNN_2，及调整后的草图鉴别器D_1和真实图鉴别器D_2，返回执行上述步骤303到305。
本申请实施例还提供一种分类训练装置，其结构示意图如图7所示，具体可以包括：
模型确定单元10，用于确定草图分类模型，所述草图分类模型中包括第一特征提取模块和第一分类模块；及确定对第二特征提取模块的输出结果进行分析的第二特征分析模型，所述第二特征提取模块属于真实图分类模型。
具体地，模型确定单元10，具体用于确定相同结构的草图分类的初始模型和真实图分类的初始模型，及确定已标记类别的草图和已标记类别的真实图；根据所述草图分类的初始模型确定所述已标记类别的草图的类别，根据真实图分类的初始模型确定已标记类别的真实图的类别，得到初始分类结果；根据所述初始分类结果计算与所述草图分类的初始模型相关的第三损失函数的函数值，及与所述真实图分类的初始模型相关的第四损失函数的函数值，并根据所述第三损失函数的函数值调整所述草图分类的初始模型中的固定参数值，根据第四损失函数的函数值调整真实图分类的初始模型中的固定参数值，以得到所述草图分类模型和真实图分类模型。
训练集单元11，用于选定训练集，所述训练集包括多个类别的草图。该训练集中还包括相应类别的真实图。
处理单元12，用于根据所述模型确定单元10确定的草图分类模型确定所述训练集单元11选定训练集中草图的类别得到第一类别处理结果；根据所述第二特征分析模型，对第一特征提取模块提取的所述草图的特征进行分析得到第二草图特征分析结果；
函数值计算单元13，用于根据所述处理单元12得到的第一类别处理结果及第二草图特征分析结果，计算所述草图分类模型的第一损失函数的函数值，其中，所述第一损失函数中包括与所述第一分类模块相关的损失函数，及所述第二特征分析模型对所述草图的特征分析的损失函数。
调整单元14，用于根据所述函数值计算单元13计算的第一损失函数的函 数值调整所述草图分类模型中的第一固定参数值。
进一步地，上述训练集单元11选定的训练集中还可以包括相应类别的真实图，所述模型确定单元10，还可以确定真实图分类模型，所述真实图分类模型包括第二特征提取模块和第二分类模块；处理单元12，还用于在所述真实图分类模型确定所述真实图的类别时，根据所述第二特征分析模型，对所述第二特征提取模块提取的所述真实图的特征进行分析得到第一真实图特征分析结果；函数值计算单元13，还用于根据所述第一真实图特征分析结果和第二草图特征分析结果，计算所述第二特征分析模型的第二对抗损失函数的函数值；调整单元14，还用于根据所述第二对抗损失函数的函数值调整所述第二特征分析模型的固定参数值。
进一步地，模型确定单元10，还用于确定对所述第一特征提取模块的输出结果进行分析的第一特征分析模型；处理单元12，还用于根据所述真实图分类模型确定所述训练集中的真实图的类别得到第二类别处理结果；根据所述第一特征分析模型，对第二特征提取模块提取的所述真实图的特征进行分析得到第二真实图特征分析结果；函数值计算单元13，还用于根据所述第二类别处理结果及第二真实图特征分析结果，计算所述真实图分类模型的第二损失函数的函数值，其中，所述第二损失函数中包括与所述第二分类模块相关的损失函数，及所述第一特征分析模型对所述真实图的特征分析的损失函数；调整单元14，还用于根据所述第二损失函数的函数值调整所述真实图分类模型中的第二固定参数值。
