WO2022205554A1 - Multi-stage training method for target recognition - Google Patents

Multi-stage training method for target recognition Download PDF

Info

Publication number
WO2022205554A1
WO2022205554A1 PCT/CN2021/091056 CN2021091056W WO2022205554A1 WO 2022205554 A1 WO2022205554 A1 WO 2022205554A1 CN 2021091056 W CN2021091056 W CN 2021091056W WO 2022205554 A1 WO2022205554 A1 WO 2022205554A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
training
sample
data set
target
Prior art date
Application number
PCT/CN2021/091056
Other languages
French (fr)
Chinese (zh)
Inventor
于效宇
李富超
刘艳
陈颖璐
Original Assignee
电子科技大学中山学院
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 电子科技大学中山学院 filed Critical 电子科技大学中山学院
Publication of WO2022205554A1 publication Critical patent/WO2022205554A1/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

Definitions

  • the invention relates to a training method for target recognition, image recognition and artificial intelligence, and in particular to a multi-stage training method for target recognition.
  • models In the field of target recognition, models rely on the ever-expanding dataset to achieve better and better performance, but in the process of training, most models do not consider the characteristics of the samples and the difficulty for the model to obtain features from the samples, and It is to load images randomly and equally to train the model. There is no targeted training for the samples. For the samples in the data set, the amount of feature information of each sample is different, and the model learns these features from the samples. The time required for information is also different, and the feature information between samples will always have a certain correlation. Most models do not actually consider the correlation between them, and disordered learning makes the model to a certain extent. It is impossible to grasp the feature information of the samples and the correlation between the samples, and the model cannot learn enough useful feature information from the samples during the whole learning process. These problems all affect the final performance of the model to a certain extent.
  • the present invention provides a multi-stage training method for target recognition.
  • the method trains the model through a multi-stage mode, and uses the feature information of the sample as the training basis.
  • Each stage uses data sets with different characteristics for training, so that the training of the model becomes orderly, and the performance of the final model is obtained. promote.
  • a multi-stage training method for target recognition the training method is optimized on the basis of the traditional training method.
  • the entire single training process is split into multiple training stages, and the training method includes four modules:
  • the data set preprocessing module preprocesses the used data set.
  • the main content of the preprocessing is to collect the basic characteristics of each data sample.
  • the basic characteristics of the data sample include the area of each target on the sample, the size of the target on the sample. The number, the size of the target on the sample, and the type of target on the sample;
  • the data set classification module classifies the data set according to the basic characteristics of the statistics of the data set preprocessing module, and classifies the samples that meet the set threshold into the same class.
  • the data set integration module is used to re-integrate the classified data subsets into a new data set.
  • each data subset is set with a weight parameter n i , where i is the data subset
  • the weight parameter is used to adjust the proportion of the corresponding data subset in the entire data set, so that each stage of training has different pertinence to the data set;
  • the data set loading module is used to load the integrated new data set into each stage for training. Before loading, the data set will be expanded.
  • the main method of expansion is to randomly mask the identified targets. Masking a target generation generates a new sample and assigns a new label.
  • the method includes the following steps:
  • Step 1 Preprocess the used data set
  • Step 2 Classify the dataset according to the result of preprocessing
  • Step 3 Remix the classified data subsets in different proportions to form a new data set
  • Step 4 Divide the process of training the model into multiple stages, and then load the new data set into different stages in turn.
  • step 1 The specific steps of step 1 are as follows:
  • the main content of preprocessing is to collect the basic characteristics of each data sample.
  • the basic characteristics of the data sample include the area of each target on the sample, the number of targets on the sample, and the size of the target on the sample. The size and kind of target on the sample.
  • the second step is to classify according to the results of preprocessing.
  • the classification criteria include one or more of the average target area on the sample, the number of targets on the sample, or the types of targets on the sample, and the data set is divided into multiple groups through pre-classification. A subset of data with specific characteristics.
  • step 2 The specific steps of step 2 are as follows:
  • Classify according to the results of preprocessing, and divide the entire data set into simple type data subsets and complex type data subsets according to the number of targets on the sample, and the samples with less than 5 identified targets but not the background are defined as simple Type data subsets, background samples and samples with more than 5 recognition targets are defined as complex type data subsets.
  • step 3 remix, set weight parameters for each data subset separately, and use the weight parameters to affect the proportion of each data subset in the entire data set, and then remix each data subset with different weights .
  • the weight is set, that is, each data subset is equipped with a separate weight, and the range of the weight is greater than zero, that is, the data volume of the data subset can be increased or decreased through the assigned weight. quantity purpose.
  • Multi-stage training in step 4 Multi-stage training is to divide the traditional one-time training into multiple stages of training, and the data sets imported in each stage are different, that is, there are multiple imported data sets, and each epoch is used as Each epoch is trained on the same or different datasets.
  • the data set preprocessing module reads and counts the number of targets on the sample. According to the statistical results, the data set classification module divides the COCO data set into two types of data sub-groups. Sets are simple type data subset A and complex type data subset B.
  • the classification standard is to classify the samples whose target number is less than 5 but not the background as simple type data subset A. Samples greater than 5 are classified as complex type data subset B;
  • the data subsets of simple type data subset A and complex type data subset B are not processed, that is, their corresponding weights are set to 1, and then only randomly mixed and then loaded into the model for training;
  • the data volume of simple type data subset A is reduced by K 1 times, and the data volume of complex type data subset B is increased by K 2 times, and in this process, the data volume of the two types of data subsets must be kept between The same as the original data set, and then load the model training after random mixing;
