WO2020232840A1 - 基于神经网络结构搜索的车辆多属性识别方法、装置、介质 - Google Patents

基于神经网络结构搜索的车辆多属性识别方法、装置、介质 Download PDF

Info

Publication number
WO2020232840A1
WO2020232840A1 PCT/CN2019/100027 CN2019100027W WO2020232840A1 WO 2020232840 A1 WO2020232840 A1 WO 2020232840A1 CN 2019100027 W CN2019100027 W CN 2019100027W WO 2020232840 A1 WO2020232840 A1 WO 2020232840A1
Authority
WO
WIPO (PCT)
Prior art keywords
network
attribute
vehicle
recognition
branch
Prior art date
Application number
PCT/CN2019/100027
Other languages
English (en)
French (fr)
Inventor
林修明
周成祖
王海滨
林淑强
张永光
Original Assignee
厦门市美亚柏科信息股份有限公司
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 厦门市美亚柏科信息股份有限公司 filed Critical 厦门市美亚柏科信息股份有限公司
Publication of WO2020232840A1 publication Critical patent/WO2020232840A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/55Clustering; Classification
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods

Definitions

  • the present disclosure relates to the field of artificial intelligence technology, in particular to a vehicle multi-attribute recognition method, device and storage medium based on neural network structure search.
  • the present disclosure proposes the following technical solutions in view of the above-mentioned defects in the prior art.
  • a vehicle multi-attribute recognition method based on neural network structure search includes:
  • the determining step is to determine the basic network structure of M neural network models as the basic unit Ci of the network search space, where 1 ⁇ i ⁇ M, and M is an integer greater than 0;
  • the benchmark model selection step is to pre-train the M neural network models for the first attribute in the vehicle's multi-attribute, and compare the recognition results of the M neural network models after training, and select the best recognition performance
  • the neural network model of as the reference neural network model Base the number of units is N, where N is an integer greater than 0;
  • the first half of the base neural network model Base is used as the low-level feature extractor of the vehicle multi-attribute recognition model, and the end of the first half of the network is used as the starting branch point of the entire vehicle multi-attribute recognition network, where the first half The network contains N/2 units;
  • the step of constructing a vehicle multi-attribute recognition network is to apply N/4 different basic units Ci as the starting branch point for the second attribute in the vehicle’s multi-attribute. Identify the attributes, compare the recognition results, select a basic unit with the best recognition performance for the second attribute as the network structure of N/4 units after the branch point, repeat the above operations, and select the best for other attributes of the vehicle Basic unit to build a multi-branch vehicle multi-attribute recognition network;
  • the attributes of the vehicle are recognized using the vehicle multi-attribute recognition network.
  • the vehicle multi-attributes are the model, body color, car model, and license plate type of the vehicle in the picture, wherein the first attribute is the vehicle model, and the second attribute is the body color.
  • a backtracking correction mechanism is used to correct the vehicle multi-attribute recognition network.
  • the attribute branch reversely reduces the number of network layers in the next search process.
  • the basic unit selection and correction mechanism is used to correct the vehicle multi-attribute recognition network.
  • an attribute branch of the vehicle multi-attribute recognition network if in a search, a specific When the basic unit has a greater impact on its recognition accuracy, special processing is performed on this unit in the next search.
  • the present disclosure also proposes a vehicle multi-attribute recognition device based on neural network structure search, which includes:
  • the determining module determines the basic network structure of M neural network models as the basic unit Ci of the network search space, where 1 ⁇ i ⁇ M, and M is an integer greater than 0;
  • the reference model selection module pre-trains the M neural network models for the first attribute of the vehicle's multi-attribute, and compares the recognition results of the M neural network models after training, and selects the best recognition performance
  • the neural network model of as the reference neural network model Base the number of units is N, where N is an integer greater than 0;
  • the branch determination module takes the first half of the reference neural network model Base as the low-level feature extractor of the vehicle multi-attribute recognition model, and the end of the first half of the network is used as the starting branch point of the entire vehicle multi-attribute recognition network, where the first half
  • the network contains N/2 units;
  • the vehicle multi-attribute recognition network building module uses N/4 different basic units Ci as the starting branch point for the second attribute in the vehicle’s multi-attribute. Identify the attributes, compare the recognition results, select a basic unit with the best recognition performance for the second attribute as the network structure of N/4 units after the branch point, repeat the above operations, and select the best for other attributes of the vehicle Basic unit to build a multi-branch vehicle multi-attribute recognition network;
  • the recognition module uses the vehicle multi-attribute recognition network to recognize the attributes of the vehicle.
  • the vehicle multi-attributes are the model, body color, car model, and license plate type of the vehicle in the picture, wherein the first attribute is the vehicle model, and the second attribute is the body color.
  • the vehicle multi-attribute recognition network building module uses a backtracking correction mechanism to correct the vehicle multi-attribute recognition network.
  • a backtracking correction mechanism to correct the vehicle multi-attribute recognition network.
  • the vehicle multi-attribute recognition network building module uses a basic unit selection correction mechanism to correct the vehicle multi-attribute recognition network.
  • a basic unit selection correction mechanism to correct the vehicle multi-attribute recognition network.
  • the present disclosure also proposes a computer-readable storage medium with computer program code stored on the storage medium, and when the computer program code is executed by a computer, any one of the above methods is executed.
  • the technical effect of the present disclosure is: first select a reference neural network model based on the vehicle attributes that are difficult to identify, and then select the first half of the network (including N/2 units) as the entire vehicle multi-attribute recognition network based on the reference neural network model
  • the low-level feature extractor of the model uses its end as a branch point to construct recognition branches for other attributes, and uses a similar dichotomy method to determine the remaining structure of each branch network, that is, using the N/ selected in the first iteration 4 basic network units, use the N/8 basic network units selected the second time, and so on.
  • the multi-branch vehicle multi-attribute recognition network constructed through the above method of constructing a multi-branch network combined with neural network structure search technology can identify multiple attributes of the vehicle at a time, such as model, color, etc., and improve the recognition efficiency.
  • the two network correction mechanisms proposed in the present disclosure can make the generated multi-branch vehicle multi-attribute recognition network have better performance and higher efficiency in network construction, which is another important disclosure point of the present disclosure.
  • Fig. 1 is a flowchart of a vehicle multi-attribute recognition method based on neural network structure search according to an embodiment of the present disclosure.
  • Figure 2 is a schematic diagram of an inception structure of googlenet.
  • Fig. 3 is a structural diagram of a vehicle multi-attribute recognition device based on neural network structure search according to an embodiment of the present disclosure.
  • Fig. 1 shows a vehicle multi-attribute recognition method based on neural network structure search according to the present disclosure.
  • the method includes:
  • the determination step S101 is to determine the basic network structure of M neural network models as the basic unit Ci of the network search space, where 1 ⁇ i ⁇ M, and M is an integer greater than 0.
  • existing network models such as VGG, googlenet, resnet, etc.
  • the network repeatability is high, usually multiple superimpositions of a single network structure, such as googlenet, which is superimposed by multiple inception structures, Figure 2 An inception of googlenet is shown.
  • Benchmark model selection step S102 pre-training the M neural network models for the first attribute of the vehicle's multi-attribute, and comparing them according to the recognition results of the M neural network models after training, and selecting the best recognition performance
  • the best neural network model is used as the reference neural network model Base, and the number of units is N, where N is an integer greater than 0.
  • the identification of the vehicle model is the most challenging compared to the color. That is, when building a vehicle multi-attribute recognition network, start with the most difficult to identify vehicle attributes, and pre-train multiple neural network models (googlenet, resnet, resneXt, SEnet, SKNet, etc.) on this single attribute and identify them Compare the results, use the best model as the base neural network model Base, and record the number of units N.
  • the first half of the base neural network model Base is taken as the low-level feature extractor of the vehicle multi-attribute recognition model, and the end of the first half of the network is used as the starting branch point of the entire vehicle multi-attribute recognition network, wherein the first half A network contains N/2 units.
  • the entire network can be regarded as a process of transforming and mapping low-level features to high-level semantic features. Therefore, the first half of the network (including N/2 units) of the Base model is taken as the low-level feature extractor of the entire vehicle multi-attribute recognition network model, and its end is used as the starting branch point of the entire vehicle multi-attribute recognition multi-label classification network.