进一步地，处理单元12，还用于根据所述第一特征分析模型，对所述第一特征提取模块提取的所述草图的特征进行分析得到第一草图特征分析结果；函数值计算单元13，还用于根据所述第一草图特征分析结果和第二真实图特征分析结果，计算所述第一特征分析模型的第一对抗损失函数的函数值；调整单元14，还用于根据所述第一对抗损失函数的函数值调整所述第一特征分析模型的固定参数值。
可见，在本实施例的装置中，训练集单元11会选定训练集，处理单元12根据草图分类模型确定训练集中草图的类别，得到第一类别处理结果，且可以根据第二特征分析模型对第一特征提取模型提取的草图特征进行分析，得到第二草图分析结果；然后函数值计算单元13根据第一类别处理结果和第二 草图分析结果得到第一损失函数的函数值；最后调整单元14根据第一损失函数的函数值对草图分类模型的第一固定参数值进行调整。且处理单元12还可以根真实图分类模型确定训练集中真实图的类别，得到第二类别处理结果，且可以根据第一特征分析模型对第二特征提取模块提取的真实图特征进行分析，得到第二真实图分析结果；然后函数值计算单元13根据第二类别处理结果和第二真实图分析结果得到第二损失函数的函数值；最后调整单元14根据第二损失函数的函数值对真实图分类模型的第二固定参数值进行调整。这样，在对某一分类模型(比如草图分类模型)的固定参数值进行调整时，不仅会参考该分类模型对相应图像进行分类的误差，还会借鉴另一分类模型(比如真实图分类模型)在分类过程中的有用信息，即对另一分类模型中特征提取模块提取的特征进行分析的特征分析模型(比如第二特征分析模型)，从而使得调整后的草图分类模型和真实图分类模型的分类计算更准确。
参考图8所示，在一个具体的实施例中，分类训练装置除了可以包括如图7所示的结构外，还可以包括判断单元15和分类单元16，其中：
判断单元15，用于判断调整单元14对所述第一固定参数值的调整是否满足预置的停止条件，如果不满足，则通知所述处理单元12针对调整所述第一固定参数值后的草图分类模型，得到所述第一类别处理结果和第二草图特征分析结果。
其中，预置的停止条件可以包括但不限于如下条件中的任何一个：
所述当前调整的第一固定参数值与上一次调整的第一固定参数值的第一差值小于第一阈值；及对第一固定参数的调整次数达到预置次数等。
进一步地，判断单元15，用于判断调整单元14对所述第二固定参数值的调整是否满足预置的停止条件，如果不满足，则通知所述处理单元12针对调整所述第二固定参数值后的真实图分类模型，得到所述第二类别处理结果和第二真实图特征分析结果。这里预置的停止条件可以包括但不限于如下条件中的任何一个：所述当前调整的第二固定参数值与上一次调整的第二固定参数值的第一差值小于第一阈值；及对第二固定参数的调整次数达到预置次数等。
分类单元16，用于获取待分类草图，根据所述调整单元14调整后的草图分类模型对所述待分类草图进行分类，得到所述待分类草图的类别。
该分类单元16，还可以获取待分类草图，及获取分类训练装置中储存的各个真实图；然后根据调整单元14调整后的草图分类模型对待分类草图进行分类得到待分类草图的类别，根据调整后的真实图分类模型分别对储存的各个真实图进行分类得到各个真实图的类别；最后再选出类别与待分类草图的类别相同的真实图。