  • the data volume of simple type data subset A is reduced by K 3 times, and the data volume of complex type data subset B is increased by K 4 times, and then loaded into the model training after random mixing.
  • the present invention uses the feature information of the samples as the training basis by adding training in stages, and uses data sets with different characteristics for training in each stage, so that the training of the model becomes orderly, and the final model is trained. Performance has been improved.
  • Fig. 1 is the step flow chart of the present invention
  • Fig. 2 is the flow chart of this method
  • Figure 3 is a block flow diagram of one example of the present invention.
  • a multi-stage training method for target recognition the training method is optimized on the basis of the traditional training method, and the training method reintegrates the data set used in the training process to form each characteristic. , and split the entire single training process into multiple training stages, the training method consists of four modules:
  • the data set preprocessing module preprocesses the used data set.
  • the main content of the preprocessing is to collect the basic characteristics of each data sample.
  • the basic characteristics of the data sample include the area of each target on the sample, the size of the target on the sample. The number, the size of the target on the sample, and the type of target on the sample;
  • the data set classification module classifies the data set according to the basic characteristics of the statistics of the data set preprocessing module, and classifies the samples that meet the set threshold into the same class.
  • the data set integration module is used to re-integrate the classified data subsets into a new data set.
  • each data subset is set with a weight parameter n i , where i is the data subset
  • the weight parameter is used to adjust the proportion of the corresponding data subset in the entire data set, so that each stage of training has different pertinence to the data set;
  • the data set loading module is used to load the integrated new data set into each stage for training. Before loading, the data set will be expanded.
  • the main method of expansion is to randomly mask the identified targets. Masking a target generation generates a new sample and assigns a new label.
  • the method includes the following steps:
  • Step 1 Preprocess the used data set
  • Step 2 Classify the dataset according to the result of preprocessing
  • Step 3 Remix the classified data subsets in different proportions to form a new data set
  • Step 4 Divide the process of training the model into multiple stages, and then load the new data set into different stages in turn.
  • step 1 The specific steps of step 1 are as follows:
  • the main content of preprocessing is to collect the basic characteristics of each data sample.
  • the basic characteristics of the data sample include the area of each target on the sample, the number of targets on the sample, and the size of the target on the sample. The size and kind of target on the sample.
  • the second step is to classify according to the results of preprocessing.
  • the classification criteria include the average target area on the sample, the number of targets on the sample, or the type of targets on the sample.
  • the data set is divided into multiple data subsets with specific characteristics through pre-classification. .
  • step 2 The specific steps of step 2 are as follows:
  • Classify according to the results of preprocessing, and divide the entire data set into simple type data subsets and complex type data subsets according to the number of targets on the sample, and the samples with less than 5 identified targets but not the background are defined as simple Type data subsets, background samples and samples with more than 5 identified targets are defined as complex type data subsets.
  • step 3 remix, set weight parameters for each data subset separately, and use the weight parameters to affect the proportion of each data subset in the entire data set, and then remix each data subset with different weights .
  • the weight is set, that is, each data subset is provided with a separate weight, and the range of the weight is greater than zero, that is, the data volume of a certain data subset can be increased by the corresponding assigned weight. Or the purpose of reducing the amount of data.
  • Multi-stage training in step 4 Multi-stage training is to divide the traditional one-time training into multiple stages of training, and the data sets imported in each stage are different, that is, there are multiple imported data sets, and each epoch is used as Each epoch is trained on the same or different datasets.
  • the entire training stage of the model is divided into n training stages, each epoch can actually be used as a stage, and the traditional training method is this training Special training methods in methods.
  • the classification methods of the dataset classification module are various, including the mean target area, the number of targets on the sample, and the types of targets on the sample.
  • multiple thresholds can be set to subdivide the samples in the classification, and then the data set integration module assigns weights to the data subsets for reintegration.
  • the data set loading module it will be judged whether to do data expansion, and then the samples will be imported into the model for training.
  • the model can obtain the characteristics of the data set from the data set in this stage.
  • a specific example of this training method is: first, the training process is divided into three stages, the data set preprocessing module reads and counts the number of targets on the sample, and according to the statistical results, the data set classification module
  • the COCO data set is divided into two types of data subsets, namely simple type data subset A and complex type data subset B.
  • the classification standard is to classify the samples whose target number is less than 5 but not the background as simple type data subsets.
  • Set A classify background samples and samples with more than 5 targets as complex type data subset B;
  • the data subsets of simple type data subset A and complex type data subset B are not processed, that is, their corresponding weights are set to 1, and then only randomly mixed and then loaded into the model for training;
  • the data volume of simple type data subset A is reduced by K 1 times, and the data volume of complex type data subset B is increased by K 2 times, and in this process, the data volume of the two types of data subsets must be kept between The same as the original data set, and then load the model training after random mixing;
  • the data subset of simple type data subset A is reduced by K 3 times, and the data subset of complex type data subset B is increased by K 4 times, and then loaded into the model training after random mixing.
  • the data subsets of simple type data set A and complex type data set B are the same as the changes in the second stage, but the reduction and increase of data are different, which is to ensure that the model can be trained to different types of data as much as possible.