  • the vehicle multi-attribute recognition network construction step S104 for the second attribute in the multi-attribute of the vehicle, respectively apply N/4 different basic units Ci as the starting branch point.
  • the basic structure of the branch neural network is for different first attributes. Recognize the two attributes, compare the recognition results, select a basic unit with the best recognition performance for the second attribute as the network structure of N/4 units after the branch point, repeat the above operation, and select the best for other attributes of the vehicle
  • the remaining branch network (that is, removing the car model attribute branch) is used to perform a binary search in the determined network search space, that is, N/ Four different basic units Ci are used as the basic structure of the branch network to identify other attributes and compare the results to determine the network structure of N/4 units after the branch point; take the color attribute branch as an example, in the first In the second iteration, N/4 predefined basic units (a total of M) are connected after the starting branch point of the Base network, so in this iteration, M different color attribute candidate branch networks can be obtained, and each candidate is compared.
  • the recognition rate of the vehicle color by the branch network from which the highest recognition rate is selected as the basic unit of the color branch.
  • a reference neural network model is first selected based on the vehicle attributes that are difficult to recognize, and then on the basis of the reference neural network model, the first half of the network (including N/2 units) is selected as the low-level feature of the entire vehicle multi-attribute recognition network model
  • the extractor uses its end as a branch point to construct recognition branches for other attributes, and uses a method similar to dichotomy to determine the structure of the network, that is, using the N/4 basic network units selected in the first iteration and using The N/8 basic network units selected for the second time, and so on.
  • a multi-branch vehicle multi-attribute recognition network can be constructed, which can identify multiple attributes of the vehicle at a time, such as model, color, etc., and improve the recognition efficiency.
  • a correction mechanism for the network is designed.
  • the backtracking correction mechanism is used to correct the vehicle multi-attribute recognition network, and the network is added to an attribute branch of the vehicle multi-attribute recognition network. If it is deep but cannot improve the recognition accuracy, the attribute branch will be reversed to reduce the number of network layers in the next search process.
  • the two network correction mechanisms proposed in the present disclosure can make the generated multi-branch vehicle multi-attribute recognition network have better performance and higher efficiency in network construction, which is another important disclosure point of the present disclosure.
  • the present disclosure provides an embodiment of a vehicle multi-attribute recognition device based on neural network structure search, which is similar to the method shown in FIG.
  • the device can be included in various electronic devices or multiple electronic devices connected via a network.
  • Fig. 3 shows the present disclosure
  • Fig. 1 shows a vehicle multi-attribute recognition device based on neural network structure search of the present disclosure, which includes:
  • the reference model selection module 302 pre-trains the M neural network models for the first attribute in the vehicle's multi-attribute, and compares the recognition results of the M neural network models after training, and selects the best recognition performance
  • the best neural network model is used as the reference neural network model Base, and the number of units is N, where N is an integer greater than 0.
  • the branch determination module 303 takes the first half of the base neural network model Base as the low-level feature extractor of the vehicle multi-attribute recognition model, and the end of the first half of the network is used as the starting branch point of the entire vehicle multi-attribute recognition network, where the first half A network contains N/2 units.
  • the entire network can be regarded as a process of transforming and mapping low-level features to high-level semantic features. Therefore, the first half of the network (including N/2 units) of the Base model is taken as the low-level feature extractor of the entire vehicle multi-attribute recognition network model, and its end is used as the starting branch point of the entire vehicle multi-attribute recognition multi-label classification network.
  • the vehicle multi-attribute recognition network construction module 304 applies N/4 different basic units Ci as the starting branch point for the second attribute in the vehicle’s multi-attribute. Recognize the two attributes, compare the recognition results, select a basic unit with the best recognition performance for the second attribute as the network structure of N/4 units after the branch point, repeat the above operation, and select the best for other attributes of the vehicle.
  • the remaining branch network (that is, removing the car model attribute branch) is used to perform a binary search in the determined network search space, that is, N/ Four different basic units Ci are used as the basic structure of the branch network to identify other attributes and compare the results to determine the network structure of N/4 units after the branch point; take the color attribute branch as an example, in the first In the second iteration, N/4 predefined basic units (a total of M) are connected after the starting branch point of the Base network, so in this iteration, M different color attribute candidate branch networks can be obtained, and each candidate is compared.
  • the recognition rate of the vehicle color by the branch network from which the highest recognition rate is selected as the basic unit of the color branch.
  • the multiple attributes of the vehicle are the model, body color, model, and license plate type of the vehicle in the picture, where the first attribute is the model of the vehicle and the second attribute is the color of the vehicle.
  • M 5
  • the neural network models are: googlenet, resnet, resneXt, SEnet and SKNet.
  • a reference neural network model is first selected based on the vehicle attributes that are difficult to recognize, and then on the basis of the reference neural network model, the first half of the network (including N/2 units) is selected as the low-level feature of the entire vehicle multi-attribute recognition network model
  • the extractor uses its end as a branch point to construct recognition branches for other attributes, and uses a method similar to dichotomy to determine the structure of the network, that is, using the N/4 basic network units selected in the first iteration and using The N/8 basic network units selected for the second time can be deduced by analogy.
  • a multi-branch vehicle multi-attribute recognition network can be constructed, which can identify multiple attributes of the vehicle at a time, such as model, color, etc., and improve the recognition efficiency.
  • the recognition module 305 uses the vehicle multi-attribute recognition network to recognize the attributes of the vehicle.
  • the vehicle multi-attribute recognition network is used to recognize the input vehicle picture, and multiple attributes of the vehicle, such as model, color, etc., can be recognized at one time, which improves the recognition efficiency.
  • a correction mechanism of the network is designed.
  • the vehicle multi-attribute recognition network construction module 304 uses a backtracking correction mechanism to correct the vehicle multi-attribute recognition network, and adds a network to an attribute branch of the vehicle multi-attribute recognition network. If it is deep but cannot improve the recognition accuracy, the attribute branch will be reversed to reduce the number of network layers in the next search process.
  • the depth of this branch network should be N/2+N/4+N/8
  • the recognition rate of the candidate network is less than the recognition rate of the candidate network obtained in the first iteration process (its network depth is N/2+N/4)
  • another correction mechanism of the network is designed.
  • the basic unit selection correction mechanism is used in the vehicle multi-attribute recognition network building module 304 to correct the vehicle multi-attribute recognition network.
  • special processing is performed on this unit in the next search. That is, different Ci is weighted by the recognition rate of the candidate network.
  • the basic unit of the worst candidate network does not participate in the next iteration process to reduce the search time. Take the color attribute branch as an example. If in the first iterative search process, if the candidate network using the inception v1 basic unit has the worst recognition rate, then in the second iterative search process, remove inception v1 from the candidate basic unit Basic unit.
  • the two network correction mechanisms proposed in the present disclosure can make the generated multi-branch vehicle multi-attribute recognition network have better performance and higher efficiency in network construction, which is another important disclosure point of the present disclosure.
  • the present disclosure applies the network structure search technology to the multi-label classification vehicle multi-attribute recognition task.
  • the network search space and path are solidified, and the time complexity of the network structure search can be limited.
  • the same method can also be extended to other multi-label classification tasks.
  • the present disclosure can be implemented by means of software plus a necessary general hardware platform.
  • the technical solution of the present disclosure essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a storage medium, such as ROM/RAM, magnetic disk , CD-ROM, etc., including a number of instructions to enable a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments of the present disclosure.