本申请实施例还提供一种服务器，其结构示意图如图9所示，该服务器可因配置或性能不同而产生比较大的差异，可以包括一个或一个以上中央处理器(central processing units，CPU)20(例如，一个或一个以上处理器)和存储器21，一个或一个以上存储应用程序221或数据222的存储介质22(例如一个或一个以上海量存储设备)。其中，存储器21和存储介质22可以是短暂存储或持久存储。存储在存储介质22的程序可以包括一个或一个以上模块(图示没标出)，每个模块可以包括对服务器中的一系列计算机可读指令。更进一步地，中央处理器20可以设置为与存储介质22通信，在服务器上执行存储介质22中的一系列计算机可读指令。该计算机可读指令被执行时，可使得中央处理器222所包括的一个或多个处理器执行一种分类训练方法。存储介质730可以是非易失性存储介质。非易失性存储介质可以是非易失性可读存储介质。
具体地，在存储介质22中储存的应用程序221包括分类训练应用程序，且该程序可以包括上述分类训练装置中的模型确定单元10，训练集单元11，处理单元12，函数值计算单元13，调整单元14，判断单元15和分类单元16，在此不进行赘述。更进一步地，中央处理器20可以设置为与存储介质22通信，在服务器上执行存储介质22中储存的分类训练应用程序对应的一系列操作。
服务器还可以包括一个或一个以上电源23，一个或一个以上有线或无线网络接口24，一个或一个以上输入输出接口25，和/或，一个或一个以上操作系统223，例如Windows ServerTM，Mac OS XTM，UnixTM,LinuxTM，FreeBSDTM等等。
上述方法实施例中所述的由分类训练装置所执行的步骤可以基于该图9所示的服务器的结构。分类训练装置中包括的每个单元模块可全部或部分通过软件、硬件或其组合来实现。
本申请实施例还提供一种存储介质，所述存储介质储存有计算机可读指令，所述计算机可读指令适于由处理器加载并执行如本申请各实施例所述的 分类训练方法。
本申请实施例还提供一种服务器，包括一个或多个处理器和存储器，所述存储器中存储有计算机可读指令，所述一个或多个处理器，用于实现各个计算机可读指令所述计算机可读指令用于由一个或多个处理器加载并执行如本申请各实施例所述的分类训练方法。
本领域普通技术人员可以理解上述实施例的各种方法中的全部或部分步骤是可以通过计算机可读指令来指令相关的硬件来完成，该计算机可读指令可以存储于一计算机可读存储介质中，存储介质可以包括：只读存储器(ROM)、随机存取存储器RAM)、磁盘或光盘等。
以上实施例的各技术特征可以进行任意的组合，为使描述简洁，未对上述实施例中的各个技术特征所有可能的组合都进行描述，然而，只要这些技术特征的组合不存在矛盾，都应当认为是本说明书记载的范围。
以上对本申请实施例所提供的分类训练方法、服务器及存储介质进行了详细介绍，本文中应用了具体个例对本申请的原理及实施方式进行了阐述，以上实施例的说明只是用于帮助理解本申请的方法及其核心思想，但并不能因此而理解为对发明专利范围的限制；同时，对于本领域的一般技术人员，依据本申请的思想，在具体实施方式及应用范围上均会有改变之处，这些都属于本申请的保护范围。综上所述，本说明书内容不应理解为对本申请的限制。