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Image Analysis (AREA)

Abstract

Disclosed in the present invention is a multi-stage training method for target recognition. The steps of the training method comprise: preprocessing datasets used for training; classifying a preprocessing result into multiple types of data subsets by using a dataset classification module; mixing the multiple types of data subsets in different proportions, so as to obtain multiple new datasets of different types; dividing model training into multiple training stages; and sequentially loading the multiple new datasets of different types onto different training stages for model training. By means of the present invention, the average accuracy of model identification of various categories is improved simply and efficiently by means of multi-stage training, without changing the training time and size of the original model.

Description

一种面向目标识别的多阶段训练方法A multi-stage training method for object recognition 技术领域technical field
本发明涉及目标识别的训练方法、图像识别、人工智能领域，尤其涉及一种面向目标识别的多阶段训练方法。The invention relates to a training method for target recognition, image recognition and artificial intelligence, and in particular to a multi-stage training method for target recognition.
背景技术Background technique
在目标识别领域中，模型依靠不断扩大的数据集取得越来越好的性能，但在训练的过程中，大部分模型并没有考虑样本的特性和模型从样本中获取特征的难易程度，并且是随机、平等地载入图片给模型进行训练，没有对样本进行针对性的训练，对于数据集中的样本来说，每个样本的特征信息量都是不一样的，模型从样本中学习这些特征信息所需的时间也是不一样的，而样本与样本间的特征信息总会存在一定的关联性，大部分模型其实并没有考虑到它们之间的关联性，无序性学习使模型在一定程度上无法把握样本的特征信息和样本间的关联性，模型在整个学习过程中不能从样本中学习到足够多有用的特征信息，这些问题都在一定程度上影响模型最终的性能。In the field of target recognition, models rely on the ever-expanding dataset to achieve better and better performance, but in the process of training, most models do not consider the characteristics of the samples and the difficulty for the model to obtain features from the samples, and It is to load images randomly and equally to train the model. There is no targeted training for the samples. For the samples in the data set, the amount of feature information of each sample is different, and the model learns these features from the samples. The time required for information is also different, and the feature information between samples will always have a certain correlation. Most models do not actually consider the correlation between them, and disordered learning makes the model to a certain extent. It is impossible to grasp the feature information of the samples and the correlation between the samples, and the model cannot learn enough useful feature information from the samples during the whole learning process. These problems all affect the final performance of the model to a certain extent.
发明内容SUMMARY OF THE INVENTION
为解决现有技术中存在的上述问题，本发明提供一种面向目标识别的多阶段训练方法。该方法通过多阶段的模式对模型进行训练，将样本的特征信息作为训练依据，每个阶段都采用具有不同特性的数据集进行训练，使得模型的训练变得有序，最终模型的性能得到了提升。In order to solve the above problems existing in the prior art, the present invention provides a multi-stage training method for target recognition. The method trains the model through a multi-stage mode, and uses the feature information of the sample as the training basis. Each stage uses data sets with different characteristics for training, so that the training of the model becomes orderly, and the performance of the final model is obtained. promote.
本发明为了解决其技术问题，采用下述技术方案：The present invention adopts following technical scheme in order to solve its technical problem:
一种面向目标识别的多阶段训练方法，该训练方法在传统训练方法的基础上进行优化，该训练方法在训练过程中重新整合所采用的数据集，形成各具特性的新数据集，并将整个单一训练过程拆分为多个训练阶段，该训练方法包括四个模块：A multi-stage training method for target recognition, the training method is optimized on the basis of the traditional training method. The entire single training process is split into multiple training stages, and the training method includes four modules:
数据集预处理模块，对所采用的数据集进行预处理，其预处理的主要内容是收集每个数据样本的基本特征，数据样本的基本特征包括样本上每个目标的面积、样本上目标的数量、样本上目标的大小以及样本上目标的种类；The data set preprocessing module preprocesses the used data set. The main content of the preprocessing is to collect the basic characteristics of each data sample. The basic characteristics of the data sample include the area of each target on the sample, the size of the target on the sample. The number, the size of the target on the sample, and the type of target on the sample;
数据集分类模块，根据数据集预处理模块统计的基本特征，对数据集进行分类，将符合设定阈值内的样本归为同一类，其分类的标准包括样本上的目标面积均值、样本上的目标数量，样本上目标的大小或者样本上目标的种类；The data set classification module classifies the data set according to the basic characteristics of the statistics of the data set preprocessing module, and classifies the samples that meet the set threshold into the same class. The number of targets, the size of the target on the sample or the type of target on the sample;
数据集整合模块，用于将分好类的数据子集进行重新整合成新的数据集，在重新整合的过程中，每个数据子集分别设置有权重参数n i，其中i为数据子集的序号，通过设定的权重参数来调节对应的数据子集在整个数据集的比重，使每个阶段的训练对数据集都有不同的针对性； The data set integration module is used to re-integrate the classified data subsets into a new data set. In the process of re-integration, each data subset is set with a weight parameter n i , where i is the data subset The weight parameter is used to adjust the proportion of the corresponding data subset in the entire data set, so that each stage of training has different pertinence to the data set;