Abstract

本公开提供了一种基于神经网络结构搜索的车辆多属性识别方法、装置及存储介质,该方法先基于难以识别的车辆属性选择基准神经网络模型,然后在该基准神经网络模型的基础上,选择前半个网络作为整个车辆多属性识别网络模型的低层特征提取器,然后,其末端作为分支点,为其他的属性构建识别分支,且采用类似二分法的方法确定网络的结构,通过上述结合神经网络结构搜索技术构建的多分支的车辆多属性识别网络,可以一次识别车辆的多个属性,如型号、颜色等等,提高了识别效率。本公开提出的两种网络更正机制可以使得生成的多分支的车辆多属性识别网络性能更佳,且在网络构建时效率更高。

Description

基于神经网络结构搜索的车辆多属性识别方法、装置、介质
相关申请
本申请要求保护在2019年5月23日提交的申请号为201910436423.0的中国专利申请的优先权,该申请的全部内容以引用的方式结合到本文中。
技术领域
本公开涉及人工智能技术领域,特别是一种基于神经网络结构搜索的车辆多属性识别方法、装置及存储介质。
背景技术
在现实生活中,人们所看到的某一景物往往具有多种标签属性,如一辆车会有不同的颜色、车型、车系以及车牌类型等等。传统的机器学习模型往往需要对单个的标签属性人工设计不同的特征以达到分类识别的目的。而随着近年来深度学习的兴起,神经网络逐渐替代传统的特征工程成为一种主流的图像分类方法。得益于其设计上的得天独厚的分层架构,使得使用多分支结构同时识别图片的多种属性标签成为可能,而这种并行的结构无疑可以有效提高效率,故使得多标签分类成为图像分类应用中一种重要的技术方向。
对于现有的借助于神经网络的多标签分类技术而言,如何去确定其分支结构成为了其瓶颈所在。工程师往往需要历时长久的网络工程才能得到一种符合应用场景的网络架构。而网络结构搜索技术被认为是解决这一问题的关键,然而现有的方法耗时良久,如谷歌公司所发布的NASNet网络需耗时1800个GPU日(即单个GPU需要运行1800天)以及基于进化算法所得的AmobaNet需耗时3150个GPU日。这些网络结构还仅仅只是应用于单任务的分类识别,显然这无法满足企业应用的需求。与此同时,对于多标签分类应用而言,不同的属性标签的分类难度不一,这种单分类任务的网络结构搜索方法不具有通用性。
公开内容
本公开针对上述现有技术中的缺陷,提出了如下技术方案。
一种基于神经网络结构搜索的车辆多属性识别方法,该方法包括:
确定步骤,确定M个神经网络模型的基本网络结构作为网络搜索空间的基本单元Ci,其中,1≤i≤M,M为大于0的整数;
基准模型选择步骤,对所述车辆的多属性中的第一属性预训练所述M个神经网络模型,并根据训练后的所述M个神经网络模型的识别结果进行比较,选择识别性能最佳的神经网 络模型作为基准神经网络模型Base,其单元数目为N,其中,N为大于0的整数;
分支确定步骤,取基准神经网络模型Base的前半个网络作为车辆多属性识别模型的低层特征提取器,前半个网络的末端作为整个车辆多属性识别网络的起始分支点,其中,所述前半个网络包含N/2个单元;
车辆多属性识别网络构建步骤,针对所述车辆的多属性中的第二属性,分别应用N/4个不同的基本单元Ci作为起始分支点后的分支神经网络的基本结构对不同的第二属性进行识别,并比较识别结果,为第二属性选择一个识别性能最佳的基本单元作为分支点后N/4个单元的网络结构,重复上述操作,为车辆多属性的其他属性选择最佳的基本单元,以构建多分支的车辆多属性识别网络;
识别步骤,使用所述车辆多属性识别网络对车辆的属性进行识别。
更进一步地,所述车辆多属性为图片中车辆的车系型号、车身颜色、车型以及车牌类型,其中第一属性为车系型号,第二属性为车身颜色。
更进一步地,所述M=5,所述神经网络模型分别为:googlenet、resnet、resneXt、SENet和SKNet。
更进一步地,所述车辆多属性识别网络构建步骤中使用回溯更正机制对车辆多属性识别网络进行校正,在对车辆多属性识别网络的一属性分支增加网络深度但无法提高识别精度,则将该属性分支在下一次的搜索过程中反向减少其网络层数。
更进一步地,所述车辆多属性识别网络构建步骤中使用基本单元选取更正机制对车辆多属性识别网络进行校正,在对车辆多属性识别网络的一属性分支中,若在一次搜索中,特定的基本单元对其识别准确率有较大影响时,则在下一次搜索对该单元做特殊处理。
本公开还提出了一种基于神经网络结构搜索的车辆多属性识别装置,该装置包括:
确定模块,确定M个神经网络模型的基本网络结构作为网络搜索空间的基本单元Ci,其中,1≤i≤M,M为大于0的整数;
基准模型选择模块,对所述车辆的多属性中的第一属性预训练所述M个神经网络模型,并根据训练后的所述M个神经网络模型的识别结果进行比较,选择识别性能最佳的神经网络模型作为基准神经网络模型Base,其单元数目为N,其中,N为大于0的整数;
分支确定模块,取基准神经网络模型Base的前半个网络作为车辆多属性识别模型的低层特征提取器,前半个网络的末端作为整个车辆多属性识别网络的起始分支点,其中,所述前半个网络包含N/2个单元;