Claims (20)

  1. 一种分类训练方法，其特征在于，包括：
    服务器确定草图分类模型，所述草图分类模型中包括第一特征提取模块和第一分类模块；及确定对第二特征提取模块的输出结果进行分析的第二特征分析模型，所述第二特征提取模块属于真实图分类模型；
    所述服务器选定训练集，所述训练集包括多个类别的草图；
    所述服务器根据所述草图分类模型确定所述训练集中草图的类别得到第一类别处理结果；根据所述第二特征分析模型，对第一特征提取模块提取的所述草图的特征进行分析得到第二草图特征分析结果；
    所述服务器根据所述第一类别处理结果及第二草图特征分析结果，计算所述草图分类模型的第一损失函数的函数值；及
    所述服务器根据所述第一损失函数的函数值调整所述草图分类模型中的第一固定参数值。
  2. 如权利要求1所述的方法，其特征在于，所述训练集中还包括相应类别的真实图，所述方法还包括：
    所述服务器确定真实图分类模型，所述真实图分类模型包括第二特征提取模块和第二分类模块；
    所述服务器在所述真实图分类模型确定所述真实图的类别时，根据所述第二特征分析模型，对所述第二特征提取模块提取的所述真实图的特征进行分析得到第一真实图特征分析结果；
    所述服务器根据所述第一真实图特征分析结果和第二草图特征分析结果，计算所述第二特征分析模型的第二对抗损失函数的函数值；及
    所述服务器根据所述第二对抗损失函数的函数值调整所述第二特征分析模型的固定参数值。
  3. 如权利要求2所述的方法，其特征在于，所述服务器确定草图分类模型和真实图分类模型，包括：
    服务器确定草图分类的初始模型和真实图分类的初始模型，及确定已标记类别的草图和已标记类别的真实图；
    所述服务器根据所述草图分类的初始模型确定所述已标记类别的草图的类别，根据所述真实图分类的初始模型确定所述已标记类别的真实图的类别， 得到初始分类结果；及
    所述服务器根据所述初始分类结果计算与所述草图分类的初始模型相关的第三损失函数的函数值，及与所述真实图分类的初始模型相关的第四损失函数的函数值，并根据所述第三损失函数的函数值调整所述草图分类的初始模型，根据所述第四损失函数的函数值调整所述真实图分类的初始模型中的固定参数值，以得到所述草图分类模型和真实图分类模型。
  4. 如权利要求2所述的方法，其特征在于，还包括：
    所述服务器确定对所述第一特征提取模块的输出结果进行分析的第一特征分析模型；
    所述服务器根据所述真实图分类模型确定所述训练集中的真实图的类别得到第二类别处理结果；根据所述第一特征分析模型，对第二特征提取模块提取的所述真实图的特征进行分析得到第二真实图特征分析结果；
    所述服务器根据所述第二类别处理结果及第二真实图特征分析结果，计算所述真实图分类模型的第二损失函数的函数值；及
    所述服务器根据所述第二损失函数的函数值调整所述真实图分类模型中的第二固定参数值。
  5. 如权利要求4所述的方法，其特征在于，还包括：
    所述服务器根据所述第一特征分析模型，对所述第一特征提取模块提取的所述草图的特征进行分析得到第一草图特征分析结果；
    所述服务器根据所述第一草图特征分析结果和第二真实图特征分析结果，计算所述第一特征分析模型的第一对抗损失函数的函数值；及
    所述服务器根据所述第一对抗损失函数的函数值调整所述第一特征分析模型的固定参数值。
  6. 如权利要求2至5任一项所述的方法，其特征在于，还包括：
    所述服务器如果对所述第一固定参数值的调整不满足预置的停止条件时，针对调整所述第一固定参数值后的草图分类模型，返回所述根据所述草图分类模型确定所述训练集中草图的类别得到第一类别处理结果以继续执行，直至对所述第一固定参数值的调整满足预置的停止条件。
  7. 如权利要求6所述的方法，其特征在于，所述预置的停止条件包括如下条件中的任何一个：
    所述当前调整的第一固定参数值与上一次调整的第一固定参数值的第一差值小于第一阈值；对所述第一固定参数值的调整次数达到预置次数。