数据集载入模块，用于把整合好的新数据集载入到各个阶段中进行训练，在载入前会对数据集进行扩充，其扩充的主要方法是对识别的目标进行随机掩盖，每掩盖一次目标生成就生成一个新的样本，并且分配新的标签。The data set loading module is used to load the integrated new data set into each stage for training. Before loading, the data set will be expanded. The main method of expansion is to randomly mask the identified targets. Masking a target generation generates a new sample and assigns a new label.
本方法包括以下步骤：The method includes the following steps:
步骤一、对采用的数据集进行预处理；Step 1: Preprocess the used data set;
步骤二、根据预处理的结果对数据集进行分类；Step 2: Classify the dataset according to the result of preprocessing;
步骤三、对分类好的数据子集按不同的比例重新混合形成新的数据集；Step 3: Remix the classified data subsets in different proportions to form a new data set;
步骤四、把训练模型的过程分成多个阶段，然后把新的数据集依次载入不同的阶段。Step 4. Divide the process of training the model into multiple stages, and then load the new data set into different stages in turn.
步骤一的具体步骤如下：The specific steps of step 1 are as follows:
对采用的数据集进行预处理，其预处理的主要内容是收集每个数据样本的基本特征，数据样本的基本特征包括样本上的每个目标的面积、样本上的目标数量、样本上目标的大小和样本上目标的种类。The main content of preprocessing is to collect the basic characteristics of each data sample. The basic characteristics of the data sample include the area of each target on the sample, the number of targets on the sample, and the size of the target on the sample. The size and kind of target on the sample.
步骤二根据预处理的结果进行分类，分类的标准包括样本上的目标面积均值、样本上的目标数量或者样本上目标的种类中的一种或者多种，通过预先的分类将数据集分成多个具有特定特征的数据子集。The second step is to classify according to the results of preprocessing. The classification criteria include one or more of the average target area on the sample, the number of targets on the sample, or the types of targets on the sample, and the data set is divided into multiple groups through pre-classification. A subset of data with specific characteristics.
步骤二的具体步骤如下：The specific steps of step 2 are as follows:
根据预处理的结果进行分类，并根据样本上的目标数目把将整个数据集分成简单类型数据子集和复杂类型数据子集，识别的目标数少于5个但并非是背景的样本定义为简单类型数据子集，背景样本和识别目标多于5的样本定义为复杂类型数据子集。Classify according to the results of preprocessing, and divide the entire data set into simple type data subsets and complex type data subsets according to the number of targets on the sample, and the samples with less than 5 identified targets but not the background are defined as simple Type data subsets, background samples and samples with more than 5 recognition targets are defined as complex type data subsets.
步骤三的过程中重新混合，对每个数据子集分别设置权重参数，通过权重参数来影响每个数据子集在整个数据集中的占比，然后将各个配有不同权重的数据子集重新混合。In the process of step 3, remix, set weight parameters for each data subset separately, and use the weight parameters to affect the proportion of each data subset in the entire data set, and then remix each data subset with different weights .
步骤三的过程中设置权重，即为每个数据子集都配有一个单独的权重，其权重的范围是大于零，即可以使数据子集的数据量通过分配的权重来达到增加或减少数据量的目的。In the process of step 3, the weight is set, that is, each data subset is equipped with a separate weight, and the range of the weight is greater than zero, that is, the data volume of the data subset can be increased or decreased through the assigned weight. quantity purpose.
步骤四的多阶段训练：多阶段训练是将传统的一次性训练划分为多个阶段的训练，每个阶段导入的数据集都不一样，即存在多次导入数据集，而且每个epoch都作为一个阶段来训练，每个epoch所训练的数据集都相同或者不相同。Multi-stage training in step 4: Multi-stage training is to divide the traditional one-time training into multiple stages of training, and the data sets imported in each stage are different, that is, there are multiple imported data sets, and each epoch is used as Each epoch is trained on the same or different datasets.
本方法的其中一个实例是将训练过程分成三个阶段，数据集预处理模块对样本上的目标数量进行读取和统计，根据该统计结果，数据集分类模块将COCO数据集分成两类数据子集，分别是简单类型数据子集A和复杂类型数据子集B，其分类的标准是把目标数量小于5但又并非是背景的样本归为简单类型数据子集A，把背景样本和目标数量大于5的样本归为复杂类型数据子集B；One example of this method is to divide the training process into three stages. The data set preprocessing module reads and counts the number of targets on the sample. According to the statistical results, the data set classification module divides the COCO data set into two types of data sub-groups. Sets are simple type data subset A and complex type data subset B. The classification standard is to classify the samples whose target number is less than 5 but not the background as simple type data subset A. Samples greater than 5 are classified as complex type data subset B;
在第一阶段中，简单类型数据子集A和复杂类型数据子集B的数据子集均不做处理，即其对应的权重都设置为1，然后仅随机混合然后载入模型训练；In the first stage, the data subsets of simple type data subset A and complex type data subset B are not processed, that is, their corresponding weights are set to 1, and then only randomly mixed and then loaded into the model for training;
在第二阶段中，简单类型数据子集A的数据量减少K 1倍，复杂类型数据子集B的数据量增多K 2倍，并且在该过程中要保持两类数据子集的数据量之和与原数据集一样，然后作随机混合后载入模型训练； In the second stage, the data volume of simple type data subset A is reduced by K 1 times, and the data volume of complex type data subset B is increased by K 2 times, and in this process, the data volume of the two types of data subsets must be kept between The same as the original data set, and then load the model training after random mixing;