车辆多属性识别网络构建模块,针对所述车辆的多属性中的第二属性,分别应用N/4个不同的基本单元Ci作为起始分支点后的分支神经网络的基本结构对不同的第二属性进行识 别,并比较识别结果,为第二属性选择一个识别性能最佳的基本单元作为分支点后N/4个单元的网络结构,重复上述操作,为车辆多属性的其他属性选择最佳的基本单元,以构建多分支的车辆多属性识别网络;
识别模块,使用所述车辆多属性识别网络对车辆的属性进行识别。
更进一步地,所述车辆多属性为图片中车辆的车系型号、车身颜色、车型以及车牌类型,其中第一属性为车系型号,第二属性为车身颜色。
更进一步地,所述M=5,所述神经网络模型分别为:googlenet、resnet、resneXt、SENet和SKNet。
更进一步地,所述车辆多属性识别网络构建模块中使用回溯更正机制对车辆多属性识别网络进行校正,在对车辆多属性识别网络的一属性分支增加网络深度但无法提高识别精度,则将该属性分支在下一次的搜索过程中反向减少其网络层数。
更进一步地,所述车辆多属性识别网络构建模块中使用基本单元选取更正机制对车辆多属性识别网络进行校正,在对车辆多属性识别网络的一属性分支中,若在一次搜索中,特定的基本单元对其识别准确率有较大影响时,则在下一次搜索对该单元做特殊处理。
本公开还提出了一种计算机可读存储介质,所述存储介质上存储有计算机程序代码,当所述计算机程序代码被计算机执行时执行上述之任一的方法。
本公开的技术效果为:先基于难以识别的车辆属性选择基准神经网络模型,然后在该基准神经网络模型的基础上,选择前半个网络(包含N/2个单元)作为整个车辆多属性识别网络模型的低层特征提取器,然后,其末端作为分支点,为其他的属性构建识别分支,且采用类似二分法的方法确定各个分支网络剩余的结构,即,使用第一次迭代选择出的N/4个基本网络单元、使用第二次选择出来的N/8个基本网络单元,依次类推。通过上述结合神经网络结构搜索技术的构建多分支网络的方法进而构建出的多分支车辆多属性识别网络,可以一次识别车辆的多个属性,如型号、颜色等等,提高了识别效率,这是本公开的重要公开点之一。本公开提出了的两种网络更正机制可以使得生成的多分支的车辆多属性识别网络性能更佳,且在网络构建时效率更高,这是本公开的另一个重要公开点。
附图说明
通过阅读参照以下附图所作的对非限制性实施例所作的详细描述,本公开的其它特征、目的和优点将会变得更明显。
图1是根据本公开的实施例的一种基于神经网络结构搜索的车辆多属性识别方法的流程图。
图2是googlenet的一个inception结构示意图。
图3是根据本公开的实施例的一种基于神经网络结构搜索的车辆多属性识别装置的结构图。
具体实施方式
下面结合附图和实施例对本公开作进一步的详细说明。可以理解的是,此处所描述的具体实施例仅仅用于解释相关公开,而非对该公开的限定。另外还需要说明的是,为了便于描述,附图中仅示出了与有关公开相关的部分。
需要说明的是,在不冲突的情况下,本公开中的实施例及实施例中的特征可以相互组合。下面将参考附图并结合实施例来详细说明本公开。
图1示出了本公开的一种基于神经网络结构搜索的车辆多属性识别方法,该方法包括:
确定步骤S101,确定M个神经网络模型的基本网络结构作为网络搜索空间的基本单元Ci,其中,1≤i≤M,M为大于0的整数。通过对现有的网络模型,如VGG,googlenet,resnet等进行分析发现,其网络重复性较高,通常为单一网络结构的多次叠加,如googlenet由用多个inception结构叠加而来,图2示出了googlenet的一个inception。
基准模型选择步骤S102,对所述车辆的多属性中的第一属性预训练所述M个神经网络模型,并根据训练后的所述M个神经网络模型的识别结果进行比较,选择识别性能最佳的神经网络模型作为基准神经网络模型Base,其单元数目为N,其中,N为大于0的整数。
在一个实施例中,车辆属性识别中,相比于颜色等,车系型号的识别是最具挑战性的。即在构建车辆多属性识别网络时,先从最难以识别的车辆属性进行入手,通过对这一单一属性预训练多种神经网络模型(googlenet,resnet,resneXt,SENet,SKNet等)并对其识别结果进行比较,将最好的模型作为基准神经网络模型Base,记录其单元数目N。
分支确定步骤S103,取基准神经网络模型Base的前半个网络作为车辆多属性识别模型的低层特征提取器,前半个网络的末端作为整个车辆多属性识别网络的起始分支点,其中,所述前半个网络包含N/2个单元。
在一个实施例中,对于神经网络而言,不同的标签分类对应着不同的高层语义特征,整个网络可以视为低层特征向高层语义特征转化映射的过程。故取Base模型的前半个网络(包含N/2个单元)作为整个车辆多属性识别网络模型的低层特征提取器,其末端作为整个车辆多属性识别多标签分类网络的起始分支点。
车辆多属性识别网络构建步骤S104,针对所述车辆的多属性中的第二属性,分别应用N/4个不同的基本单元Ci作为起始分支点后的分支神经网络的基本结构对不同的第二属性 进行识别,并比较识别结果,为第二属性选择一个识别性能最佳的基本单元作为分支点后N/4个单元的网络结构,重复上述操作,为车辆多属性的其他属性选择最佳的基本单元,以构建多分支的车辆多属性识别网络。