  8. 如权利要求1至3任一项所述的方法，其特征在于，还包括：
    所述服务器获取待分类草图，根据所述调整后的草图分类模型对所述待分类草图进行分类，得到所述待分类草图的类别。
  9. 一个或多个存储介质，其特征在于，所述一个或多个存储介质储存有计算机可读指令，所述计算机可读指令由处理器加载并执行如下步骤：
    确定真实图分类模型，所述真实图分类模型包括第二特征提取模块和第二分类模块；
    在所述真实图分类模型确定所述真实图的类别时，根据所述第二特征分析模型，对所述第二特征提取模块提取的所述真实图的特征进行分析得到第一真实图特征分析结果；
    根据所述第一真实图特征分析结果和第二草图特征分析结果，计算所述第二特征分析模型的第二对抗损失函数的函数值；及
    根据所述第二对抗损失函数的函数值调整所述第二特征分析模型的固定参数值。
  10. 如权利要求9所述的存储介质，其特征在于，所述确定草图分类模型和真实图分类模型，包括：
    确定草图分类的初始模型和真实图分类的初始模型，及确定已标记类别的草图和已标记类别的真实图；
    根据所述草图分类的初始模型确定所述已标记类别的草图的类别，根据所述真实图分类的初始模型确定所述已标记类别的真实图的类别，得到初始分类结果；及
    根据所述初始分类结果计算与所述草图分类的初始模型相关的第三损失函数的函数值，及与所述真实图分类的初始模型相关的第四损失函数的函数值，并根据所述第三损失函数的函数值调整所述草图分类的初始模型，根据所述第四损失函数的函数值调整所述真实图分类的初始模型中的固定参数值，以得到所述草图分类模型和真实图分类模型。
  11. 如权利要求10所述的存储介质，其特征在于，所述计算机可读指令由处理器加载并执行如下步骤：
    确定对所述第一特征提取模块的输出结果进行分析的第一特征分析模型；
    根据所述真实图分类模型确定所述训练集中的真实图的类别得到第二类别处理结果；根据所述第一特征分析模型，对第二特征提取模块提取的所述真实图的特征进行分析得到第二真实图特征分析结果；
    根据所述第二类别处理结果及第二真实图特征分析结果，计算所述真实图分类模型的第二损失函数的函数值；及
    根据所述第二损失函数的函数值调整所述真实图分类模型中的第二固定参数值。
  12. 如权利要求10所述的存储介质，其特征在于，所述计算机可读指令由处理器加载并执行如下步骤：
    根据所述第一特征分析模型，对所述第一特征提取模块提取的所述草图的特征进行分析得到第一草图特征分析结果；
    根据所述第一草图特征分析结果和第二真实图特征分析结果，计算所述第一特征分析模型的第一对抗损失函数的函数值；及
    根据所述第一对抗损失函数的函数值调整所述第一特征分析模型的固定参数值。
  13. 一种服务器，其特征在于，包括一个或多个处理器和存储器，所述存储器中存储有计算机可读指令，所述处理器，用于实现各个计算机可读指令；所述计算机可读指令由一个或多个处理器加载并执行以下步骤：
    确定草图分类模型，所述草图分类模型中包括第一特征提取模块和第一分类模块；及确定对第二特征提取模块的输出结果进行分析的第二特征分析模型，所述第二特征提取模块属于真实图分类模型；
    选定训练集，所述训练集包括多个类别的草图；
    根据所述草图分类模型确定所述训练集中草图的类别得到第一类别处理结果；根据所述第二特征分析模型，对第一特征提取模块提取的所述草图的特征进行分析得到第二草图特征分析结果；
    根据所述第一类别处理结果及第二草图特征分析结果，计算所述草图分类模型的第一损失函数的函数值；及
    根据所述第一损失函数的函数值调整所述草图分类模型中的第一固定参 数值。
  14. 如权利要求13所述的服务器，其特征在于，所述训练集中还包括相应类别的真实图，所述计算机可读指令由一个或多个处理器加载并执行以下步骤：
    确定真实图分类模型，所述真实图分类模型包括第二特征提取模块和第二分类模块；