在第三阶段中，简单类型数据子集A的数据量减少K 3倍，复杂类型数据子集B的数据量增多K 4倍，然后作随机混合后载入模型训练。 In the third stage, the data volume of simple type data subset A is reduced by K 3 times, and the data volume of complex type data subset B is increased by K 4 times, and then loaded into the model training after random mixing.
本发明的有益效果是：本发明通过加入分阶段训练，以样本的特 征信息作为训练依据，每个阶段都采用具有不同特性的数据集进行训练，使得模型的训练变得有序，最终模型的性能得到了提升。The beneficial effects of the present invention are as follows: the present invention uses the feature information of the samples as the training basis by adding training in stages, and uses data sets with different characteristics for training in each stage, so that the training of the model becomes orderly, and the final model is trained. Performance has been improved.
附图说明Description of drawings
下面结合附图和实施例对本发明进一步说明。The present invention will be further described below in conjunction with the accompanying drawings and embodiments.
图1是本发明的步骤流程图；Fig. 1 is the step flow chart of the present invention;
图2是本方法的流程框图；Fig. 2 is the flow chart of this method;
图3是本发明的一个实例的流程框图。Figure 3 is a block flow diagram of one example of the present invention.
具体实施方式Detailed ways
参照图1至图3，一种面向目标识别的多阶段训练方法，该训练方法在传统训练方法的基础上进行优化，该训练方法在训练过程中重新整合所采用的数据集，形成各具特性的新数据集，并将整个单一训练过程拆分为多个训练阶段，该训练方法包括四个模块：Referring to Figures 1 to 3, a multi-stage training method for target recognition, the training method is optimized on the basis of the traditional training method, and the training method reintegrates the data set used in the training process to form each characteristic. , and split the entire single training process into multiple training stages, the training method consists of four modules:
数据集预处理模块，对所采用的数据集进行预处理，其预处理的主要内容是收集每个数据样本的基本特征，数据样本的基本特征包括样本上每个目标的面积、样本上目标的数量、样本上目标的大小以及样本上目标的种类；The data set preprocessing module preprocesses the used data set. The main content of the preprocessing is to collect the basic characteristics of each data sample. The basic characteristics of the data sample include the area of each target on the sample, the size of the target on the sample. The number, the size of the target on the sample, and the type of target on the sample;
数据集分类模块，根据数据集预处理模块统计的基本特征，对数据集进行分类，将符合设定阈值内的样本归为同一类，其分类的标准包括样本上的目标面积均值、样本上的目标数量，样本上目标的大小或者样本上目标的种类；The data set classification module classifies the data set according to the basic characteristics of the statistics of the data set preprocessing module, and classifies the samples that meet the set threshold into the same class. The number of targets, the size of the target on the sample or the type of target on the sample;
数据集整合模块，用于将分好类的数据子集进行重新整合成新的 数据集，在重新整合的过程中，每个数据子集分别设置有权重参数n i，其中i为数据子集的序号，通过设定的权重参数来调节对应的数据子集在整个数据集的比重，使每个阶段的训练对数据集都有不同的针对性； The data set integration module is used to re-integrate the classified data subsets into a new data set. In the process of re-integration, each data subset is set with a weight parameter n i , where i is the data subset The weight parameter is used to adjust the proportion of the corresponding data subset in the entire data set, so that each stage of training has different pertinence to the data set;
数据集载入模块，用于把整合好的新数据集载入到各个阶段中进行训练，在载入前会对数据集进行扩充，其扩充的主要方法是对识别的目标进行随机掩盖，每掩盖一次目标生成就生成一个新的样本，并且分配新的标签。The data set loading module is used to load the integrated new data set into each stage for training. Before loading, the data set will be expanded. The main method of expansion is to randomly mask the identified targets. Masking a target generation generates a new sample and assigns a new label.
本方法包括以下步骤：The method includes the following steps:
步骤一、对采用的数据集进行预处理；Step 1: Preprocess the used data set;
步骤二、根据预处理的结果对数据集进行分类；Step 2: Classify the dataset according to the result of preprocessing;
步骤三、对分类好的数据子集按不同的比例重新混合形成新的数据集；Step 3: Remix the classified data subsets in different proportions to form a new data set;
步骤四、把训练模型的过程分成多个阶段，然后把新的数据集依次载入不同的阶段。Step 4. Divide the process of training the model into multiple stages, and then load the new data set into different stages in turn.
步骤一的具体步骤如下：The specific steps of step 1 are as follows:
对采用的数据集进行预处理，其预处理的主要内容是收集每个数据样本的基本特征，数据样本的基本特征包括样本上的每个目标的面积、样本上的目标数量、样本上目标的大小和样本上目标的种类。The main content of preprocessing is to collect the basic characteristics of each data sample. The basic characteristics of the data sample include the area of each target on the sample, the number of targets on the sample, and the size of the target on the sample. The size and kind of target on the sample.
步骤二根据预处理的结果进行分类，分类的标准包括样本上的目标面积均值、样本上的目标数量或者样本上目标的种类，通过预先的分类将数据集分成多个具有特定特征的数据子集。The second step is to classify according to the results of preprocessing. The classification criteria include the average target area on the sample, the number of targets on the sample, or the type of targets on the sample. The data set is divided into multiple data subsets with specific characteristics through pre-classification. .
步骤二的具体步骤如下：The specific steps of step 2 are as follows:
根据预处理的结果进行分类，并根据样本上的目标数目把将整个数据集分成简单类型数据子集和复杂类型数据子集，识别的目标数少于5个但并非是背景的样本定义为简单类型数据子集，背景样本和识别的目标数多于5的样本定义为复杂类型数据子集。Classify according to the results of preprocessing, and divide the entire data set into simple type data subsets and complex type data subsets according to the number of targets on the sample, and the samples with less than 5 identified targets but not the background are defined as simple Type data subsets, background samples and samples with more than 5 identified targets are defined as complex type data subsets.