在一个实施例中,得到Base模型及网络不同分支的起始点后,对于剩余分支网络(即刨除车系型号属性分支)使用二分法在确定的网络搜索空间进行折半搜索,即首先分别应用N/4个不同的基本单元Ci作为分支网络的基本结构对另外的属性进行识别,并比较其结果,以此确定分支点后N/4个单元的网络结构;以颜色属性分支为例,在第一次迭代中,在Base网络的起始分支点后接续N/4个预定义的基本单元(共有M个),故在本次迭代中可得到M个不同的颜色属性候选分支网络,比较各个候选分支网络对车辆颜色的识别率,从中选取识别率最高的作为颜色分支的基本单元。
在一个实施例中,所述车辆多属性为图片中车辆的车系型号、车身颜色、车型以及车牌类型,其中第一属性为车系型号,第二属性为车身颜色。M=5,所述神经网络模型分别为:googlenet、resnet、resneXt、SENet和SKNet。
本公开中先基于难以识别的车辆属性选择基准神经网络模型,然后在该基准神经网络模型的基础上,选择前半个网络(包含N/2个单元)作为整个车辆多属性识别网络模型的低层特征提取器,然后,其末端作为分支点,为其他的属性构建识别分支,且采用类似二分法的方法确定网络的结构,即,使用第一次迭代选择出的N/4个基本网络单元、使用第二次选择出来的N/8个基本网络单元,依次类推。如N=8时,在不考虑回溯更正机制情况下,以颜色属性分支为例,第二次迭代过程所需接续的基本单元个数为N/8=1,故其需要重复上述操作的次数为2。通过上述结合神经网络结构搜索技术的构建多分支网络的方法进而构建出多分支的车辆多属性识别网络,可以一次识别车辆的多个属性,如型号、颜色等等,提高了识别效率,这是本公开的重要公开点之一。
识别步骤S105,使用所述车辆多属性识别网络对车辆的属性进行识别。使用车辆多属性识别网络对输入的车辆图片进行识别,一次可以识别出来车辆的多个属性,如型号、颜色等等,提高了识别效率。
在一个实施例中,设计了网络的更正机制,所述车辆多属性识别网络构建步骤S104中使用回溯更正机制对车辆多属性识别网络进行校正,在对车辆多属性识别网络的一属性分支增加网络深度但无法提高识别精度,则将该属性分支在下一次的搜索过程中反向减少其网络层数。以颜色属性分支为例,使用类似二叉树搜索中的宽度优先算法,若在第二次迭代过程中(此时这一分支网络的深度应为N/2+N/4+N/8)所得的候选网络的识别率均小于第一 次迭代过程中所得的候选网络(其网络深度为N/2+N/4)的识别率,则在第三次搜索迭代过程中其候选网络的深度应为N/2+N/4-N/8=N/2+N/8。
在一个实施例中,设计了网络的另一种更正机制,所述车辆多属性识别网络构建步骤S104中使用基本单元选取更正机制对车辆多属性识别网络进行校正,在对车辆多属性识别网络的一属性分支中,若在一次搜索中,特定的基本单元对其识别准确率有较大影响时,则在下一次搜索对该单元做特殊处理。即通过候选网络的识别率来对不同的Ci做不同的加权处理,如最差候选网络的基本单元不参与下一次迭代过程,以减少搜索时间。以颜色属性分支为例,若在第一次迭代搜索过程中,若使用了inception v1基本单元的候选网络的识别率最差,则在第二次迭代搜索过程中在候选基本单元中剔除inception v1基本单元。
本公开提出了的两种网络更正机制可以使得生成的多分支的车辆多属性识别网络性能更佳,且在网络构建时效率更高,这是本公开的另一个重要公开点。
进一步参考图3,作为对上述图1所示方法的实现,本公开提供了一种基于神经网络结构搜索的车辆多属性识别装置的一个实施例,该装置实施例与图1所示的方法实施例相对应,该装置具体可以包含于各种电子设备或通过网络连接的多个电子设备中。
图3示出了本公开的图1示出了本公开的一种基于神经网络结构搜索的车辆多属性识别装置,该装置包括:
确定模块301,确定M个神经网络模型的基本网络结构作为网络搜索空间的基本单元Ci,其中,1≤i≤M,M为大于0的整数。通过对现有的网络模型,如VGG,googlenet,resnet等进行分析发现,其网络重复性较高,通常为单一网络结构的多次叠加,如googlenet由用多个inception结构叠加而来,图2示出了googlenet的一个inception。
基准模型选择模块302,对所述车辆的多属性中的第一属性预训练所述M个神经网络模型,并根据训练后的所述M个神经网络模型的识别结果进行比较,选择识别性能最佳的神经网络模型作为基准神经网络模型Base,其单元数目为N,其中,N为大于0的整数。
在一个实施例中,车辆属性识别中,相比于颜色等,车系型号的识别是最具挑战性的。即在构建车辆多属性识别网络时,先从最难以识别的车辆属性进行入手,通过对这一单一属性预训练多种神经网络模型(googlenet,resnet,resneXt,SENet,SKNet等)并对其识别结果进行比较,将最好的模型作为基准神经网络模型Base,记录其单元数目N。
分支确定模块303,取基准神经网络模型Base的前半个网络作为车辆多属性识别模型的低层特征提取器,前半个网络的末端作为整个车辆多属性识别网络的起始分支点,其中,所述前半个网络包含N/2个单元。