    在所述真实图分类模型确定所述真实图的类别时，根据所述第二特征分析模型，对所述第二特征提取模块提取的所述真实图的特征进行分析得到第一真实图特征分析结果；
    根据所述第一真实图特征分析结果和第二草图特征分析结果，计算所述第二特征分析模型的第二对抗损失函数的函数值；及
    根据所述第二对抗损失函数的函数值调整所述第二特征分析模型的固定参数值。
  15. 如权利要求14所述的服务器，其特征在于，所述确定草图分类模型和真实图分类模型，包括：
    确定草图分类的初始模型和真实图分类的初始模型，及确定已标记类别的草图和已标记类别的真实图；
    根据所述草图分类的初始模型确定所述已标记类别的草图的类别，根据所述真实图分类的初始模型确定所述已标记类别的真实图的类别，得到初始分类结果；及
    根据所述初始分类结果计算与所述草图分类的初始模型相关的第三损失函数的函数值，及与所述真实图分类的初始模型相关的第四损失函数的函数值，并根据所述第三损失函数的函数值调整所述草图分类的初始模型，根据所述第四损失函数的函数值调整所述真实图分类的初始模型中的固定参数值，以得到所述草图分类模型和真实图分类模型。
  16. 如权利要求14所述的服务器，其特征在于，所述计算机可读指令由一个或多个处理器加载并执行以下步骤：
    确定对所述第一特征提取模块的输出结果进行分析的第一特征分析模型；
    根据所述真实图分类模型确定所述训练集中的真实图的类别得到第二类 别处理结果；根据所述第一特征分析模型，对第二特征提取模块提取的所述真实图的特征进行分析得到第二真实图特征分析结果；
    根据所述第二类别处理结果及第二真实图特征分析结果，计算所述真实图分类模型的第二损失函数的函数值；及
    根据所述第二损失函数的函数值调整所述真实图分类模型中的第二固定参数值。
  17. 如权利要求16所述的服务器，其特征在于，所述计算机可读指令由一个或多个处理器加载并执行以下步骤：
    根据所述第一特征分析模型，对所述第一特征提取模块提取的所述草图的特征进行分析得到第一草图特征分析结果；
    根据所述第一草图特征分析结果和第二真实图特征分析结果，计算所述第一特征分析模型的第一对抗损失函数的函数值；及
    根据所述第一对抗损失函数的函数值调整所述第一特征分析模型的固定参数值。
  18. 如权利要求13至17任一项所述的服务器，其特征在于，所述计算机可读指令由一个或多个处理器加载并执行以下步骤：
    如果对所述第一固定参数值的调整不满足预置的停止条件时，针对调整所述第一固定参数值后的草图分类模型，返回所述根据所述草图分类模型确定所述训练集中草图的类别得到第一类别处理结果以继续执行，直至对所述第一固定参数值的调整满足预置的停止条件。
  19. 如权利要求18所述的服务器，其特征在于，所述预置的停止条件包括如下条件中的任何一个：
    所述当前调整的第一固定参数值与上一次调整的第一固定参数值的第一差值小于第一阈值；对所述第一固定参数值的调整次数达到预置次数。
  20. 如权利要求13至15任一项所述的服务器，其特征在于，所述计算机可读指令由一个或多个处理器加载并执行以下步骤：
    获取待分类草图，根据所述调整后的草图分类模型对所述待分类草图进行分类，得到所述待分类草图的类别。
PCT/CN2018/117158 2017-12-12 2018-11-23 分类训练方法、服务器及存储介质 WO2019114523A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
EP18887500.9A EP3726426A4 (en) 2017-12-12 2018-11-23 CLASSIFICATION TRAINING PROCEDURES, SERVER AND STORAGE MEDIUM
US16/696,361 US11017220B2 (en) 2017-12-12 2019-11-26 Classification model training method, server, and storage medium