步骤三的过程中重新混合，对每个数据子集分别设置权重参数，通过权重参数来影响每个数据子集在整个数据集中的占比，然后将各个配有不同权重的数据子集重新混合。In the process of step 3, remix, set weight parameters for each data subset separately, and use the weight parameters to affect the proportion of each data subset in the entire data set, and then remix each data subset with different weights .
步骤三的过程中设置权重，即为每个数据子集都配有一个单独的权重，其权重的范围是大于零，即可以使某个数据子集的数据量通过对应分配的权重来达到增加或减少数据量的目的。In the process of step 3, the weight is set, that is, each data subset is provided with a separate weight, and the range of the weight is greater than zero, that is, the data volume of a certain data subset can be increased by the corresponding assigned weight. Or the purpose of reducing the amount of data.
步骤四的多阶段训练：多阶段训练是将传统的一次性训练划分为多个阶段的训练，每个阶段导入的数据集都不一样，即存在多次导入数据集，而且每个epoch都作为一个阶段来训练，每个epoch所训练的数据集都相同或者不相同。Multi-stage training in step 4: Multi-stage training is to divide the traditional one-time training into multiple stages of training, and the data sets imported in each stage are different, that is, there are multiple imported data sets, and each epoch is used as Each epoch is trained on the same or different datasets.
如图2所示，该训练过程的多阶段训练方法的通用实例，模型的整个训练阶段被分成n个训练分阶段，每个epoch其实都可以作为一个分阶段，而传统的训练方法是本训练方法中的特殊训练方法。数据集分类模块的分类方法是多样的，可以包括目标面积均值、样本上的目标数量、样本上目标的种类等。对于其中一个数据集分类模块来说，可以设定多个阈值将该分类中的样本再进行划分，然后通过数据集整合模块给数据子集分配权重，进行重新整合。在数据集载入模块中会 判断是否做数据扩充，然后才将样本导入模型进行训练,对于每个训练阶段，模型都能从该阶段的数据集中获得该数据集的特性。As shown in Figure 2, a general example of the multi-stage training method of the training process, the entire training stage of the model is divided into n training stages, each epoch can actually be used as a stage, and the traditional training method is this training Special training methods in methods. The classification methods of the dataset classification module are various, including the mean target area, the number of targets on the sample, and the types of targets on the sample. For one of the data set classification modules, multiple thresholds can be set to subdivide the samples in the classification, and then the data set integration module assigns weights to the data subsets for reintegration. In the data set loading module, it will be judged whether to do data expansion, and then the samples will be imported into the model for training. For each training stage, the model can obtain the characteristics of the data set from the data set in this stage.
如图3所示，本训练方法的一个具体实例是：首先将训练过程分成三个阶段，数据集预处理模块对样本上的目标数量进行读取和统计，根据该统计结果，数据集分类模块将COCO数据集分成两类数据子集，分别是简单类型数据子集A和复杂类型数据子集B，其分类的标准是把目标数量小于5但又并非是背景的样本归为简单类型数据子集A，把背景样本和目标数量大于5的样本归为复杂类型数据子集B；As shown in Figure 3, a specific example of this training method is: first, the training process is divided into three stages, the data set preprocessing module reads and counts the number of targets on the sample, and according to the statistical results, the data set classification module The COCO data set is divided into two types of data subsets, namely simple type data subset A and complex type data subset B. The classification standard is to classify the samples whose target number is less than 5 but not the background as simple type data subsets. Set A, classify background samples and samples with more than 5 targets as complex type data subset B;
在第一阶段中，简单类型数据子集A和复杂类型数据子集B的数据子集均不做处理，即其对应的权重设置为1，然后仅随机混合然后载入模型训练；In the first stage, the data subsets of simple type data subset A and complex type data subset B are not processed, that is, their corresponding weights are set to 1, and then only randomly mixed and then loaded into the model for training;
在第二阶段中，简单类型数据子集A的数据量减少K 1倍，复杂类型数据子集B的数据量增多K 2倍，并且在该过程中要保持两类数据子集的数据量之和与原数据集一样，然后作随机混合后载入模型训练； In the second stage, the data volume of simple type data subset A is reduced by K 1 times, and the data volume of complex type data subset B is increased by K 2 times, and in this process, the data volume of the two types of data subsets must be kept between The same as the original data set, and then load the model training after random mixing;
在第三阶段中，简单类型数据子集A的数据子集减少K 3倍，复杂类型数据子集B的数据子集增多K 4倍，然后作随机混合后载入模型训练,虽然该阶段中简单类型数据集A和复杂类型数据集B的数据子集与第二阶段的变化一样，但其减少和增加的数据是有差别，这是为了保证模型能尽可能训练到不同的类型的数据。 In the third stage, the data subset of simple type data subset A is reduced by K 3 times, and the data subset of complex type data subset B is increased by K 4 times, and then loaded into the model training after random mixing. The data subsets of simple type data set A and complex type data set B are the same as the changes in the second stage, but the reduction and increase of data are different, which is to ensure that the model can be trained to different types of data as much as possible.
以上的实施方式不能限定本发明创造的保护范围，专业技术领域的人员在不脱离本发明创造整体构思的情况下，所做的均等修饰与变化，均仍属于本发明创造涵盖的范围之内。The above embodiments cannot limit the protection scope of the present invention, and those in the technical field can make equal modifications and changes without departing from the overall concept of the present invention. All still fall within the scope of the present invention.