在一个实施例中,对于神经网络而言,不同的标签分类对应着不同的高层语义特征,整个网络可以视为低层特征向高层语义特征转化映射的过程。故取Base模型的前半个网络(包含N/2个单元)作为整个车辆多属性识别网络模型的低层特征提取器,其末端作为整个车辆多属性识别多标签分类网络的起始分支点。
车辆多属性识别网络构建模块304,针对所述车辆的多属性中的第二属性,分别应用N/4个不同的基本单元Ci作为起始分支点后的分支神经网络的基本结构对不同的第二属性进行识别,并比较识别结果,为第二属性选择一个识别性能最佳的基本单元作为分支点后N/4个单元的网络结构,重复上述操作,为车辆多属性的其他属性选择最佳的基本单元,以构建多分支的车辆多属性识别网络。
在一个实施例中,得到Base模型及网络不同分支的起始点后,对于剩余分支网络(即刨除车系型号属性分支)使用二分法在确定的网络搜索空间进行折半搜索,即首先分别应用N/4个不同的基本单元Ci作为分支网络的基本结构对另外的属性进行识别,并比较其结果,以此确定分支点后N/4个单元的网络结构;以颜色属性分支为例,在第一次迭代中,在Base网络的起始分支点后接续N/4个预定义的基本单元(共有M个),故在本次迭代中可得到M个不同的颜色属性候选分支网络,比较各个候选分支网络对车辆颜色的识别率,从中选取识别率最高的作为颜色分支的基本单元。
在一个实施例中,所述车辆多属性为图片中车辆的车系型号、车身颜色、车型以及车牌类型,其中第一属性为车系型号,第二属性为车身颜色。M=5,所述神经网络模型分别为:googlenet、resnet、resneXt、SENet和SKNet。
本公开中先基于难以识别的车辆属性选择基准神经网络模型,然后在该基准神经网络模型的基础上,选择前半个网络(包含N/2个单元)作为整个车辆多属性识别网络模型的低层特征提取器,然后,其末端作为分支点,为其他的属性构建识别分支,且采用类似二分法的方法确定网络的结构,即,使用第一次迭代选择出的N/4个基本网络单元、使用第二次选择出来的N/8个基本网络单元依次类推。如N=8时,在不考虑回溯更正机制情况下,以颜色属性分支为例,第二次迭代过程所需接续的基本单元个数为N/8=1,故其需要重复上述操作的次数为2。通过上述结合神经网络结构搜索技术的构建多分支网络的方法进而构建出多分支的车辆多属性识别网络,可以一次识别车辆的多个属性,如型号、颜色等等,提高了识别效率,这是本公开的重要公开点之一。
识别模块305,使用所述车辆多属性识别网络对车辆的属性进行识别。使用车辆多属性识别网络对输入的车辆图片进行识别,一次可以识别出来车辆的多个属性,如型号、颜色等等,提高了识别效率。
在一个实施例中,设计了网络的更正机制,所述车辆多属性识别网络构建模块304中使用回溯更正机制对车辆多属性识别网络进行校正,在对车辆多属性识别网络的一属性分支增加网络深度但无法提高识别精度,则将该属性分支在下一次的搜索过程中反向减少其网络层数。以颜色属性分支为例,使用类似二叉树搜索中的宽度优先算法,若在第二次迭代过程中(此时这一分支网络的深度应为N/2+N/4+N/8)所得的候选网络的识别率均小于第一次迭代过程中所得的候选网络(其网络深度为N/2+N/4)的识别率,则在第三次搜索迭代过程中其候选网络的深度应为N/2+N/4-N/8=N/2+N/8。
在一个实施例中,设计了网络的另一种更正机制,所述车辆多属性识别网络构建模块304中使用基本单元选取更正机制对车辆多属性识别网络进行校正,在对车辆多属性识别网络的一属性分支中,若在一次搜索中,特定的基本单元对其识别准确率有较大影响时,则在下一次搜索对该单元做特殊处理。即通过候选网络的识别率来对不同的Ci做不同的加权处理,如最差候选网络的基本单元不参与下一次迭代过程,以减少搜索时间。以颜色属性分支为例,若在第一次迭代搜索过程中,若使用了inception v1基本单元的候选网络的识别率最差,则在第二次迭代搜索过程中在候选基本单元中剔除inception v1基本单元。
本公开提出了的两种网络更正机制可以使得生成的多分支的车辆多属性识别网络性能更佳,且在网络构建时效率更高,这是本公开的另一个重要公开点。
本公开将网络结构搜索技术应用于多标签分类车辆多属性识别任务中,通过引入人工先验知识,固化网络搜索空间及路径,可有限降低网络结构搜索的时间复杂度。同样的方法也可推广到其他的多标签分类任务中。
为了描述的方便,描述以上装置时以功能分为各种单元分别描述。当然,在实施本公开时可以把各单元的功能在同一个或多个软件和/或硬件中实现。
通过以上的实施方式的描述可知,本领域的技术人员可以清楚地了解到本公开可借助软件加必需的通用硬件平台的方式来实现。基于这样的理解,本公开的技术方案本质上或者说对现有技术做出贡献的部分可以以软件产品的形式体现出来,该计算机软件产品可以存储在存储介质中,如ROM/RAM、磁碟、光盘等,包括若干指令用以使得一台计算机设备(可以是个人计算机,服务器,或者网络设备等)执行本公开各个实施例或者实施例的某些部分所述的方法。
最后所应说明的是:以上实施例仅以说明而非限制本公开的技术方案,尽管参照上述实施例对本公开进行了详细说明,本领域的普通技术人员应当理解:依然可以对本公开进行修改或者等同替换,而不脱离本公开的精神和范围的任何修改或局部替换,其均应涵盖在本公开的权利要求范围当中。