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201711322612.2 2017-12-12
CN201711322612.2A CN108090508B (zh) 2017-12-12 2017-12-12 一种分类训练方法、装置及存储介质

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US16/696,361 Continuation US11017220B2 (en) 2017-12-12 2019-11-26 Classification model training method, server, and storage medium

Publications (1)

Publication Number Publication Date
WO2019114523A1 true WO2019114523A1 (zh) 2019-06-20

Family

ID=62175384

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/117158 WO2019114523A1 (zh) 2017-12-12 2018-11-23 分类训练方法、服务器及存储介质

Country Status (4)

Country Link
US (1) US11017220B2 (zh)
EP (1) EP3726426A4 (zh)
CN (2) CN110633745B (zh)
WO (1) WO2019114523A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111461352A (zh) * 2020-04-17 2020-07-28 支付宝(杭州)信息技术有限公司 模型训练、业务节点识别方法、装置及电子设备
CN113313022A (zh) * 2021-05-27 2021-08-27 北京百度网讯科技有限公司 文字识别模型的训练方法和识别图像中文字的方法

Families Citing this family (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110633745B (zh) 2017-12-12 2022-11-29 腾讯科技（深圳）有限公司 一种基于人工智能的图像分类训练方法、装置及存储介质
WO2020059446A1 (ja) * 2018-09-20 2020-03-26 富士フイルム株式会社 学習装置及び学習方法
CN110163222B (zh) * 2018-10-08 2023-01-24 腾讯科技（深圳）有限公司 一种图像识别的方法、模型训练的方法以及服务器
CN109977262B (zh) * 2019-03-25 2021-11-16 北京旷视科技有限公司 从视频中获取候选片段的方法、装置及处理设备
US11048932B2 (en) * 2019-08-26 2021-06-29 Adobe Inc. Transformation of hand-drawn sketches to digital images
US11886514B2 (en) 2019-10-11 2024-01-30 Kinaxis Inc. Machine learning segmentation methods and systems
US11526899B2 (en) 2019-10-11 2022-12-13 Kinaxis Inc. Systems and methods for dynamic demand sensing
CN111159397B (zh) * 2019-12-04 2023-04-18 支付宝(杭州)信息技术有限公司 文本分类方法和装置、服务器
CN111143552B (zh) * 2019-12-05 2023-06-27 支付宝(杭州)信息技术有限公司 文本信息的类别预测方法和装置、服务器
CN111079813B (zh) * 2019-12-10 2023-07-07 北京百度网讯科技有限公司 基于模型并行的分类模型计算方法和装置
CN111951128B (zh) * 2020-08-31 2022-01-28 江苏工程职业技术学院 一种节能环保的建筑施工方法和装置
CN111931865B (zh) * 2020-09-17 2021-01-26 平安科技（深圳）有限公司 图像分类模型的训练方法、装置、计算机设备及存储介质
US20220091873A1 (en) * 2020-09-24 2022-03-24 Google Llc Systems and methods for cross media reporting by fast merging of data sources
CN112288098A (zh) * 2020-11-02 2021-01-29 平安数字信息科技(深圳)有限公司 预训练模型的获取方法、装置以及计算机设备
CN112529978B (zh) * 2020-12-07 2022-10-14 四川大学 一种人机交互式抽象画生成方法
CN112862110B (zh) * 2021-02-11 2024-01-30 脸萌有限公司 模型生成方法、装置和电子设备
TWI815492B (zh) * 2022-06-06 2023-09-11 中國鋼鐵股份有限公司 鋼帶表面缺陷辨識方法與系統

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106250866A (zh) * 2016-08-12 2016-12-21 广州视源电子科技股份有限公司 基于神经网络的图像特征提取建模、图像识别方法及装置
CN106683048A (zh) * 2016-11-30 2017-05-17 浙江宇视科技有限公司 一种图像超分辨率方法及设备
CN107220277A (zh) * 2017-04-14 2017-09-29 西北大学 基于手绘草图的图像检索算法
CN108090508A (zh) * 2017-12-12 2018-05-29 腾讯科技（深圳）有限公司 一种分类训练方法、装置及存储介质

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8140448B2 (en) * 2008-05-09 2012-03-20 International Business Machines Corporation System and method for classifying data streams with very large cardinality
US8289287B2 (en) * 2008-12-30 2012-10-16 Nokia Corporation Method, apparatus and computer program product for providing a personalizable user interface
US8718375B2 (en) * 2010-12-03 2014-05-06 Massachusetts Institute Of Technology Sketch recognition system
CN102184395B (zh) * 2011-06-08 2012-12-19 天津大学 基于字符串核的草图识别方法
US20140270489A1 (en) * 2013-03-12 2014-09-18 Microsoft Corporation Learned mid-level representation for contour and object detection
CN104680120B (zh) * 2013-12-02 2018-10-19 华为技术有限公司 一种人脸检测的强分类器的生成方法及装置
WO2017168125A1 (en) * 2016-03-31 2017-10-05 Queen Mary University Of London Sketch based search methods
CN106126581B (zh) * 2016-06-20 2019-07-05 复旦大学 基于深度学习的手绘草图图像检索方法
CN107122396B (zh) * 2017-03-13 2019-10-29 西北大学 基于深度卷积神经网络的三维模型检索方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106250866A (zh) * 2016-08-12 2016-12-21 广州视源电子科技股份有限公司 基于神经网络的图像特征提取建模、图像识别方法及装置
CN106683048A (zh) * 2016-11-30 2017-05-17 浙江宇视科技有限公司 一种图像超分辨率方法及设备
CN107220277A (zh) * 2017-04-14 2017-09-29 西北大学 基于手绘草图的图像检索算法
CN108090508A (zh) * 2017-12-12 2018-05-29 腾讯科技（深圳）有限公司 一种分类训练方法、装置及存储介质