Claims (9)

  1. 一种面向目标识别的多阶段训练方法，该训练方法在传统训练方法的基础上进行优化，其特征在于该训练方法在训练过程中重新整合所采用的数据集，形成各具特性的新数据集，并将整个单一训练过程拆分为多个训练阶段，该训练方法包括四个模块：A multi-stage training method for target recognition, the training method is optimized on the basis of the traditional training method, and is characterized in that the training method re-integrates the data set used in the training process to form a new data set with each characteristic , and split the entire single training process into multiple training stages, the training method includes four modules:
    数据集预处理模块，对所采用的数据集进行预处理，其预处理的主要内容是收集每个数据样本的基本特征，数据样本的基本特征包括样本上每个目标的面积、样本上目标的数量、样本上目标的大小以及样本上目标的种类；The data set preprocessing module preprocesses the used data set. The main content of the preprocessing is to collect the basic characteristics of each data sample. The basic characteristics of the data sample include the area of each target on the sample, the size of the target on the sample. The number, the size of the target on the sample, and the type of target on the sample;
    数据集分类模块，根据数据集预处理模块统计的基本特征，对数据集进行分类，将符合设定阈值内的样本归为同一类，其分类的标准包括样本上的目标面积均值、样本上的目标数量，样本上目标的大小或者样本上目标的种类；The data set classification module classifies the data set according to the basic characteristics of the statistics of the data set preprocessing module, and classifies the samples that meet the set threshold into the same class. The number of targets, the size of the target on the sample or the type of target on the sample;
    数据集整合模块，用于将分好类的数据子集进行重新整合成新的数据集，在重新整合的过程中，每个数据子集分别设置有权重参数n i，其中i为数据子集的序号，通过设定的权重参数来调节对应的数据子集在整个数据集的比重，使每个阶段的训练对数据集都有不同的针对性； The data set integration module is used to re-integrate the classified data subsets into a new data set. In the process of re-integration, each data subset is set with a weight parameter n i , where i is the data subset The weight parameter is used to adjust the proportion of the corresponding data subset in the entire data set, so that each stage of training has different pertinence to the data set;
    数据集载入模块，用于把整合好的新数据集载入到各个阶段中进行训练，在载入前会对数据集进行扩充，其扩充的主要方法是对识别的目标进行随机掩盖，每掩盖一次目标生成就生成一个新的样本，并且分配新的标签。The data set loading module is used to load the integrated new data set into each stage for training. Before loading, the data set will be expanded. The main method of expansion is to randomly mask the identified targets. Masking a target generation generates a new sample and assigns a new label.
  2. 根据权利要求1所述的面向目标识别的多阶段训练方法，其特征在于包括以下步骤：The multi-stage training method for target recognition according to claim 1, is characterized in that comprising the following steps:
    步骤一、对采用的数据集进行预处理；Step 1: Preprocess the used data set;
    步骤二、根据预处理的结果对数据集进行分类；Step 2: Classify the dataset according to the result of preprocessing;
    步骤三、对分类好的数据子集按不同的比例重新混合形成新的数据集；Step 3: Remix the classified data subsets in different proportions to form a new data set;
    步骤四、把训练模型的过程分成多个阶段，然后把新的数据集依次载入不同的阶段。Step 4. Divide the process of training the model into multiple stages, and then load the new data set into different stages in turn.
  3. 根据权利要求2所述的面向目标识别的多阶段训练方法，其特征在于步骤一的具体步骤如下：The multi-stage training method for target recognition according to claim 2 is characterized in that the concrete steps of step 1 are as follows:
    对采用的数据集进行预处理，其预处理的主要内容是收集每个数据样本的基本特征，数据样本的基本特征包括样本上的每个目标的面积、样本上的目标数量、样本上目标的大小和样本上目标的种类。The main content of preprocessing is to collect the basic characteristics of each data sample. The basic characteristics of the data sample include the area of each target on the sample, the number of targets on the sample, and the size of the target on the sample. The size and kind of target on the sample.
  4. 根据权利要求3所述的面向目标识别的多阶段训练方法，其特征在于步骤二根据预处理的结果进行分类，分类的标准包括样本上的目标面积均值、样本上的目标数量、样本上目标的大小或者样本上目标的种类，通过预先的分类将数据集分成多个具有特定特征的数据子集。The multi-stage training method for target recognition according to claim 3, characterized in that the second step is to classify according to the result of preprocessing, and the classification criteria include the mean value of the target area on the sample, the number of targets on the sample, and the number of targets on the sample. The size or the type of target on the sample, the data set is divided into multiple data subsets with specific characteristics through pre-classification.
  5. 根据权利要求3所述的面向目标识别的多阶段训练方法，其特征在于步骤二的具体步骤如下：The multi-stage training method for target recognition according to claim 3 is characterized in that the concrete steps of step 2 are as follows:
    根据预处理的结果进行分类，并根据样本上的目标数目把将整个数据集分成简单类型数据子集和复杂类型数据子集，识别的目标数少于5个但并非是背景的样本定义为简单类型数据子集，背景样本和识别的 目标数多于5的样本定义为复杂类型数据子集。Classify according to the results of preprocessing, and divide the entire data set into simple type data subsets and complex type data subsets according to the number of targets on the sample, and the samples with less than 5 identified targets but not the background are defined as simple Type data subsets, background samples and samples with more than 5 identified targets are defined as complex type data subsets.
  6. 根据权利要求5所述的面向目标识别的多阶段训练方法，其特征在于步骤三的过程中重新混合，对每个数据子集分别设置权重参数，通过权重参数来影响每个数据子集在整个数据集中的占比，然后将各个配有不同权重的数据子集重新混合。The multi-stage training method for target recognition according to claim 5, characterized in that in the process of step 3, remixing is performed, weight parameters are set for each data subset respectively, and each data subset is affected by the weight parameter in the whole proportions in the data set, and then remix each subset of the data with different weights.
  7. 根据权利要求6所述的面向目标识别的多阶段训练方法，其特征在于步骤三的过程中设置权重，即为每个数据子集都配有一个单独的权重，其权重的范围是大于零，即可以使数据子集的数据量通过分配的权重来达到增加或减少数据量的目的。The multi-stage training method for target recognition according to claim 6, wherein the weight is set in the process of step 3, that is, each data subset is provided with a separate weight, and the range of the weight is greater than zero, That is, the data amount of the data subset can be increased or decreased by the assigned weight.
  8. 根据权利要求7所述的面向目标识别的多阶段训练方法，其特征在于步骤四的多阶段训练：多阶段训练是将传统的一次性训练划分为多个阶段的训练，每个阶段导入的数据集不一样，即存在多次导入数据集，而且每个epoch都作为一个阶段来训练，每个epoch所训练的数据集都相同或者不相同。The multi-stage training method for target recognition according to claim 7, characterized in that the multi-stage training in step 4: the multi-stage training is to divide the traditional one-time training into a plurality of stages of training, and the data imported in each stage The sets are different, that is, there are multiple imported data sets, and each epoch is trained as a stage, and the data sets trained by each epoch are the same or different.
  9. 根据权利要求1或2所述的面向目标识别的多阶段训练方法，其特征在于将训练过程分成三个阶段，数据集预处理模块对样本上的目标数量进行读取和统计，根据该统计结果，数据集分类模块将COCO数据集分成两类数据子集，分别是简单类型数据子集A和复杂类型数据子集B，其分类的标准是把目标数量小于5的样本但又并非是背景样本归为简单类型数据子集A，把背景样本和目标数量大于5的样本归为复杂类型数据子集B；The multi-stage training method for target recognition according to claim 1 or 2, wherein the training process is divided into three stages, the data set preprocessing module reads and counts the number of targets on the sample, and according to the statistical result , the data set classification module divides the COCO data set into two types of data subsets, namely simple type data subset A and complex type data subset B. The classification standard is to classify samples with a target number of less than 5 but not background samples. It is classified as a simple type data subset A, and the background samples and samples with a target number greater than 5 are classified as a complex type data subset B;
    在第一阶段中，简单类型数据子集A和复杂类型数据子集B的数 据子集均不做处理，即其对应的权重都设置为1，然后仅作随机混合后载入模型训练；In the first stage, the data subsets of simple type data subset A and complex type data subset B are not processed, that is, their corresponding weights are set to 1, and then they are only randomly mixed and loaded into the model for training;
    在第二阶段中，简单类型数据子集A的数据量减少K 1倍，复杂类型数据子集B的数据量增多K 2倍，并且在该过程中要保持两类数据子集的数据量之和与原数据集一样，然后作随机混合后载入模型训练； In the second stage, the data volume of simple type data subset A is reduced by K 1 times, and the data volume of complex type data subset B is increased by K 2 times, and in this process, the data volume of the two types of data subsets must be kept between The same as the original data set, and then load the model training after random mixing;
    在第三阶段中，简单类型数据子集A的数据量减少K 3倍，复杂类型数据子集B的数据量增多K 4倍，然后作随机混合后载入模型训练。 In the third stage, the data volume of simple type data subset A is reduced by K 3 times, and the data volume of complex type data subset B is increased by K 4 times, and then loaded into the model training after random mixing.
PCT/CN2021/091056 2021-03-31 2021-04-29 Multi-stage training method for target recognition WO2022205554A1 (en)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110348354.5A CN112990337B (en) 2021-03-31 2021-03-31 Multi-stage training method for target identification
CN202110348354.5 2021-03-31