Claims (11)

  1. 一种基于神经网络结构搜索的车辆多属性识别方法,其特征在于,该方法包括:
    确定步骤,确定M个神经网络模型的基本网络结构作为网络搜索空间的基本单元Ci,其中,1≤i≤M,M为大于0的整数;
    基准模型选择步骤,对所述车辆的多属性中的第一属性预训练所述M个神经网络模型,并根据训练后的所述M个神经网络模型的识别结果进行比较,选择识别性能最佳的神经网络模型作为基准神经网络模型Base,其单元数目为N,其中,N为大于0的整数;
    分支确定步骤,取基准神经网络模型Base的前半个网络作为车辆多属性识别模型的低层特征提取器,前半个网络的末端作为整个车辆多属性识别网络的起始分支点,其中,所述前半个网络包含N/2个单元;
    车辆多属性识别网络构建步骤,针对所述车辆的多属性中的第二属性,分别应用N/4个不同的基本单元Ci作为起始分支点后的分支神经网络的基本结构对不同的第二属性进行识别,并比较识别结果,为第二属性选择一个识别性能最佳的基本单元作为分支点后N/4个单元的网络结构,重复上述操作,为车辆多属性的其他属性选择最佳的基本单元,以构建多分支的车辆多属性识别网络;
    识别步骤,使用所述车辆多属性识别网络对车辆的属性进行识别。
  2. 根据权利要求1所述的方法,其特征在于,所述车辆多属性为图片中车辆的车系型号、车身颜色、车型以及车牌类型,其中第一属性为车系型号,第二属性为车身颜色。
  3. 根据权利要求2所述的方法,其特征在于,所述M=5,所述神经网络模型分别为:googlenet、resnet、resneXt、SENet和SKNet。
  4. 根据权利要求3所述的方法,其特征在于,所述车辆多属性识别网络构建步骤中使用回溯更正机制对车辆多属性识别网络进行校正,在对车辆多属性识别网络的一属性分支增加网络深度但无法提高识别精度,则将该属性分支在下一次的搜索过程中反向减少其网络层数。
  5. 根据权利要求3任一项所述的方法,其特征在于,所述车辆多属性识别网络构建步骤中使用基本单元选取更正机制对车辆多属性识别网络进行校正,在对车辆多属性识别网络的一属性分支中,若在一次搜索中,特定的基本单元对其识别准确率有较大影响时,则在下一次搜索对该单元做特殊处理。
  6. 一种基于神经网络结构搜索的车辆多属性识别装置,其特征在于,该装置包括:
    确定模块,确定M个神经网络模型的基本网络结构作为网络搜索空间的基本单元Ci,其中,1≤i≤M,M为大于0的整数;
    基准模型选择模块,对所述车辆的多属性中的第一属性预训练所述M个神经网络模型,并根据训练后的所述M个神经网络模型的识别结果进行比较,选择识别性能最佳的神经网络模型作为基准神经网络模型Base,其单元数目为N,其中,N为大于0的整数;
    分支确定模块,取基准神经网络模型Base的前半个网络作为车辆多属性识别模型的低层特征提取器,前半个网络的末端作为整个车辆多属性识别网络的起始分支点,其中,所述前半个网络包含N/2个单元;
    车辆多属性识别网络构建模块,针对所述车辆的多属性中的第二属性,分别应用N/4个不同的基本单元Ci作为起始分支点后的分支神经网络的基本结构对不同的第二属性进行识别,并比较识别结果,为第二属性选择一个识别性能最佳的基本单元作为分支点后N/4个单元的网络结构,重复上述操作,为车辆多属性的其他属性选择最佳的基本单元,以构建多分支的车辆多属性识别网络;
    识别模块,使用所述车辆多属性识别网络对车辆的属性进行识别。
  7. 根据权利要求6所述的装置,其特征在于,所述车辆多属性为图片中车辆的车系型号、车身颜色、车型以及车牌类型,其中第一属性为车系型号,第二属性为车身颜色。
  8. 根据权利要求7所述的装置,其特征在于,所述M=5,所述神经网络模型分别为:googlenet、resnet、resneXt、SENet和SKNet。
  9. 根据权利要求8所述的装置,其特征在于,所述车辆多属性识别网络构建模块中使用回溯更正机制对车辆多属性识别网络进行校正,在对车辆多属性识别网络的一属性分支增加网络深度但无法提高识别精度,则将该属性分支在下一次的搜索过程中反向减少其网络层数。
  10. 根据权利要求3任一项所述的装置,其特征在于,所述车辆多属性识别网络构建模块中使用基本单元选取更正机制对车辆多属性识别网络进行校正,在对车辆多属性识别网络的一属性分支中,若在一次搜索中,特定的基本单元对其识别准确率有较大影响时,则在下一次搜索对该单元做特殊处理。
  11. 一种计算机可读存储介质,其特征在于,所述存储介质上存储有计算机程序代码,当所述计算机程序代码被计算机执行时执行权利要求1-5之任一的方法。
PCT/CN2019/100027 2019-05-23 2019-08-09 基于神经网络结构搜索的车辆多属性识别方法、装置、介质 WO2020232840A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910436423.0A CN110209857B (zh) 2019-05-23 2019-05-23 基于神经网络结构搜索的车辆多属性识别方法、装置、介质
CN201910436423.0 2019-05-23

Publications (1)

Publication Number Publication Date
WO2020232840A1 true WO2020232840A1 (zh) 2020-11-26

Family

ID=67788451

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/100027 WO2020232840A1 (zh) 2019-05-23 2019-08-09 基于神经网络结构搜索的车辆多属性识别方法、装置、介质

Country Status (2)