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP3726426A4 *

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111461352A (zh) * 2020-04-17 2020-07-28 支付宝(杭州)信息技术有限公司 模型训练、业务节点识别方法、装置及电子设备
CN111461352B (zh) * 2020-04-17 2023-05-09 蚂蚁胜信(上海)信息技术有限公司 模型训练、业务节点识别方法、装置及电子设备
CN113313022A (zh) * 2021-05-27 2021-08-27 北京百度网讯科技有限公司 文字识别模型的训练方法和识别图像中文字的方法
CN113313022B (zh) * 2021-05-27 2023-11-10 北京百度网讯科技有限公司 文字识别模型的训练方法和识别图像中文字的方法

Also Published As

Publication number Publication date
CN108090508A (zh) 2018-05-29
CN110633745A (zh) 2019-12-31
US20200097709A1 (en) 2020-03-26
CN108090508B (zh) 2020-01-31
CN110633745B (zh) 2022-11-29
EP3726426A4 (en) 2021-08-04
US11017220B2 (en) 2021-05-25
EP3726426A1 (en) 2020-10-21

Similar Documents

Publication Publication Date Title
WO2019114523A1 (zh) 分类训练方法、服务器及存储介质
WO2020221298A1 (zh) 文本检测模型训练方法、文本区域、内容确定方法和装置
US11631248B2 (en) Video watermark identification method and apparatus, device, and storage medium
US9424493B2 (en) Generic object detection in images
CN106845421B (zh) 基于多区域特征与度量学习的人脸特征识别方法及系统
WO2018219016A1 (zh) 一种人脸检测训练方法、装置及电子设备
CN107784288B (zh) 一种基于深度神经网络的迭代定位式人脸检测方法
WO2017166586A1 (zh) 基于卷积神经网络的图片鉴别方法、系统和电子设备
US10540988B2 (en) Method and apparatus for sound event detection robust to frequency change
WO2019200782A1 (zh) 样本数据分类方法、模型训练方法、电子设备及存储介质
WO2017096758A1 (zh) 图像分类方法、电子设备和存储介质
CN110008899B (zh) 一种可见光遥感图像候选目标提取与分类方法
US10163000B2 (en) Method and apparatus for determining type of movement of object in video
CN113762049B (zh) 内容识别方法、装置、存储介质和终端设备
CN112991281B (zh) 视觉检测方法、系统、电子设备及介质
CN114358279A (zh) 图像识别网络模型剪枝方法、装置、设备及存储介质
KR101584091B1 (ko) 공유 지역 특징을 이용한 다시점 객체 검출 방법
CN111950545B (zh) 一种基于MSDNet和空间划分的场景文本检测方法
CN116415020A (zh) 一种图像检索的方法、装置、电子设备及存储介质
Zulkarnain et al. Fire Detection based on Smoke Image using Convolutional Neural Network (CNN)
Ishraq et al. Assessment of Building Damage on Post-Hurricane Satellite Imagery using improved CNN
CN113205082B (zh) 基于采集不确定性解耦的鲁棒虹膜识别方法
CN114663965B (zh) 一种基于双阶段交替学习的人证比对方法和装置
US20240161460A1 (en) Self-supervised point cloud ordering using machine learning models
Aravind et al. Evaluation and Comparison of Novel Yolo V5 Algorithm in Identifying Land Cover Area of Geographical Regions over Single Shot Detector Algorithm Accuracy

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: 18887500

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 2018887500

Country of ref document: EP

Effective date: 20200713