Publications (1)

Publication Number Publication Date
WO2022205554A1 true WO2022205554A1 (en) 2022-10-06

Family

ID=76338647

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/091056 WO2022205554A1 (en) 2021-03-31 2021-04-29 Multi-stage training method for target recognition

Country Status (2)

Country Link
CN (1) CN112990337B (en)
WO (1) WO2022205554A1 (en)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP0588074A2 (en) * 1992-08-18 1994-03-23 Eastman Kodak Company Method and apparatus for character recognition with supervised training
CN106778853A (en) * 2016-12-07 2017-05-31 中南大学 Unbalanced data sorting technique based on weight cluster and sub- sampling
CN108595585A (en) * 2018-04-18 2018-09-28 平安科技（深圳）有限公司 Sample data sorting technique, model training method, electronic equipment and storage medium
CN112215303A (en) * 2020-11-05 2021-01-12 北京理工大学 Image understanding method and system based on self-learning attribute

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20120209880A1 (en) * 2011-02-15 2012-08-16 General Electric Company Method of constructing a mixture model
CN104217219A (en) * 2014-09-15 2014-12-17 西安电子科技大学 Polarization SAR image classification method based on matching pursuit selection integration
CN108345904A (en) * 2018-01-26 2018-07-31 华南理工大学 A kind of Ensemble Learning Algorithms of the unbalanced data based on the sampling of random susceptibility
CN108460523B (en) * 2018-02-12 2020-08-21 阿里巴巴集团控股有限公司 Wind control rule generation method and device
CN110414587A (en) * 2019-07-23 2019-11-05 南京邮电大学 Depth convolutional neural networks training method and system based on progressive learning
CN111950630B (en) * 2020-08-12 2022-08-02 深圳市烨嘉为技术有限公司 Small sample industrial product defect classification method based on two-stage transfer learning

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP0588074A2 (en) * 1992-08-18 1994-03-23 Eastman Kodak Company Method and apparatus for character recognition with supervised training
CN106778853A (en) * 2016-12-07 2017-05-31 中南大学 Unbalanced data sorting technique based on weight cluster and sub- sampling
CN108595585A (en) * 2018-04-18 2018-09-28 平安科技（深圳）有限公司 Sample data sorting technique, model training method, electronic equipment and storage medium
CN112215303A (en) * 2020-11-05 2021-01-12 北京理工大学 Image understanding method and system based on self-learning attribute

Also Published As

Publication number Publication date
CN112990337A (en) 2021-06-18
CN112990337B (en) 2022-11-29

Similar Documents

Publication Publication Date Title
WO2021134871A1 (en) Forensics method for synthesized face image based on local binary pattern and deep learning
CN109800778B (en) Faster RCNN target detection method based on difficultly-divided sample mining
CN108256482B (en) Face age estimation method for distributed learning based on convolutional neural network
CN108197643B (en) Transfer learning method based on unsupervised clustering and metric learning
CN111523579B (en) Vehicle type recognition method and system based on improved deep learning
CN109635830B (en) Screening method for effective data for estimating automobile quality
CN109740603A (en) Based on the vehicle character identifying method under CNN convolutional neural networks
CN108021908B (en) Face age group identification method and device, computer device and readable storage medium
CN104850890A (en) Method for adjusting parameter of convolution neural network based on example learning and Sadowsky distribution
CN106845525A (en) A kind of depth confidence network image bracket protocol based on bottom fusion feature
CN105304078A (en) Target sound data training device and target sound data training method
CN110348448A (en) A kind of license plate character recognition method based on convolutional neural networks
CN103077399A (en) Biological microscopic image classification method based on integrated cascade structure
Gao et al. Bottom-up and top-down: Bidirectional additive net for edge detection
CN111310820A (en) Foundation meteorological cloud chart classification method based on cross validation depth CNN feature integration
CN104537385A (en) Method for evaluating DAGSVM classification accuracy
WO2022205554A1 (en) Multi-stage training method for target recognition
CN104468276B (en) Network flow identification method based on random sampling multi-categorizer
CN117037006A (en) Unmanned aerial vehicle tracking method with high endurance capacity
CN110610203A (en) Electric energy quality disturbance classification method based on DWT and extreme learning machine
Huo et al. Traffic sign recognition based on resnet-20 and deep mutual learning
Yao et al. Research and comparison of ship classification algorithms based on variant CNNs
CN114332491A (en) Saliency target detection algorithm based on feature reconstruction
He et al. Classification of Galaxy Morphology Based on Multi-Channel Deep Residual Networks
Cui Research on Recognition and Classification Technology Based on Deep Convolutional Neural Network

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: 21934205

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: 21934205

Country of ref document: EP

Kind code of ref document: A1