Country Link
CN (1) CN110209857B (zh)
WO (1) WO2020232840A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113378984A (zh) * 2021-07-05 2021-09-10 国药(武汉)医学实验室有限公司 一种医学图像分类方法、系统、终端以及存储介质

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP4030347A4 (en) * 2019-09-24 2022-11-16 Huawei Technologies Co., Ltd. METHOD AND DEVICE FOR CONSTRUCTING NEURONAL NETWORK, AND METHOD AND DEVICE FOR IMAGE PROCESSING
CN112580794A (zh) * 2019-09-29 2021-03-30 佳能株式会社 属性识别装置、方法和系统及识别对象属性的神经网络
CN112598076B (zh) * 2020-12-29 2023-09-19 北京易华录信息技术股份有限公司 一种机动车属性识别方法及系统
CN113313079B (zh) * 2021-07-16 2021-11-12 深圳市安软科技股份有限公司 一种车辆属性识别模型的训练方法、系统及相关设备

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105740906A (zh) * 2016-01-29 2016-07-06 中国科学院重庆绿色智能技术研究院 一种基于深度学习的车辆多属性联合分析方法
CN106203330A (zh) * 2016-07-08 2016-12-07 西安理工大学 一种基于卷积神经网络的车辆分类方法
US20170140253A1 (en) * 2015-11-12 2017-05-18 Xerox Corporation Multi-layer fusion in a convolutional neural network for image classification
CN106778583A (zh) * 2016-12-07 2017-05-31 北京理工大学 基于卷积神经网络的车辆属性识别方法与装置
CN107633220A (zh) * 2017-09-13 2018-01-26 吉林大学 一种基于卷积神经网络的车辆前方目标识别方法
CN108549926A (zh) * 2018-03-09 2018-09-18 中山大学 一种用于精细化识别车辆属性的深度神经网络及训练方法

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102478668A (zh) * 2010-11-30 2012-05-30 中国石油天然气集团公司 一种应用地震多属性参数预测煤层厚度的方法
CN105809251B (zh) * 2016-03-01 2018-01-30 北京交通大学 一种多属性融合的动车组系统关键部件辨识方法
CN106570477B (zh) * 2016-10-28 2019-07-16 中科视语(北京)科技有限公司 基于深度学习的车型识别模型构建方法及车型识别方法
US10565494B2 (en) * 2016-12-31 2020-02-18 Via Alliance Semiconductor Co., Ltd. Neural network unit with segmentable array width rotator
CN108898060A (zh) * 2018-05-30 2018-11-27 珠海亿智电子科技有限公司 车载环境下基于卷积神经网络的车型识别方法
US10176405B1 (en) * 2018-06-18 2019-01-08 Inception Institute Of Artificial Intelligence Vehicle re-identification techniques using neural networks for image analysis, viewpoint-aware pattern recognition, and generation of multi- view vehicle representations

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170140253A1 (en) * 2015-11-12 2017-05-18 Xerox Corporation Multi-layer fusion in a convolutional neural network for image classification
CN105740906A (zh) * 2016-01-29 2016-07-06 中国科学院重庆绿色智能技术研究院 一种基于深度学习的车辆多属性联合分析方法
CN106203330A (zh) * 2016-07-08 2016-12-07 西安理工大学 一种基于卷积神经网络的车辆分类方法
CN106778583A (zh) * 2016-12-07 2017-05-31 北京理工大学 基于卷积神经网络的车辆属性识别方法与装置
CN107633220A (zh) * 2017-09-13 2018-01-26 吉林大学 一种基于卷积神经网络的车辆前方目标识别方法
CN108549926A (zh) * 2018-03-09 2018-09-18 中山大学 一种用于精细化识别车辆属性的深度神经网络及训练方法

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113378984A (zh) * 2021-07-05 2021-09-10 国药(武汉)医学实验室有限公司 一种医学图像分类方法、系统、终端以及存储介质
CN113378984B (zh) * 2021-07-05 2023-05-02 国药(武汉)医学实验室有限公司 一种医学图像分类方法、系统、终端以及存储介质

Also Published As

Publication number Publication date
CN110209857B (zh) 2022-04-22
CN110209857A (zh) 2019-09-06

Similar Documents

Publication Publication Date Title
WO2020232840A1 (zh) 基于神经网络结构搜索的车辆多属性识别方法、装置、介质
Liu et al. Open compound domain adaptation
CN113591902B (zh) 基于多模态预训练模型的跨模态理解与生成方法和装置
US9619735B1 (en) Pure convolutional neural network localization
CA3022998A1 (en) Method and device for generative adversarial network training
CN111738016B (zh) 多意图识别方法及相关设备
CN111461226A (zh) 对抗样本生成方法、装置、终端及可读存储介质
Mannhardt et al. Unsupervised event abstraction using pattern abstraction and local process models
Ji et al. Unsupervised few-shot feature learning via self-supervised training
KR20180044295A (ko) 트레이닝된 머신 학습 모델의 성능을 개선시키는 방법
KR20170106338A (ko) 모델 압축 및 미세-튜닝
US10832149B2 (en) Automatic segmentation of data derived from learned features of a predictive statistical model
CN109783666A (zh) 一种基于迭代精细化的图像场景图谱生成方法
CN112232395B (zh) 一种基于联合训练生成对抗网络的半监督图像分类方法
CN110414005A (zh) 意图识别方法、电子设备及存储介质
CN113033410B (zh) 基于自动数据增强的域泛化行人重识别方法、系统及介质
Wei et al. Efficient robustness assessment via adversarial spatial-temporal focus on videos
WO2021180243A1 (zh) 基于机器学习的图像信息识别的优化方法及装置
CN113254645A (zh) 一种文本分类方法、装置、计算机设备及可读存储介质
US9864930B2 (en) Clustering technique for optimized search over high-dimensional space
Li et al. Deep selective feature learning for action recognition
Balgi et al. Contradistinguisher: a vapnik’s imperative to unsupervised domain adaptation
Piergiovanni et al. Unsupervised action segmentation for instructional videos
US10546247B2 (en) Switching leader-endorser for classifier decision combination
Aghera et al. MnasNet based lightweight CNN for facial expression recognition

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: 19929917

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: 19929917

Country of ref document: EP

Kind code of ref document: A1